Method: 1) Narrative dynamic extraction to maintain coherent user profiles with dialogue-grounded information only; 2) Structured writing pipeline that batches, retrieves, clusters, and consolidates memory entries; 3) Multimodal memory module with ReAct-style reasoning (observe-think-act process) for video understanding.

Result: TeleMem outperforms state-of-the-art Mem0 baseline with 19% higher accuracy, 43% fewer tokens, and 2.1x speedup on the ZH-4O long-term role-play gaming benchmark.

Conclusion: TeleMem provides an effective solution for long-term multimodal memory management in LLMs, addressing key limitations of existing approaches through its unified architecture and efficient memory operations.

Abstract: Large language models (LLMs) excel at many NLP tasks but struggle to sustain long-term interactions due to limited attention over extended dialogue histories. Retrieval-augmented generation (RAG) mitigates this issue but lacks reliable mechanisms for updating or refining stored memories, leading to schema-driven hallucinations, inefficient write operations, and minimal support for multimodal reasoning.To address these challenges, we propose TeleMem, a unified long-term and multimodal memory system that maintains coherent user profiles through narrative dynamic extraction, ensuring that only dialogue-grounded information is preserved. TeleMem further introduces a structured writing pipeline that batches, retrieves, clusters, and consolidates memory entries, substantially improving storage efficiency, reducing token usage, and accelerating memory operations. Additionally, a multimodal memory module combined with ReAct-style reasoning equips the system with a closed-loop observe, think, and act process that enables accurate understanding of complex video content in long-term contexts. Experimental results show that TeleMem surpasses the state-of-the-art Mem0 baseline with 19% higher accuracy, 43% fewer tokens, and a 2.1x speedup on the ZH-4O long-term role-play gaming benchmark.

Method: Proposes VEJA framework as a new paradigm for data curation focusing on four core concepts: Values, Experiences, Judgments, and Abilities. Conducts pilot study comparing manually curated VEJA-grounded dataset against state-of-the-art synthetic baseline using LLM-as-judge evaluation.

Result: VEJA framework demonstrates significant quality gap over synthetic baseline, suggesting conceptually grounded data curation is necessary for creating roleplaying agents with genuine depth and narrative continuity.

Conclusion: Shift toward conceptually grounded data curation (VEJA framework) is essential for developing roleplaying agents with authentic character depth and narrative continuity, addressing systemic limitations of current approaches.

Abstract: Modern roleplaying models are increasingly sophisticated, yet they consistently struggle to capture the essence of believable, engaging characters. We argue this failure stems from training paradigms that overlook the dynamic interplay of a character’s internal world. Current approaches, including Retrieval-Augmented Generation (RAG), fact-based priming, literature-based learning, and synthetic data generation, exhibit recurring limitations in modeling the deliberative, value-conflicted reasoning that defines human interaction. In this paper, we identify four core concepts essential for character authenticity: Values, Experiences, Judgments, and Abilities (VEJA). We propose the VEJA framework as a new paradigm for data curation that addresses these systemic limitations. To illustrate the qualitative ceiling enabled by our framework, we present a pilot study comparing a manually curated, VEJA-grounded dataset against a state-of-the-art synthetic baseline. Using an LLM-as-judge evaluation, our findings demonstrate a significant quality gap, suggesting that a shift toward conceptually grounded data curation, as embodied by VEJA, is necessary for creating roleplaying agents with genuine depth and narrative continuity. The full dataset is available at https://github.com/HyouinKyoumaIRL/Operation-Veja

Method: Corpus-driven analysis using Vacaspati (literature) and IndicCorp (newspaper) corpora, measuring TTR, HLR, Bigram diversity, syllable/word lengths, Zipf’s Law adherence, n-gram perplexity, and readability indices (Flesch, Coleman-Liau).

Result: Literary corpus exhibits higher lexical richness, structural variation, perplexity, entropy, and complexity; adheres better to Zipf’s Law; and integrating literary data improves downstream task performance.

Conclusion: Bangla literary texts are linguistically richer and more complex than newspaper texts, and including literary data enhances NLP model capabilities, with findings generalizable to English corpora as well.

Abstract: In this paper, we present a comprehensive corpus-driven analysis of Bangla literary and newspaper texts to investigate their lexical diversity, structural complexity and readability. We undertook Vacaspati and IndicCorp, which are the most extensive literature and newspaper-only corpora for Bangla. We examine key linguistic properties, including the type-token ratio (TTR), hapax legomena ratio (HLR), Bigram diversity, average syllable and word lengths, and adherence to Zipfs Law, for both newspaper (IndicCorp) and literary corpora (Vacaspati).For all the features, such as Bigram Diversity and HLR, despite its smaller size, the literary corpus exhibits significantly higher lexical richness and structural variation. Additionally, we tried to understand the diversity of corpora by building n-gram models and measuring perplexity. Our findings reveal that literary corpora have higher perplexity than newspaper corpora, even for similar sentence sizes. This trend can also be observed for the English newspaper and literature corpus, indicating its generalizability. We also examined how the perfor- mance of models on downstream tasks is influenced by the inclusion of literary data alongside newspaper data. Our findings suggest that inte- grating literary data with newspapers improves the performance of models on various downstream tasks. We have also demonstrated that a literary corpus adheres more closely to global word distribution proper- ties, such as Zipfs law, than a newspaper corpus or a merged corpus of both literary and newspaper texts. Literature corpora also have higher entropy and lower redundancy values compared to a newspaper corpus. We also further assess the readability using Flesch and Coleman-Liau in- dices, showing that literary texts are more complex.

Method: Sample-level soft reinforcement learning compression that penalizes inefficiently long rollouts, but only on problems where the model has already mastered and produced a more concise rollout.

Result: Reduces average response length by 20-40% with comparable or higher accuracy. Shows strong cross-domain generalization - training on math leads to spontaneous shortening on unseen tasks (code, instruction following, QA).

Conclusion: Demonstrates a stable post-training curriculum (accuracy-compression-accuracy) that produces more accurate and concise reasoning models. Argues such compression should be standard in developing efficient reasoning models.

Abstract: Chain-of-thought reasoning in large language models often creates an “overthinking trap,” leading to excessive computational cost and latency for unreliable accuracy gains. Prior work has typically relied on global, static controls that risk penalizing necessary reasoning. We introduce a sample-level, soft reinforcement learning compression method that penalizes inefficiently long rollouts, but only on problems where the model has already mastered and already produced a more concise rollout. Our experiments show that this method reduces average response length by 20-40% with comparable or higher accuracy. Crucially, the compression exhibits strong cross-domain generalization; a model trained on math spontaneously shortens responses on unseen tasks like code, instruction following, and general knowledge QA, with stable or improved accuracy. We demonstrate a stable post-training curriculum (accuracy-compression-accuracy) that can ultimately produce models that are more accurate and reason more concisely, arguing that such compression method should be a standard phase in developing efficient reasoning models.

Method: A multi-stage workflow using fine-tuned reasoning LLMs with evaluation of different fine-tuning strategies (SFT and GRPO), training small LLMs to generate reasoning tokens before final response, and analyzing the impact of different reward function combinations in GRPO training.

Result: The paper presents a novel approach for automatic compliance issue identification, compares effectiveness of different fine-tuning strategies, evaluates reasoning token generation in small LLMs, and assesses how reward function choices affect GRPO performance.

Conclusion: The proposed multi-stage workflow with fine-tuned reasoning LLMs offers an effective approach for automated marketing content compliance review, with specific insights on optimal fine-tuning strategies and reward function configurations.

Abstract: Reasoning Large Language Models (LLMs) have shown promising results when tasked with solving complex problems. In this paper, we propose and evaluate a multi-stage workflow that leverages the capabilities of fine-tuned reasoning LLMs to assist in the review process of marketing content, making sure they comply with a given list of requirements. The contributions of this paper are the following: (i) we present a novel approach – that does not rely on any external knowledge representation – for the automatic identification of compliance issues in textual content; (ii) compare the effectiveness of different fine-tuning strategies like Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) in training models to solve this problem; (iii) we evaluate the effectiveness of training small LLMs to generate reasoning tokens before providing their final response; (iv) we evaluate how the choice and combinations of different reward functions affects the performance of a model trained with GRPO.

Method: Proposes SIFT paradigm where frozen LLM generates supervision signals from textual speech representations. AZeroS implements this by training only two lightweight projection modules (23.8M params each) connecting frozen Qwen2.5-7B-Instruct LLM with audio encoders, using ~25K hours of ASR speech and ~3K hours of paralinguistic data.

Result: AZeroS achieves state-of-the-art performance on VoiceBench, AIR-Bench Foundation (Speech), and AIR-Bench Chat (Speech) benchmarks for both semantic and paralinguistic tasks, despite minimal training cost and modest data scale.

Conclusion: SIFT paradigm provides theoretically optimal generalization to unseen tasks by eliminating need for task-specific instruction data, enabling efficient training of high-performance speech-LLMs with frozen components and minimal parameter updates.

Abstract: Extending large language models (LLMs) to the speech domain has recently gained significant attention. A typical approach connects a pretrained LLM with an audio encoder through a projection module and trains the resulting model on large-scale, task-specific instruction-tuning datasets. However, curating such instruction-tuning data for specific requirements is time-consuming, and models trained in this manner often generalize poorly to unseen tasks. In this work, we first formulate that the strongest generalization of a speech-LLM is achieved when it is trained with Self-Generated Instruction-Free Tuning (SIFT), in which supervision signals are generated by a frozen LLM using textual representations of speech as input. Our proposed SIFT paradigm eliminates the need for collecting task-specific question-answer pairs and yields the theoretically best generalization to unseen tasks. Building upon this paradigm, we introduce AZeroS (Auden Zero-instruction-tuned Speech-LLM), which is trained on speech-text pairs derived from publicly available corpora, including approximately 25,000 hours of speech with ASR transcripts and 3,000 hours of speech with paralinguistic labels. Built upon Qwen2.5-7B-Instruct, the model updates only two lightweight projection modules (23.8 million parameters each), while keeping both the LLM and audio encoders frozen. Despite the minimal training cost and modest data scale, AZeroS achieves state-of-the-art performance on both semantic and paralinguistic benchmarks, including VoiceBench, AIR-Bench Foundation (Speech), and AIR-Bench Chat (Speech).

Method: Used parallel Bhagavad Gita verses in Sanskrit, English, and Hindi with transliteration. Tested multiple tokenizers (SentencePiece, GPT models, Gemini, GPT-4o) and measured token count, characters per token, and tokens per character.

Result: Sanskrit shows ~2x fewer tokens than English/Hindi under unbiased baseline. English/Hindi translations of Sanskrit commentary had ~20x more tokens. GPT-4o and Gemini reduce bias but still don’t fully capture Sanskrit’s compactness.

Conclusion: There’s significant tokenization bias against non-English languages, inflating costs. Sanskrit demonstrates potential for highly compact encoding, providing foundation for improved tokenizer design to reduce bias and improve efficiency.

Abstract: Tokens are the basic units of Large Language Models (LLMs). LLMs rely on tokenizers to segment text into these tokens, and tokenization is the primary determinant of computational and inference cost. Sanskrit, one of the oldest languages, is hypothesized to express more meaning per token due to its morphology and grammar rules; however, no prior work has quantified this. We use a dataset of 701 parallel verses of the Bhagavad Gita, which comprises three languages-Sanskrit, English, and Hindi along with transliteration of Sanskrit into English. We test tokenizers including SentencePiece (SPM), older GPT models, and the latest generation tokenizers from Gemini and GPT. We use metrics of token count, characters per token (token efficiency), and tokens per character (token cost). Results show a ~2x difference in token counts between Sanskrit and English/Hindi under the unbiased SPM baseline. English/Hindi translations of Sanskrit commentary resulted in an approximately 20x increase in token count. GPT o200k base (latest, used by GPT-4o) and Gemini (latest) reduce bias by a significant degree compared to GPT cl100k base (used until GPT-4), but still fail to fully capture Sanskrit’s compactness. This matters because there might be a penalty bias for non-English users, which inflates the token count. This research provides a foundation for improving future tokenizer design and shows the potential of Sanskrit for highly compact encoding, saving on cost while speeding up training and inference. The code and dataset are available at https://github.com/anshulkr713/sanskrit-token-efficiency

Method: Amory organizes conversational fragments into episodic narratives, consolidates memories with momentum, and semanticizes peripheral facts into semantic memory. It employs coherence-driven reasoning over narrative structures during retrieval time, actively constructing structured memory representations through agentic reasoning during offline periods.

Result: On the LOCOMO benchmark for long-term reasoning, Amory achieves considerable improvements over previous state-of-the-art, with performance comparable to full context reasoning while reducing response time by 50%. Momentum-aware consolidation significantly enhances response quality, and coherence-driven retrieval provides superior memory coverage compared to embedding-based approaches.

Conclusion: Amory demonstrates that actively constructing structured memory representations through agentic reasoning during offline time can effectively address the scalability challenges of long-term conversational agents while better capturing the coherence and subtlety of human memory.

Abstract: Long-term conversational agents face a fundamental scalability challenge as interactions extend over time: repeatedly processing entire conversation histories becomes computationally prohibitive. Current approaches attempt to solve this through memory frameworks that predominantly fragment conversations into isolated embeddings or graph representations and retrieve relevant ones in a RAG style. While computationally efficient, these methods often treat memory formation minimally and fail to capture the subtlety and coherence of human memory. We introduce Amory, a working memory framework that actively constructs structured memory representations through enhancing agentic reasoning during offline time. Amory organizes conversational fragments into episodic narratives, consolidates memories with momentum, and semanticizes peripheral facts into semantic memory. At retrieval time, the system employs coherence-driven reasoning over narrative structures. Evaluated on the LOCOMO benchmark for long-term reasoning, Amory achieves considerable improvements over previous state-of-the-art, with performance comparable to full context reasoning while reducing response time by 50%. Analysis shows that momentum-aware consolidation significantly enhances response quality, while coherence-driven retrieval provides superior memory coverage compared to embedding-based approaches.

Method: Introduced a Chain-of-Thought prompting framework that formalizes expert chemists’ reasoning steps (DBE analysis, neutral loss identification, fragment assembly) into structured prompts. Evaluated multiple state-of-the-art LLMs (Claude-3.5-Sonnet, GPT-4o-mini, Llama-3 series) in zero-shot setting using MassSpecGym dataset.

Result: LLMs can produce syntactically valid and partially plausible structures, but fail to achieve chemical accuracy or link reasoning to correct molecular predictions. Evaluation across SMILES validity, formula consistency, and structural similarity metrics reveals current limitations.

Conclusion: Findings highlight both interpretive potential and current limitations of LLM-based reasoning for molecular elucidation. Provides foundation for future work combining domain knowledge and reinforcement learning to achieve chemically grounded AI reasoning.

Abstract: Mass spectrometry (MS) is a powerful analytical technique for identifying small molecules, yet determining complete molecular structures directly from tandem mass spectra (MS/MS) remains a long-standing challenge due to complex fragmentation patterns and the vast diversity of chemical space. Recent progress in large language models (LLMs) has shown promise for reasoning-intensive scientific tasks, but their capability for chemical interpretation is still unclear. In this work, we introduce a Chain-of-Thought (CoT) prompting framework and benchmark that evaluate how LLMs reason about mass spectral data to predict molecular structures. We formalize expert chemists’ reasoning steps-such as double bond equivalent (DBE) analysis, neutral loss identification, and fragment assembly-into structured prompts and assess multiple state-of-the-art LLMs (Claude-3.5-Sonnet, GPT-4o-mini, and Llama-3 series) in a zero-shot setting using the MassSpecGym dataset. Our evaluation across metrics of SMILES validity, formula consistency, and structural similarity reveals that while LLMs can produce syntactically valid and partially plausible structures, they fail to achieve chemical accuracy or link reasoning to correct molecular predictions. These findings highlight both the interpretive potential and the current limitations of LLM-based reasoning for molecular elucidation, providing a foundation for future work that combines domain knowledge and reinforcement learning to achieve chemically grounded AI reasoning.

Method: Introduces eligibility criteria amendment prediction as an NLP task, releases AMEND++ benchmark suite with two datasets (AMEND and AMEND_LLM), and proposes Change-Aware Masked Language Modeling (CAMLM) - a revision-aware pretraining strategy that leverages historical edits.

Result: CAMLM consistently improves amendment prediction across diverse baselines, enabling more robust and cost-effective clinical trial design.

Conclusion: The proposed approach provides an effective method for predicting clinical trial eligibility criteria amendments, potentially reducing trial delays and costs through better trial design.

Abstract: Clinical trial amendments frequently introduce delays, increased costs, and administrative burden, with eligibility criteria being the most commonly amended component. We introduce \textit{eligibility criteria amendment prediction}, a novel NLP task that aims to forecast whether the eligibility criteria of an initial trial protocol will undergo future amendments. To support this task, we release $\texttt{AMEND++}$, a benchmark suite comprising two datasets: $\texttt{AMEND}$, which captures eligibility-criteria version histories and amendment labels from public clinical trials, and $\verb|AMEND_LLM|$, a refined subset curated using an LLM-based denoising pipeline to isolate substantive changes. We further propose $\textit{Change-Aware Masked Language Modeling}$ (CAMLM), a revision-aware pretraining strategy that leverages historical edits to learn amendment-sensitive representations. Experiments across diverse baselines show that CAMLM consistently improves amendment prediction, enabling more robust and cost-effective clinical trial design.

Method: The paper analyzes LoRA’s spectral properties and introduces Regularized Low-Rank Adaptation (RoRA) with three components: clean-strengthened regularization to increase spectral strength, trigger-insensitive constraints to improve spectral alignment, and post-training spectral rescaling.

Result: Experiments across multiple NLP benchmarks and attack settings show RoRA substantially reduces attack success rates while maintaining clean accuracy.

Conclusion: LoRA’s vulnerability to backdoors is fundamentally spectral, not just due to low rank. RoRA effectively addresses these spectral issues and improves forgetting of backdoor behaviors during fine-tuning.

Abstract: Low-Rank Adaptation (LoRA) is widely used for parameter-efficient fine-tuning of large language models, but it is notably ineffective at removing backdoor behaviors from poisoned pretrained models when fine-tuning on clean dataset. Contrary to the common belief that this weakness is caused primarily by low rank, we show that LoRA’s vulnerability is fundamentally spectral. Our analysis identifies two key factors: LoRA updates (i) possess insufficient spectral strength, with singular values far below those of pretrained weights, and (ii) exhibit unfavorable spectral alignment, weakly matching clean-task directions while retaining overlap with trigger-sensitive subspaces. We further establish a critical scaling threshold beyond which LoRA can theoretically suppress trigger-induced activations, and we show empirically that standard LoRA rarely reaches this regime. We introduce Regularized Low-Rank Adaptation (RoRA), which improves forgetting by increasing spectral strength and correcting alignment through clean-strengthened regularization, trigger-insensitive constraints, and post-training spectral rescaling. Experiments across multiple NLP benchmarks and attack settings show that RoRA substantially reduces attack success rates while maintaining clean accuracy.

Method: Unified architecture integrating BanglaBERT embeddings with multiple parallel processing branches using GRUs and CNNs, followed by attention mechanisms and dense layers for final classification. This hybrid approach captures both contextual semantics and local linguistic patterns.

Result: Achieved competitive performance: 0.7345 micro F1-Score (2nd place) in Subtask 1A and 0.7317 micro F1-Score (5th place) in Subtask 1B of the BLP-2025 competition.

Conclusion: The proposed unified architecture effectively addresses Bangla hate speech detection, demonstrating that combining pre-trained language model embeddings with parallel GRU/CNN branches and attention mechanisms yields robust performance across different classification subtasks.

Abstract: This paper describes our system used in the BLP-2025 Task 1: Hate Speech Detection. We participated in Subtask 1A and Subtask 1B, addressing hate speech classification in Bangla text. Our approach employs a unified architecture that integrates BanglaBERT embeddings with multiple parallel processing branches based on GRUs and CNNs, followed by attention and dense layers for final classification. The model is designed to capture both contextual semantics and local linguistic cues, enabling robust performance across subtasks. The proposed system demonstrated high competitiveness, obtaining 0.7345 micro F1-Score (2nd place) in Subtask 1A and 0.7317 micro F1-Score (5th place) in Subtask 1B.

Method: GRPO-style fine-tuning using Machine Translation Quality Estimation (MTQE) models as reward functions to train models to better translate idioms. Experiments conducted using Chinese and Hindi idiom datasets.

Result: Idiom translation abilities improve by ~14 points, general non-idiomatic translation implicitly improves by ~8 points, and cross-lingual translation abilities (trained on one language, evaluated on another) improves by ~6 points.

Conclusion: The work quantifies the non-compositional translation gap and offers insights for developing LLMs with stronger cross-cultural and figurative language understanding.

Abstract: Non-compositional expressions (e.g., idioms, proverbs, and metaphors) pose significant challenges for neural machine translation systems because their meanings cannot be derived from individual words alone. These expressions encode rich, cultural meaning, and have both figurative and literal meanings, making accurate translation difficult. Because models are fairly good at translating compositional text, we investigate GRPO-style fine-tuning using Machine Translation Quality Estimation (MTQE) models as reward functions to train models to better translate idioms. Using Chinese and Hindi idiom datasets, we find that idiom translation abilities improve by ~14 points, general, non-idiomatic translation implicitly improves by ~8 points, and cross-lingual translation abilities (trained on one language, evaluated on another) improves by ~6 points. Overall, our work quantifies the non-compositional translation gap and offers insights for developing LLMs with stronger cross-cultural and figurative language understanding.

Method: Created W&C-Sent dataset with detailed data collection, annotation, and quality-control procedures; evaluated range of LLMs on identifying trust, sociability, and competence in text.

Result: Dataset includes over 1,600 English sentence-target pairs annotated along three dimensions (trust, sociability, competence) from social media, providing new resource for NLP analysis.

Conclusion: W&C-Sent enables analysis of warmth and competence in language and supports future research at intersection of NLP and computational social science.

Abstract: Warmth (W) (often further broken down into Trust (T) and Sociability (S)) and Competence (C) are central dimensions along which people evaluate individuals and social groups (Fiske, 2018). While these constructs are well established in social psychology, they are only starting to get attention in NLP research through word-level lexicons, which do not completely capture their contextual expression in larger text units and discourse. In this work, we introduce Warmth and Competence Sentences (W&C-Sent), the first sentence-level dataset annotated for warmth and competence. The dataset includes over 1,600 English sentence–target pairs annotated along three dimensions: trust and sociability (components of warmth), and competence. The sentences in W&C-Sent are from social media and often express attitudes and opinions about specific individuals or social groups (the targets of our annotations). We describe the data collection, annotation, and quality-control procedures in detail, and evaluate a range of large language models (LLMs) on their ability to identify trust, sociability, and competence in text. W&C-Sent provides a new resource for analyzing warmth and competence in language and supports future research at the intersection of NLP and computational social science.

Method: TELEVAL evaluates SLMs through two core aspects: 1) Reliable Content Fulfillment (semantic understanding and correct responses), and 2) Interactional Appropriateness (socially capable, human-like responses with paralinguistic cues).

Result: Experiments show current SLMs perform well on semantic/knowledge tasks but struggle to produce natural, interactionally appropriate responses, revealing a gap in interaction-faithful evaluation.

Conclusion: There’s a need for more interaction-focused evaluation benchmarks like TELEVAL to better assess SLMs’ performance in realistic spoken conversations beyond just semantic accuracy.

Abstract: Spoken language models (SLMs) have advanced rapidly in recent years, accompanied by a growing number of evaluation benchmarks. However, most existing benchmarks emphasize task completion and capability scaling, while remaining poorly aligned with how users interact with SLMs in real-world spoken conversations. Effective spoken interaction requires not only accurate understanding of user intent and content, but also the ability to respond with appropriate interactional strategies. In this paper, we present TELEVAL, a dynamic, user-centered benchmark for evaluating SLMs in realistic Chinese spoken interaction scenarios. TELEVAL consolidates evaluation into two core aspects. Reliable Content Fulfillment assesses whether models can comprehend spoken inputs and produce semantically correct responses. Interactional Appropriateness evaluates whether models act as socially capable interlocutors, requiring them not only to generate human-like, colloquial responses, but also to implicitly incorporate paralinguistic cues for natural interaction. Experiments reveal that, despite strong performance on semantic and knowledge-oriented tasks, current SLMs still struggle to produce natural and interactionally appropriate responses, highlighting the need for more interaction-faithful evaluation.

Method: Proposes a variety of likelihood- and generative-based evaluation methods as alternatives to naive global token perplexity for assessing spoken language models.

Result: The proposed evaluations more faithfully reflect perceived generation quality with stronger correlations to human-rated mean opinion scores (MOS). Under new metrics, the performance gap between best models and human topline is significantly reduced.

Conclusion: Appropriate evaluation is critical for accurately assessing progress in spoken language modeling, as current text-based metrics may underestimate model capabilities.

Abstract: Generative spoken language models pretrained on large-scale raw audio can continue a speech prompt with appropriate content while preserving attributes like speaker and emotion, serving as foundation models for spoken dialogue. In prior literature, these models are often evaluated using ``global token perplexity’’, which directly applies the text perplexity formulation to speech tokens. However, this practice overlooks fundamental differences between speech and text modalities, possibly leading to an underestimation of the speech characteristics. In this work, we propose a variety of likelihood- and generative-based evaluation methods that serve in place of naive global token perplexity. We demonstrate that the proposed evaluations more faithfully reflect perceived generation quality, as evidenced by stronger correlations with human-rated mean opinion scores (MOS). When assessed under the new metrics, the relative performance landscape of spoken language models is reshaped, revealing a significantly reduced gap between the best-performing model and the human topline. Together, these results suggest that appropriate evaluation is critical for accurately assessing progress in spoken language modeling.

Method: Conducted extensive experiments on architectures, transformer backbones, training objectives, and data composition across many languages, then used these insights to develop Otter, a universal multilingual NER model.

Result: Otter achieves 5.3pp F1 improvement over GLiNER-x-base and competitive performance with large generative models like Qwen3-32B while being substantially more efficient.

Conclusion: The systematic analysis approach enables better understanding of what improves multilingual NER performance, and Otter demonstrates strong results across 100+ languages with released resources for reproducibility.

Abstract: Recent progress in universal multilingual named entity recognition (NER) has been driven by advances in multilingual transformer models and task-specific architectures, loss functions, and training datasets. Despite substantial prior work, we find that many critical design decisions for such models are made without systematic justification, with architectural components, training objectives, and data sources evaluated only in combination rather than in isolation. We argue that these decisions impede progress in the field by making it difficult to identify which choices improve model performance. In this work, we conduct extensive experiments around architectures, transformer backbones, training objectives, and data composition across a wide range of languages. Based on these insights, we introduce Otter, a universal multilingual NER model supporting over 100 languages. Otter achieves consistent improvements over strong multilingual NER baselines, outperforming GLiNER-x-base by 5.3pp in F1 and achieves competitive performance compared to large generative models such as Qwen3-32B, while being substantially more efficient. We release model checkpoints, training and evaluation code to facilitate reproducibility and future research.

Method: Constructed growing word-adjacency networks from Chinese and English literary works across different periods, treating punctuation marks as ordinary words. Analyzed average shortest path length L(N) vs network size N for different epochs, individual novels, and translations. Compared empirical results with a growing network model.

Result: Empirical results show satisfactory agreement with the growing network model. L(N) behaves asymptotically similar for both languages when punctuation marks are included, but becomes significantly larger for Chinese when punctuation marks are neglected.

Conclusion: Punctuation marks play a crucial role in network topology analysis of literary works, particularly affecting Chinese more than English, and should be treated as linguistic elements in network-based stylometric studies.

Abstract: Complex networks provide powerful tools for analyzing and understanding the intricate structures present in various systems, including natural language. Here, we analyze topology of growing word-adjacency networks constructed from Chinese and English literary works written in different periods. Unconventionally, instead of considering dictionary words only, we also include punctuation marks as if they were ordinary words. Our approach is based on two arguments: (1) punctuation carries genuine information related to emotional state, allows for logical grouping of content, provides a pause in reading, and facilitates understanding by avoiding ambiguity, and (2) our previous works have shown that punctuation marks behave like words in a Zipfian analysis and, if considered together with regular words, can improve authorship attribution in stylometric studies. We focus on a functional dependence of the average shortest path length $L(N)$ on a network size $N$ for different epochs and individual novels in their original language as well as for translations of selected novels into the other language. We approximate the empirical results with a growing network model and obtain satisfactory agreement between the two. We also observe that $L(N)$ behaves asymptotically similar for both languages if punctuation marks are included but becomes sizably larger for Chinese if punctuation marks are neglected.

Method: Analyzes examples from folktales where extraordinary objects possess language capabilities, examining how these narratives present diverse non-human entities with speech and intelligence.

Result: Folktales demonstrate that language capacity and intelligence are not inherently connected to humanness, offering diverse memorable examples of extraordinary objects with speech capabilities.

Conclusion: Folktales provide valuable inspiration and insight for designing speech and language interfaces that don’t rely on anthropomorphization, suggesting alternative approaches for human-technology interaction.

Abstract: Speech and language are valuable for interacting with technology. It would be ideal to be able to decouple their use from anthropomorphization, which has recently met an important moment of reckoning. In the world of folktales, language is everywhere and talking to extraordinary objects is not unusual. This overview presents examples of the analogies that folktales offer. Extraordinary objects in folktales are diverse and also memorable. Language capacity and intelligence are not always connected to humanness. Consideration of folktales can offer inspiration and insight for using speech and language for interacting with technology.

Method: Continued pre-training (CPT) on 26B tokens across 20 African languages using five base models (Llama 3.1, Gemma 3, Qwen 3). Systematic variation of data composition mixtures including math, code, and synthetic translated data, with comprehensive evaluation on multilingual benchmarks.

Result: Data composition is the primary driver of CPT gains. Adding math, code, and synthetic translated data yields consistent improvements, including on reasoning tasks. Architectural choices dominate scale when comparing across model families, and strong multilingual base performance doesn’t reliably predict post-CPT outcomes.

Conclusion: Robust architectures coupled with task-aligned data provide a more dependable recipe for multilingual adaptation than relying on base model multilingual capabilities. The best models improve long-context performance including document-level translation, and models are publicly released.

Abstract: Large language models (LLMs) are increasingly multilingual, yet open models continue to underperform relative to proprietary systems, with the gap most pronounced for African languages. Continued pre-training (CPT) offers a practical route to language adaptation, but improvements on demanding capabilities such as mathematical reasoning often remain limited. This limitation is driven in part by the uneven domain coverage and missing task-relevant knowledge that characterize many low-resource language corpora. We present \texttt{AfriqueLLM}, a suite of open LLMs adapted to 20 African languages through CPT on 26B tokens. We perform a comprehensive empirical study across five base models spanning sizes and architectures, including Llama 3.1, Gemma 3, and Qwen 3, and systematically analyze how CPT data composition shapes downstream performance. In particular, we vary mixtures that include math, code, and synthetic translated data, and evaluate the resulting models on a range of multilingual benchmarks. Our results identify data composition as the primary driver of CPT gains. Adding math, code, and synthetic translated data yields consistent improvements, including on reasoning-oriented evaluations. Within a fixed architecture, larger models typically improve performance, but architectural choices dominate scale when comparing across model families. Moreover, strong multilingual performance in the base model does not reliably predict post-CPT outcomes; robust architectures coupled with task-aligned data provide a more dependable recipe. Finally, our best models improve long-context performance, including document-level translation. Models have been released on Huggingface.

Method: Developed MITRA framework with MITRA-parallel pipeline for multilingual parallel passage mining, created a 1.74M parallel sentence corpus, and built domain-specific pretrained language models Gemma 2 MITRA (including MT and embedding variants).

Result: Gemma 2 MITRA-MT achieves state-of-the-art machine translation performance for Sanskrit, Chinese, and Tibetan into English, outperforming larger open-source models. Gemma 2 MITRA-E shows SOTA performance on semantic embedding benchmark. Resources made openly available.

Conclusion: The MITRA framework successfully addresses the challenge of analyzing massive multilingual Buddhist literature through automated parallel passage mining and domain-specific language models, providing valuable resources for both NLP research and philological studies.

Abstract: Ancient Buddhist literature features frequent, yet often unannotated, textual parallels spread across diverse languages: Sanskrit, Pāli, Buddhist Chinese, Tibetan, and more. The scale of this material makes manual examination prohibitive. We present the MITRA framework, which consists of a novel pipeline for multilingual parallel passage mining, MITRA-parallel, a large-scale corpus of 1.74 million parallel sentence pairs between Sanskrit, Chinese, and Tibetan, and the development of the domain-specific pretrained language model Gemma 2 MITRA. We present Gemma 2 MITRA-MT, a version of this base model fine-tuned on machine translation tasks, reaching state-of-the-art performance for machine translation of these languages into English and outperforming even much larger open-source models. We also present Gemma 2 MITRA-E, a semantic embedding model that shows state-of-the-art performance on a novel, detailed semantic embedding benchmark. We make the parallel dataset, model weights, and semantic similarity benchmark openly available to aid both NLP research and philological studies in Buddhist and classical Asian literature.

Method: Contrast logits from target and default system prompts to isolate behavioral signal unique to target persona, then amplify this signal by scalar factor alpha for continuous control over prompt adherence.

Result: Substantial improvements across five benchmarks: +8.5 strict accuracy on IFEval, +45 percentage points refusal rate on OffTopicEval, and +13% steerability on Prompt-Steering.

Conclusion: Enables practitioners to modulate system prompt strength for dynamic control over model behavior without retraining, providing continuous control over prompt adherence.

Abstract: Large language models excel at complex instructions yet struggle to deviate from their helpful assistant persona, as post-training instills strong priors that resist conflicting instructions. We introduce system prompt strength, a training-free method that treats prompt adherence as a continuous control. By contrasting logits from target and default system prompts, we isolate and amplify the behavioral signal unique to the target persona by a scalar factor alpha. Across five diverse benchmarks spanning constraint satisfaction, behavioral control, pluralistic alignment, capability modulation, and stylistic control, our method yields substantial improvements: up to +8.5 strict accuracy on IFEval, +45pp refusal rate on OffTopicEval, and +13% steerability on Prompt-Steering. Our approach enables practitioners to modulate system prompt strength, providing dynamic control over model behavior without retraining.

Method: Introduces a decision-theoretic framework using Value of Information (VoI) that enables agents to dynamically weigh expected utility gain from asking questions against the cognitive cost imposed on users. This inference-time method requires no hyperparameter tuning and adapts across contexts.

Result: Experiments across four diverse domains (20 Questions, medical diagnosis, flight booking, and e-commerce) show VoI consistently matches or exceeds best manually-tuned baselines, achieving up to 1.36 utility points higher in high-cost settings.

Conclusion: Provides a parameter-free framework for adaptive agent communication that explicitly balances task risk, query ambiguity, and user effort, enabling agents to make optimal decisions about when to seek clarification.

Abstract: Large Language Model (LLM) agents deployed for real-world tasks face a fundamental dilemma: user requests are underspecified, yet agents must decide whether to act on incomplete information or interrupt users for clarification. Existing approaches either rely on brittle confidence thresholds that require task-specific tuning, or fail to account for the varying stakes of different decisions. We introduce a decision-theoretic framework that resolves this trade-off through the Value of Information (VoI), enabling agents to dynamically weigh the expected utility gain from asking questions against the cognitive cost imposed on users. Our inference-time method requires no hyperparameter tuning and adapts seamlessly across contexts-from casual games to medical diagnosis. Experiments across four diverse domains (20 Questions, medical diagnosis, flight booking, and e-commerce) show that VoI consistently matches or exceeds the best manually-tuned baselines, achieving up to 1.36 utility points higher in high-cost settings. This work provides a parameter-free framework for adaptive agent communication that explicitly balances task risk, query ambiguity, and user effort.

Method: Proposes Structured Episodic Event Memory (SEEM) with two layers: graph memory for relational facts and dynamic episodic memory for narrative progression. Uses Episodic Event Frames (EEFs) with provenance pointers, plus associative fusion and Reverse Provenance Expansion (RPE) mechanisms.

Result: SEEM significantly outperforms baselines on LoCoMo and LongMemEval benchmarks, enabling agents to maintain superior narrative coherence and logical consistency.

Conclusion: SEEM addresses limitations of current memory approaches by providing a cognitive-inspired hierarchical framework that better organizes and reconstructs coherent narratives from fragmented evidence in long-term agent interactions.

Abstract: Current approaches to memory in Large Language Models (LLMs) predominantly rely on static Retrieval-Augmented Generation (RAG), which often results in scattered retrieval and fails to capture the structural dependencies required for complex reasoning. For autonomous agents, these passive and flat architectures lack the cognitive organization necessary to model the dynamic and associative nature of long-term interaction. To address this, we propose Structured Episodic Event Memory (SEEM), a hierarchical framework that synergizes a graph memory layer for relational facts with a dynamic episodic memory layer for narrative progression. Grounded in cognitive frame theory, SEEM transforms interaction streams into structured Episodic Event Frames (EEFs) anchored by precise provenance pointers. Furthermore, we introduce an agentic associative fusion and Reverse Provenance Expansion (RPE) mechanism to reconstruct coherent narrative contexts from fragmented evidence. Experimental results on the LoCoMo and LongMemEval benchmarks demonstrate that SEEM significantly outperforms baselines, enabling agents to maintain superior narrative coherence and logical consistency.

Method: Proposed method enables a language agent to give feedback to a vision-language model (VLM) to adapt text generation to the agent’s preferences, allowing the VLM to generate multimodal scene descriptions that help the LLM better understand multimodal context.

Result: LLM preference feedback significantly enhances VLM descriptions, achieving up to 13% absolute accuracy improvement over baseline multimodal approach. Human study shows 64.6% preference alignment between LLM choices and human judgments.

Conclusion: Unimodal LLMs can effectively provide feedback to optimize multimodal models, demonstrating a scalable approach for enhancing multimodal capabilities. Extensive experiments provide insights into how and why the method works and its limitations.

Abstract: To explore a more scalable path for adding multimodal capabilities to existing LLMs, this paper addresses a fundamental question: Can a unimodal LLM, relying solely on text, reason about its own informational needs and provide effective feedback to optimize a multimodal model? To answer this, we propose a method that enables a language agent to give feedback to a vision-language model (VLM) to adapt text generation to the agent’s preferences. Our results from different experiments affirm this hypothesis, showing that LLM preference feedback significantly enhances VLM descriptions. Using our proposed method, we find that the VLM can generate multimodal scene descriptions to help the LLM better understand multimodal context, leading to improvements of maximum 13% in absolute accuracy compared to the baseline multimodal approach. Furthermore, a human study validated our AI-driven feedback, showing a 64.6% preference alignment rate between the LLM’s choices and human judgments. Extensive experiments provide insights on how and why the method works and its limitations.

Method: Based on IBM Natural Conversation Framework (NCF), NC-Bench has three sets: 1) Basic Conversation Competence (fundamental sequence management like answering, repairing, closing), 2) RAG set (same patterns with retrieval-augmented generation), 3) Complex Request set (intricate sequence management patterns). Evaluates 14 interaction patterns across 6 open-source models.

Result: Models perform well on basic answering tasks, struggle with repair tasks (especially repeat), have mixed performance on closing sequences, and find complex multi-turn requests most challenging. Qwen models excel on Basic set, Granite models on RAG and Complex Request sets.

Conclusion: NC-Bench provides a lightweight, extensible, theory-grounded framework for assessing and improving LLMs’ conversational abilities by operationalizing fundamental principles of human conversation, moving beyond content-focused evaluation.

Abstract: The Natural Conversation Benchmark (NC-Bench) introduce a new approach to evaluating the general conversational competence of large language models (LLMs). Unlike prior benchmarks that focus on the content of model behavior, NC-Bench focuses on the form and structure of natural conversation. Grounded in the IBM Natural Conversation Framework (NCF), NC-Bench comprises three distinct sets. The Basic Conversation Competence set evaluates fundamental sequence management practices, such as answering inquiries, repairing responses, and closing conversational pairs. The RAG set applies the same sequence management patterns as the first set but incorporates retrieval-augmented generation (RAG). The Complex Request set extends the evaluation to complex requests involving more intricate sequence management patterns. Each benchmark tests a model’s ability to produce contextually appropriate conversational actions in response to characteristic interaction patterns. Initial evaluations across 6 open-source models and 14 interaction patterns show that models perform well on basic answering tasks, struggle more with repair tasks (especially repeat), have mixed performance on closing sequences, and find complex multi-turn requests most challenging, with Qwen models excelling on the Basic set and Granite models on the RAG set and the Complex Request set. By operationalizing fundamental principles of human conversation, NC-Bench provides a lightweight, extensible, and theory-grounded framework for assessing and improving the conversational abilities of LLMs beyond topical or task-specific benchmarks.

Method: Introduces Time Travel Engine (TTE) - an interpretability framework that projects diachronic linguistic patterns onto a shared chronological manifold, directly modulating latent representations to induce era-specific stylistic, lexical, and conceptual shifts.

Result: Temporal information is organized as continuous, traversable geometry rather than discrete clusters; TTE enables fluid navigation through period-specific “zeitgeists” while restricting future knowledge; Chinese and English LLMs show topological isomorphism in temporal subspaces, indicating universal geometric logic of historical evolution.

Conclusion: Bridges historical linguistics with mechanistic interpretability, offering a novel paradigm for controlling temporal reasoning in neural networks through geometric manipulation of latent temporal representations.

Abstract: Time functions as a fundamental dimension of human cognition, yet the mechanisms by which Large Language Models (LLMs) encode chronological progression remain opaque. We demonstrate that temporal information in their latent space is organized not as discrete clusters but as a continuous, traversable geometry. We introduce the Time Travel Engine (TTE), an interpretability-driven framework that projects diachronic linguistic patterns onto a shared chronological manifold. Unlike surface-level prompting, TTE directly modulates latent representations to induce coherent stylistic, lexical, and conceptual shifts aligned with target eras. By parameterizing diachronic evolution as a continuous manifold within the residual stream, TTE enables fluid navigation through period-specific “zeitgeists” while restricting access to future knowledge. Furthermore, experiments across diverse architectures reveal topological isomorphism between the temporal subspaces of Chinese and English-indicating that distinct languages share a universal geometric logic of historical evolution. These findings bridge historical linguistics with mechanistic interpretability, offering a novel paradigm for controlling temporal reasoning in neural networks.

Method: Propose VISTA Space - a high-dimensional representational framework for narrative orchestration that unifies human and model perspectives. Introduce LitVISTA benchmark with structural annotations from literary texts for systematic evaluation.

Result: Oracle evaluations on GPT, Claude, Grok, and Gemini reveal systematic deficiencies: models fail to construct unified global narrative views and struggle to jointly capture narrative function and structure. Advanced thinking modes provide only limited gains.

Conclusion: Existing LLMs have fundamental limitations in narrative orchestration despite strong causal coherence capabilities, highlighting the need for frameworks like VISTA Space to bridge the gap between human and model narrative understanding.

Abstract: Computational narrative analysis aims to capture rhythm, tension, and emotional dynamics in literary texts. Existing large language models can generate long stories but overly focus on causal coherence, neglecting the complex story arcs and orchestration inherent in human narratives. This creates a structural misalignment between model- and human-generated narratives. We propose VISTA Space, a high-dimensional representational framework for narrative orchestration that unifies human and model narrative perspectives. We further introduce LitVISTA, a structurally annotated benchmark grounded in literary texts, enabling systematic evaluation of models’ narrative orchestration capabilities. We conduct oracle evaluations on a diverse selection of frontier LLMs, including GPT, Claude, Grok, and Gemini. Results reveal systematic deficiencies: existing models fail to construct a unified global narrative view, struggling to jointly capture narrative function and structure. Furthermore, even advanced thinking modes yield only limited gains for such literary narrative understanding.

Method: PRISP uses a Text-to-LoRA hypernetwork to generate task-aware LoRA parameters from task descriptions, then optimizes only a small subset of these parameters plus minimal additional modules using few-shot user data.

Result: Experiments on a few-shot variant of LaMP benchmark show PRISP achieves strong overall performance compared to prior approaches while reducing computational overhead and eliminating privacy risks.

Conclusion: PRISP provides an effective solution for practical LLM personalization under realistic constraints of limited data, computational resources, and privacy requirements.

Abstract: Large language model (LLM) personalization aims to adapt general-purpose models to individual users. Most existing methods, however, are developed under data-rich and resource-abundant settings, often incurring privacy risks. In contrast, realistic personalization typically occurs after deployment under (i) extremely limited user data, (ii) constrained computational resources, and (iii) strict privacy requirements. We propose PRISP, a lightweight and privacy-safe personalization framework tailored to these constraints. PRISP leverages a Text-to-LoRA hypernetwork to generate task-aware LoRA parameters from task descriptions, and enables efficient user personalization by optimizing a small subset of task-aware LoRA parameters together with minimal additional modules using few-shot user data. Experiments on a few-shot variant of the LaMP benchmark demonstrate that PRISP achieves strong overall performance compared to prior approaches, while reducing computational overhead and eliminating privacy risks.

Method: Created IndRegBias dataset with 25,000 comments from Reddit and YouTube discussing Indian regional issues. Developed multilevel annotation strategy for bias severity. Evaluated open-source LLMs and Indic Language Models using zero-shot, few-shot, and fine-tuning strategies for bias detection and severity classification.

Result: Zero-shot and few-shot approaches showed lower accuracy in detecting regional biases and severity across most LLMs and ILMs. However, fine-tuning significantly enhanced LLM performance in detecting Indian regional bias and its severity levels.

Conclusion: The paper successfully creates a valuable dataset for studying Indian regional biases and demonstrates that fine-tuning is crucial for effective regional bias detection in language models, addressing an important gap in bias research.

Abstract: Warning: This paper consists of examples representing regional biases in Indian regions that might be offensive towards a particular region. While social biases corresponding to gender, race, socio-economic conditions, etc., have been extensively studied in the major applications of Natural Language Processing (NLP), biases corresponding to regions have garnered less attention. This is mainly because of (i) difficulty in the extraction of regional bias datasets, (ii) disagreements in annotation due to inherent human biases, and (iii) regional biases being studied in combination with other types of social biases and often being under-represented. This paper focuses on creating a dataset IndRegBias, consisting of regional biases in an Indian context reflected in users’ comments on popular social media platforms, namely Reddit and YouTube. We carefully selected 25,000 comments appearing on various threads in Reddit and videos on YouTube discussing trending topics on regional issues in India. Furthermore, we propose a multilevel annotation strategy to annotate the comments describing the severity of regional biased statements. To detect the presence of regional bias and its severity in IndRegBias, we evaluate open-source Large Language Models (LLMs) and Indic Language Models (ILMs) using zero-shot, few-shot, and fine-tuning strategies. We observe that zero-shot and few-shot approaches show lower accuracy in detecting regional biases and severity in the majority of the LLMs and ILMs. However, the fine-tuning approach significantly enhances the performance of the LLM in detecting Indian regional bias along with its severity.

Method: A tool-augmented vision-language agent with interleaved multimodal chain-of-thought reasoning, trained via two-stage post-training: cold-start supervised fine-tuning on expert inspection trajectories followed by outcome-based reinforcement learning on rare-type verification tasks.

Result: Achieves state-of-the-art performance on five rare-object identification tasks from LAMOST, boosting macro-F1 from 28.3 to 76.5 with a 7B parameter model, outperforming proprietary VLMs and specialized deep models, and demonstrates strong generalization to unseen tasks across survey shifts.

Conclusion: Spec-o3 successfully bridges the gap between automated classification and expert inspection, providing transparent, physically consistent reasoning that supports trustworthy decision-making for rare celestial object identification at scale.

Abstract: Due to the limited generalization and interpretability of deep learning classifiers, The final vetting of rare celestial object candidates still relies on expert visual inspection–a manually intensive process. In this process, astronomers leverage specialized tools to analyze spectra and construct reliable catalogs. However, this practice has become the primary bottleneck, as it is fundamentally incapable of scaling with the data deluge from modern spectroscopic surveys. To bridge this gap, we propose Spec-o3, a tool-augmented vision-language agent that performs astronomer-aligned spectral inspection via interleaved multimodal chain-of-thought reasoning. Spec-o3 is trained with a two-stage post-training recipe: cold-start supervised fine-tuning on expert inspection trajectories followed by outcome-based reinforcement learning on rare-type verification tasks. Evaluated on five rare-object identification tasks from LAMOST, Spec-o3 establishes a new State-of-the-Art, boosting the macro-F1 score from 28.3 to 76.5 with a 7B parameter base model and outperforming both proprietary VLMs and specialized deep models. Crucially, the agent demonstrates strong generalization to unseen inspection tasks across survey shifts (from LAMOST to SDSS/DESI). Expert evaluations confirm that its reasoning traces are coherent and physically consistent, supporting transparent and trustworthy decision-making. Code, data, and models are available at \href{https://github.com/Maxwell-Jia/spec-o3}{Project HomePage}.

Method: Decomposes generated answers into atomic claims, estimates claim support by combining evidence-grounded natural language inference with biomedical knowledge-graph consistency signals, aggregates claim decisions for answer-level diagnostics, and distills the pipeline into compact biomedical models with ensemble verification and class-specific reliability weighting.

Result: Experiments on four biomedical QA benchmarks show MedRAGChecker reliably flags unsupported and contradicted claims and reveals distinct risk profiles across different generators, particularly on safety-critical biomedical relations.

Conclusion: MedRAGChecker provides an effective framework for claim-level verification in biomedical RAG systems, enabling scalable evaluation and identification of safety-critical errors through its diagnostic capabilities.

Abstract: Biomedical retrieval-augmented generation (RAG) can ground LLM answers in medical literature, yet long-form outputs often contain isolated unsupported or contradictory claims with safety implications. We introduce MedRAGChecker, a claim-level verification and diagnostic framework for biomedical RAG. Given a question, retrieved evidence, and a generated answer, MedRAGChecker decomposes the answer into atomic claims and estimates claim support by combining evidence-grounded natural language inference (NLI) with biomedical knowledge-graph (KG) consistency signals. Aggregating claim decisions yields answer-level diagnostics that help disentangle retrieval and generation failures, including faithfulness, under-evidence, contradiction, and safety-critical error rates. To enable scalable evaluation, we distill the pipeline into compact biomedical models and use an ensemble verifier with class-specific reliability weighting. Experiments on four biomedical QA benchmarks show that MedRAGChecker reliably flags unsupported and contradicted claims and reveals distinct risk profiles across generators, particularly on safety-critical biomedical relations.

Method: Introduce Atomic-SNLI, a novel dataset constructed by decomposing SNLI and enriching it with carefully curated atomic-level examples through linguistically informed generation strategies. Models are fine-tuned on this dataset to improve atomic reasoning capabilities.

Result: Models fine-tuned on Atomic-SNLI achieve significant improvements in atomic reasoning capabilities while maintaining strong sentence-level performance. This enables both accurate judgments and transparent, explainable results at the fact level.

Conclusion: Atomic-SNLI successfully addresses the limitation of poor atomic-level reasoning in NLI models, enabling explainable inference at the fact level while preserving sentence-level accuracy, thus bridging the gap between fine-grained reasoning and overall NLI performance.

Abstract: Current Natural Language Inference (NLI) systems primarily operate at the sentence level, providing black-box decisions that lack explanatory power. While atomic-level NLI offers a promising alternative by decomposing hypotheses into individual facts, we demonstrate that the conventional assumption that a hypothesis is entailed only when all its atomic facts are entailed fails in practice due to models’ poor performance on fine-grained reasoning. Our analysis reveals that existing models perform substantially worse on atomic level inference compared to sentence level tasks. To address this limitation, we introduce Atomic-SNLI, a novel dataset constructed by decomposing SNLI and enriching it with carefully curated atomic level examples through linguistically informed generation strategies. Experimental results demonstrate that models fine-tuned on Atomic-SNLI achieve significant improvements in atomic reasoning capabilities while maintaining strong sentence level performance, enabling both accurate judgements and transparent, explainable results at the fact level.

Method: Created Exposía dataset from a Computer Science “Introduction to Scientific Work” course, containing student research proposals, peer/instructor feedback with comments and reviews, and human assessment scores based on a pedagogically-grounded schema. Used this dataset to benchmark open-source LLMs on two tasks: automated scoring of proposals and student reviews.

Result: LLMs achieve high agreement on scoring aspects requiring little domain knowledge but degrade on content evaluation dimensions, matching human agreement patterns. LLMs align better with instructors giving high scores. A prompting strategy scoring multiple writing aspects together proved most effective for classroom deployment.

Conclusion: Exposía enables research on educationally grounded academic writing evaluation. LLMs show promise for automated scoring but have limitations in content evaluation, with multi-aspect prompting being most effective. The dataset supports development of better automated assessment tools for academic writing education.

Abstract: We present Exposía, the first public dataset that connects writing and feedback assessment in higher education, enabling research on educationally grounded approaches to academic writing evaluation. Exposía includes student research project proposals and peer and instructor feedback consisting of comments and free-text reviews. The dataset was collected in the “Introduction to Scientific Work” course of the Computer Science undergraduate program that focuses on teaching academic writing skills and providing peer feedback on academic writing. Exposía reflects the multi-stage nature of the academic writing process that includes drafting, providing and receiving feedback, and revising the writing based on the feedback received. Both the project proposals and peer feedback are accompanied by human assessment scores based on a fine-grained, pedagogically-grounded schema for writing and feedback assessment that we develop. We use Exposía to benchmark state-of-the-art open-source large language models (LLMs) for two tasks: automated scoring of (1) the proposals and (2) the student reviews. The strongest LLMs attain high agreement on scoring aspects that require little domain knowledge but degrade on dimensions evaluating content, in line with human agreement values. We find that LLMs align better with the human instructors giving high scores. Finally, we establish that a prompting strategy that scores multiple aspects of the writing together is the most effective, an important finding for classroom deployment.

Method: Multi-stage fine-tuning framework: two stages of supervised fine-tuning (SFT) followed by direct preference optimization (DPO) on curated SimPy queueing data to enhance code generation quality.

Result: Both fine-tuned models show substantial improvements in executability, output-format compliance, and instruction-code consistency, making them reliable SimPy simulation generators.

Conclusion: Domain-specific fine-tuning can transform compact open-source code models into practical alternatives to closed-source LLMs for SimPy queueing simulations in education, research, and decision support.

Abstract: The Python package SimPy is widely used for modeling queueing systems due to its flexibility, simplicity, and smooth integration with modern data analysis and optimization frameworks. Recent advances in large language models (LLMs) have shown strong ability in generating clear and executable code, making them powerful and suitable tools for writing SimPy queueing simulation code. However, directly employing closed-source models like GPT-4o to generate such code may lead to high computational costs and raise data privacy concerns. To address this, we fine-tune two open-source LLMs, Qwen-Coder-7B and DeepSeek-Coder-6.7B, on curated SimPy queueing data, which enhances their code-generating performance in executability, output-format compliance, and instruction-code consistency. Particularly, we proposed a multi-stage fine-tuning framework comprising two stages of supervised fine-tuning (SFT) and one stage of direct preference optimization (DPO), progressively enhancing the model’s ability in SimPy-based queueing simulation code generation. Extensive evaluations demonstrate that both fine-tuned models achieve substantial improvements in executability, output-format compliance, and instruct consistency. These results confirm that domain-specific fine-tuning can effectively transform compact open-source code models into reliable SimPy simulation generators which provide a practical alternative to closed-source LLMs for education, research, and operational decision support.

Method: CSR-RAG (Contextual, Structural, and Relational Retrieval Augmented Generation) - a hybrid RAG system that combines three types of retrieval: contextual (semantic understanding), structural (schema organization), and relational (table relationships) to efficiently identify relevant database tables.

Result: Achieves up to 40% precision and over 80% recall with negligible average query generation latency of only 30ms on commodity hardware, making it suitable for enterprise-scale LLM systems.

Conclusion: CSR-RAG provides an effective solution for enterprise-scale Text-to-SQL by enabling efficient table retrieval before SQL generation, balancing accuracy with computational efficiency for practical deployment.

Abstract: Natural language to SQL translation (Text-to-SQL) is one of the long-standing problems that has recently benefited from advances in Large Language Models (LLMs). While most academic Text-to-SQL benchmarks request schema description as a part of natural language input, enterprise-scale applications often require table retrieval before SQL query generation. To address this need, we propose a novel hybrid Retrieval Augmented Generation (RAG) system consisting of contextual, structural, and relational retrieval (CSR-RAG) to achieve computationally efficient yet sufficiently accurate retrieval for enterprise-scale databases. Through extensive enterprise benchmarks, we demonstrate that CSR-RAG achieves up to 40% precision and over 80% recall while incurring a negligible average query generation latency of only 30ms on commodity data center hardware, which makes it appropriate for modern LLM-based enterprise-scale systems.

Method: Dynamic evaluation using template pools for instructions, predefined intervals for numeric parameters, and validators to verify outcomes. Contains 107 tasks (62 atomic, 45 composite) with modular architecture for rapid development. Executes scripts on forked EVM chain with snapshot isolation; composite tasks apply step-efficiency decay.

Result: Evaluation of 20 models reveals large performance gaps, with split scores showing persistent asymmetry between single-action precision and multi-step workflow completion.

Conclusion: EVM-QuestBench addresses critical gaps in LLM evaluation for blockchain transaction scenarios, highlighting the need for better models that can handle both atomic operations and complex multi-step workflows safely and accurately.

Abstract: Large language models are increasingly applied to various development scenarios. However, in on-chain transaction scenarios, even a minor error can cause irreversible loss for users. Existing evaluations often overlook execution accuracy and safety. We introduce EVM-QuestBench, an execution-grounded benchmark for natural-language transaction-script generation on EVM-compatible chains. The benchmark employs dynamic evaluation: instructions are sampled from template pools, numeric parameters are drawn from predefined intervals, and validators verify outcomes against these instantiated values. EVM-QuestBench contains 107 tasks (62 atomic, 45 composite). Its modular architecture enables rapid task development. The runner executes scripts on a forked EVM chain with snapshot isolation; composite tasks apply step-efficiency decay. We evaluate 20 models and find large performance gaps, with split scores revealing persistent asymmetry between single-action precision and multi-step workflow completion. Code: https://anonymous.4open.science/r/bsc_quest_bench-A9CF/.

Method: Proposes using contrastive learning on a hypersphere to induce circular emotion representations within language model embeddings, aligning with psychological circumplex geometry.

Result: Circular alignment offers superior interpretability and robustness against dimensionality reduction, but underperforms compared to conventional designs in high-dimensional settings and fine-grained classification tasks.

Conclusion: The findings elucidate trade-offs involved in applying psychological circumplex models to deep learning architectures, highlighting the balance between interpretability/robustness and performance in different settings.

Abstract: Psychological research has long utilized circumplex models to structure emotions, placing similar emotions adjacently and opposing ones diagonally. Although frequently used to interpret deep learning representations, these models are rarely directly incorporated into the representation learning of language models, leaving their geometric validity unexplored. This paper proposes a method to induce circular emotion representations within language model embeddings via contrastive learning on a hypersphere. We show that while this circular alignment offers superior interpretability and robustness against dimensionality reduction, it underperforms compared to conventional designs in high-dimensional settings and fine-grained classification. Our findings elucidate the trade-offs involved in applying psychological circumplex models to deep learning architectures.

Method: Proposed a stylistic similarity framework comparing lexico-structural, pragmatic, psycholinguistic, and encoder-derived features across years of informal digital text messages to quantify temporal variation in Singlish.

Result: Found significant diachronic changes in tone, expressivity, and sentence construction over the decade. LLMs could generate superficially realistic Singlish but failed to produce temporally neutral outputs, with residual temporal signals remaining detectable despite prompting and fine-tuning.

Conclusion: Singlish shows dynamic evolution over time, and current LLMs have limitations in modeling sociolectal and temporal variations in colloquial languages, highlighting both their capabilities and constraints.

Abstract: Singlish is a creole rooted in Singapore’s multilingual environment and continues to evolve alongside social and technological change. This study investigates the evolution of Singlish over a decade of informal digital text messages. We propose a stylistic similarity framework that compares lexico-structural, pragmatic, psycholinguistic, and encoder-derived features across years to quantify temporal variation. Our analysis reveals notable diachronic changes in tone, expressivity and sentence construction over the years. Conversely, while some LLMs were able to generate superficially realistic Singlish messages, they do not produce temporally neutral outputs, and residual temporal signals remain detectable despite prompting and fine-tuning. Our findings highlight the dynamic evolution of Singlish, as well as the capabilities and limitations of current LLMs in modeling sociolectal and temporal variations in the colloquial language.

Method: Train a classifier to determine if text is authored by an LLM or human, deployed on an online CPU-based platform, with three key innovations: no reliance on auxiliary information like watermarks or specific LLM knowledge, improved human-LLM distinction, and enabling statistical inference.

Result: The classifier achieves higher classification accuracy than existing detectors while maintaining type-I error control, high statistical power, and computational efficiency.

Conclusion: The proposed detector effectively addresses practical concerns about LLM-generated text by providing a reliable, statistically sound, and computationally efficient solution that outperforms existing methods.

Abstract: Large language models (LLMs) such as GPT, Claude, Gemini, and Grok have been deeply integrated into our daily life. They now support a wide range of tasks – from dialogue and email drafting to assisting with teaching and coding, serving as search engines, and much more. However, their ability to produce highly human-like text raises serious concerns, including the spread of fake news, the generation of misleading governmental reports, and academic misconduct. To address this practical problem, we train a classifier to determine whether a piece of text is authored by an LLM or a human. Our detector is deployed on an online CPU-based platform https://huggingface.co/spaces/stats-powered-ai/StatDetectLLM, and contains three novelties over existing detectors: (i) it does not rely on auxiliary information, such as watermarks or knowledge of the specific LLM used to generate the text; (ii) it more effectively distinguishes between human- and LLM-authored text; and (iii) it enables statistical inference, which is largely absent in the current literature. Empirically, our classifier achieves higher classification accuracy compared to existing detectors, while maintaining type-I error control, high statistical power, and computational efficiency.

Method: The researchers measure two key geometric properties: (1) the directional change (θ) between truth vectors with and without context, and (2) the relative magnitude of truth vectors upon adding context. They conduct experiments across four LLMs and four datasets.

Result: Three main findings: (1) Truth vectors are orthogonal in early layers, converge in middle layers, and stabilize/increase in later layers; (2) Adding context generally increases truth vector magnitude, amplifying separation between true/false representations; (3) Larger models distinguish relevant from irrelevant context through directional changes, while smaller models use magnitude differences. Context conflicting with parametric knowledge produces larger geometric changes.

Conclusion: This is the first work to provide a geometric characterization of how context transforms truth vectors in LLM activation space, revealing systematic patterns across model layers and sizes, with implications for understanding how LLMs process contextual information.

Abstract: Large Language Models (LLMs) often encode whether a statement is true as a vector in their residual stream activations. These vectors, also known as truth vectors, have been studied in prior work, however how they change when context is introduced remains unexplored. We study this question by measuring (1) the directional change ($θ$) between the truth vectors with and without context and (2) the relative magnitude of the truth vectors upon adding context. Across four LLMs and four datasets, we find that (1) truth vectors are roughly orthogonal in early layers, converge in middle layers, and may stabilize or continue increasing in later layers; (2) adding context generally increases the truth vector magnitude, i.e., the separation between true and false representations in the activation space is amplified; (3) larger models distinguish relevant from irrelevant context mainly through directional change ($θ$), while smaller models show this distinction through magnitude differences. We also find that context conflicting with parametric knowledge produces larger geometric changes than parametrically aligned context. To the best of our knowledge, this is the first work that provides a geometric characterization of how context transforms the truth vector in the activation space of LLMs.

Method: Created a comprehensive evaluation framework using a manually annotated dataset of 200 short videos across four health domains with fine-grained annotations for three deceptive patterns. Evaluated eight frontier MLLMs across five modality settings.

Result: Gemini-2.5-Pro achieved highest performance in multimodal setting (71.5/100 belief score), while o3 performed worst (35.2). Models are susceptible to biases like authoritative channel IDs that induce false beliefs.

Conclusion: MLLMs show significant performance gaps in detecting misinformation in short videos and remain vulnerable to social cues and cognitive biases, highlighting the need for more robust multimodal misinformation detection systems.

Abstract: Short-video platforms have become major channels for misinformation, where deceptive claims frequently leverage visual experiments and social cues. While Multimodal Large Language Models (MLLMs) have demonstrated impressive reasoning capabilities, their robustness against misinformation entangled with cognitive biases remains under-explored. In this paper, we introduce a comprehensive evaluation framework using a high-quality, manually annotated dataset of 200 short videos spanning four health domains. This dataset provides fine-grained annotations for three deceptive patterns, experimental errors, logical fallacies, and fabricated claims, each verified by evidence such as national standards and academic literature. We evaluate eight frontier MLLMs across five modality settings. Experimental results demonstrate that Gemini-2.5-Pro achieves the highest performance in the multimodal setting with a belief score of 71.5/100, while o3 performs the worst at 35.2. Furthermore, we investigate social cues that induce false beliefs in videos and find that models are susceptible to biases like authoritative channel IDs.

Method: N2N-GQA constructs dynamic evidence graphs from noisy retrieval outputs by modeling documents as graph nodes with semantic relationships as edges. This allows identification of bridge documents connecting reasoning steps, which is absent in list-based retrieval approaches.

Result: On OTT-QA, graph-based evidence curation provides a 19.9-point EM improvement over strong baselines. N2N-GQA achieves 48.80 EM, matching fine-tuned retrieval models (CORE: 49.0 EM) and approaching heavily optimized systems (COS: 56.9 EM) without any task-specific training.

Conclusion: Graph-structured evidence organization is essential for scalable, zero-shot multi-hop QA systems. Simple, interpretable graph construction can rival sophisticated fine-tuned approaches, demonstrating the critical importance of organizing retrieval results as structured graphs for multi-hop reasoning.

Abstract: Multi-hop question answering over hybrid table-text data requires retrieving and reasoning across multiple evidence pieces from large corpora, but standard Retrieval-Augmented Generation (RAG) pipelines process documents as flat ranked lists, causing retrieval noise to obscure reasoning chains. We introduce N2N-GQA. To our knowledge, it is the first zeroshot framework for open-domain hybrid table-text QA that constructs dynamic evidence graphs from noisy retrieval outputs. Our key insight is that multi-hop reasoning requires understanding relationships between evidence pieces: by modeling documents as graph nodes with semantic relationships as edges, we identify bridge documents connecting reasoning steps, a capability absent in list-based retrieval. On OTT-QA, graph-based evidence curation provides a 19.9-point EM improvement over strong baselines, demonstrating that organizing retrieval results as structured graphs is critical for multihop reasoning. N2N-GQA achieves 48.80 EM, matching finetuned retrieval models (CORE: 49.0 EM) and approaching heavily optimized systems (COS: 56.9 EM) without any task specific training. This establishes graph-structured evidence organization as essential for scalable, zero-shot multi-hop QA systems and demonstrates that simple, interpretable graph construction can rival sophisticated fine-tuned approaches.

Method: Curated dataset of 200 verses with thematic tags, uses IndicBERT sentence embeddings for semantic retrieval, then Mistral LLM for generating transliterations, translations, and contextual explanations.

Result: Semantic retrieval significantly outperforms keyword matching in precision and relevance; user studies show improved accessibility through generated summaries.

Conclusion: First attempt at integrating retrieval and generation for Sanskrit Subhasitas, bridging cultural heritage with modern applied AI.

Abstract: Sanskrit Subhasitas encapsulate centuries of cultural and philosophical wisdom, yet remain underutilized in the digital age due to linguistic and contextual barriers. In this work, we present Pragya, a retrieval-augmented generation (RAG) framework for semantic recommendation of Subhasitas. We curate a dataset of 200 verses annotated with thematic tags such as motivation, friendship, and compassion. Using sentence embeddings (IndicBERT), the system retrieves top-k verses relevant to user queries. The retrieved results are then passed to a generative model (Mistral LLM) to produce transliterations, translations, and contextual explanations. Experimental evaluation demonstrates that semantic retrieval significantly outperforms keyword matching in precision and relevance, while user studies highlight improved accessibility through generated summaries. To our knowledge, this is the first attempt at integrating retrieval and generation for Sanskrit Subhasitas, bridging cultural heritage with modern applied AI.

Method: Frames NEL accuracy estimation as constrained optimization to minimize annotation cost subject to target Margin of Error. Adapts Stratified Two-Stage Cluster Sampling (STWCS) to NEL setting with label-based strata and global surface-form clusters independent of NEL annotations.

Result: Applied to 11,184 NEL annotations in GutBrainIE corpus, achieved MoE ≤ 0.05 by annotating only 2,749 triples (24.6%), estimating overall accuracy at 0.915 ± 0.0473. Reduced expert annotation time by about 29% compared to Simple Random Sampling baseline.

Conclusion: The framework provides scalable, statistically robust accuracy assessment for NEL benchmarks and IE pipelines, enabling quality evaluation with constrained annotation budgets while maintaining statistical guarantees.

Abstract: Named Entity Linking (NEL) is a core component of biomedical Information Extraction (IE) pipelines, yet assessing its quality at scale is challenging due to the high cost of expert annotations and the large size of corpora. In this paper, we present a sampling-based framework to estimate the NEL accuracy of large-scale IE corpora under statistical guarantees and constrained annotation budgets. We frame NEL accuracy estimation as a constrained optimization problem, where the objective is to minimize expected annotation cost subject to a target Margin of Error (MoE) for the corpus-level accuracy estimate. Building on recent works on knowledge graph accuracy estimation, we adapt Stratified Two-Stage Cluster Sampling (STWCS) to the NEL setting, defining label-based strata and global surface-form clusters in a way that is independent of NEL annotations. Applied to 11,184 NEL annotations in GutBrainIE – a new biomedical corpus openly released in fall 2025 – our framework reaches a MoE $\leq 0.05$ by manually annotating only 2,749 triples (24.6%), leading to an overall accuracy estimate of $0.915 \pm 0.0473$. A time-based cost model and simulations against a Simple Random Sampling (SRS) baseline show that our design reduces expert annotation time by about 29% at fixed sample size. The framework is generic and can be applied to other NEL benchmarks and IE pipelines that require scalable and statistically robust accuracy assessment.

Method: MultiCalibrated Subjective Task Learner (MC-STL) clusters annotations using three approaches: similarity of annotator rationales, expert-value taxonomies, or rater’s sociocultural descriptors, then learns cluster-specific embeddings for calibration.

Result: MC-STL consistently outperforms baselines across multiple datasets (toxic chatbot conversations, offensive social media posts, human preference alignment) in discrimination, value-specific calibration, and disagreement-aware metrics.

Conclusion: Accounting for latent value structure in annotations through clustering and cluster-specific calibration significantly improves performance on subjective NLP tasks across various learning settings.

Abstract: Building NLP systems for subjective tasks requires one to ensure their alignment to contrasting human values. We propose the MultiCalibrated Subjective Task Learner framework (MC-STL), which clusters annotations into identifiable human value clusters by three approaches (similarity of annotator rationales, expert-value taxonomies or rater’s sociocultural descriptors) and calibrates predictions for each value cluster by learning cluster-specific embeddings. We demonstrate MC-STL on several subjective learning settings, including ordinal, binary, and preference learning predictions, and evaluate it on multiple datasets covering toxic chatbot conversations, offensive social media posts, and human preference alignment. The results show that MC-STL consistently outperforms the baselines that ignore the latent value structure of the annotations, delivering gains in discrimination, value-specific calibration, and disagreement-aware metrics.

Method: 1) Introduce MedEinst benchmark with 5,383 paired clinical cases across 49 diseases (control case + “trap” case with altered discriminative evidence). 2) Propose ECR-Agent with two components: Dynamic Causal Inference (DCI) for structured reasoning through dual-pathway perception, dynamic causal graph reasoning, and evidence audit; and Critic-Driven Graph and Memory Evolution (CGME) for iterative refinement via exemplar base and illness graphs.

Result: Extensive evaluation of 17 LLMs shows frontier models achieve high baseline accuracy but severe bias trap rates (probability of misdiagnosing traps despite correctly diagnosing controls).

Conclusion: Current LLMs exhibit dangerous Einstellung Effect in clinical diagnosis. The proposed ECR-Agent framework effectively aligns LLM reasoning with Evidence-Based Medicine standards to address this critical failure mode.

Abstract: Despite achieving high accuracy on medical benchmarks, LLMs exhibit the Einstellung Effect in clinical diagnosis–relying on statistical shortcuts rather than patient-specific evidence, causing misdiagnosis in atypical cases. Existing benchmarks fail to detect this critical failure mode. We introduce MedEinst, a counterfactual benchmark with 5,383 paired clinical cases across 49 diseases. Each pair contains a control case and a “trap” case with altered discriminative evidence that flips the diagnosis. We measure susceptibility via Bias Trap Rate–probability of misdiagnosing traps despite correctly diagnosing controls. Extensive Evaluation of 17 LLMs shows frontier models achieve high baseline accuracy but severe bias trap rates. Thus, we propose ECR-Agent, aligning LLM reasoning with Evidence-Based Medicine standard via two components: (1) Dynamic Causal Inference (DCI) performs structured reasoning through dual-pathway perception, dynamic causal graph reasoning across three levels (association, intervention, counterfactual), and evidence audit for final diagnosis; (2) Critic-Driven Graph and Memory Evolution (CGME) iteratively refines the system by storing validated reasoning paths in an exemplar base and consolidating disease-specific knowledge into evolving illness graphs. Source code is to be released.

Method: SpikeATE architecture uses ternary spiking neurons and direct spike training fine-tuned with pseudo-gradients. It leverages SNNs’ sparse activations and event-driven inference to capture temporal dependencies between words for ATE as sequence labeling.

Result: Evaluated on four benchmark SemEval datasets, SpikeATE achieves performance comparable to state-of-the-art DNNs while demonstrating significantly lower energy consumption.

Conclusion: SNNs present a practical and sustainable alternative to DNNs for Aspect Term Extraction tasks, offering comparable accuracy with substantially reduced energy requirements.

Abstract: Aspect Term Extraction (ATE) identifies aspect terms in review sentences, a key subtask of sentiment analysis. While most existing approaches use energy-intensive deep neural networks (DNNs) for ATE as sequence labeling, this paper proposes a more energy-efficient alternative using Spiking Neural Networks (SNNs). Using sparse activations and event-driven inferences, SNNs capture temporal dependencies between words, making them suitable for ATE. The proposed architecture, SpikeATE, employs ternary spiking neurons and direct spike training fine-tuned with pseudo-gradients. Evaluated on four benchmark SemEval datasets, SpikeATE achieves performance comparable to state-of-the-art DNNs with significantly lower energy consumption. This highlights the use of SNNs as a practical and sustainable choice for ATE tasks.

Method: Introduces a two-hop question answering setting requiring inference over two multilingual documents, evaluates model sensitivity to language variation, analyzes reasoning decomposition failures, and proposes SUBQ prompting to guide multi-step reasoning with sub-questions.

Result: Models are more sensitive to language variation in answer-span documents than bridging documents; up to 33% of multilingual cases show failed bridging inference but correct final answers; 18% composition failure occurs; SUBQ prompting boosts accuracy from 10.1% to 66.5%.

Conclusion: Language models lack faithful step-by-step reasoning in multilingual settings, leading to composition failures, but explicit decomposition through sub-question prompting can dramatically improve performance in multilingual multi-hop QA.

Abstract: The real-world information sources are inherently multilingual, which naturally raises a question about whether language models can synthesize information across languages. In this paper, we introduce a simple two-hop question answering setting, where answering a question requires making inferences over two multilingual documents. We find that language models are more sensitive to language variation in answer-span documents than in those providing bridging information, despite the equal importance of both documents for answering a question. Under a step-by-step sub-question evaluation, we further show that in up to 33% of multilingual cases, models fail to infer the bridging information in the first step yet still answer the overall question correctly. This indicates that reasoning in language models, especially in multilingual settings, does not follow a faithful step-by-step decomposition. Subsequently, we show that the absence of reasoning decomposition leads to around 18% composition failure, where both sub-questions are answered correctly but fail for the final two-hop questions. To mitigate this, we propose a simple three-stage SUBQ prompting method to guide the multi-step reasoning with sub-questions, which boosts accuracy from 10.1% to 66.5%.

Method: Analyzed 2,858 films/series from Movie Scripts Corpus; identified protagonists via AI-assisted annotation; quantified warmth and competence using LLM_annotate package (gpt-4.1-mini); conducted preregistered Bayesian regression analyses to examine associations with IMDb ratings and genre interactions.

Result: Small but theory-consistent associations between both warmth and competence and audience ratings; genre interactions didn’t meaningfully improve predictions; male protagonists were slightly less warm than female protagonists; male-led movies received higher ratings overall (stronger effect than warmth/competence relationships).

Conclusion: Audiences favor warm, competent characters but effects on movie ratings are modest; character personality is only one of many factors shaping evaluations; AI-assisted annotation with LLMs is effective for large-scale analysis but occasionally falls short of manual annotation quality.

Abstract: Drawing on psychological and literary theory, we investigated whether the warmth and competence of movie protagonists predict IMDb ratings, and whether these effects vary across genres. Using 2,858 films and series from the Movie Scripts Corpus, we identified protagonists via AI-assisted annotation and quantified their warmth and competence with the LLM_annotate package ([1]; human-LLM agreement: r = .83). Preregistered Bayesian regression analyses revealed theory-consistent but small associations between both warmth and competence and audience ratings, while genre-specific interactions did not meaningfully improve predictions. Male protagonists were slightly less warm than female protagonists, and movies with male leads received higher ratings on average (an association that was multiple times stronger than the relationships between movie ratings and warmth/competence). These findings suggest that, although audiences tend to favor warm, competent characters, the effects on movie evaluations are modest, indicating that character personality is only one of many factors shaping movie ratings. AI-assisted annotation with LLM_annotate and gpt-4.1-mini proved effective for large-scale analyses but occasionally fell short of manually generated annotations.

Method: 1) Controlled data synthesis pipeline generating high-quality data with explicit evidence, fine-grained error type labels, justifications, and corrections; 2) Construction of large-scale training data and manually verified benchmark InFi-Check-FG; 3) Development of InFi-Checker model that jointly provides evidence, classifies error types, and produces justifications with corrections.

Result: InFi-Checker achieves state-of-the-art performance on InFi-Check-FG benchmark and demonstrates strong generalization across various downstream tasks, significantly improving utility and trustworthiness of factuality evaluation.

Conclusion: The InFi-Check framework provides a more interpretable and fine-grained approach to fact-checking LLM outputs, addressing limitations of binary classification methods and enhancing the reliability of factuality assessment.

Abstract: Large language models (LLMs) often hallucinate, yet most existing fact-checking methods treat factuality evaluation as a binary classification problem, offering limited interpretability and failing to capture fine-grained error types. In this paper, we introduce InFi-Check, a framework for interpretable and fine-grained fact-checking of LLM outputs. Specifically, we first propose a controlled data synthesis pipeline that generates high-quality data featuring explicit evidence, fine-grained error type labels, justifications, and corrections. Based on this, we further construct large-scale training data and a manually verified benchmark InFi-Check-FG for fine-grained fact-checking of LLM outputs. Building on these high-quality training data, we further propose InFi-Checker, which can jointly provide supporting evidence, classify fine-grained error types, and produce justifications along with corrections. Experiments show that InFi-Checker achieves state-of-the-art performance on InFi-Check-FG and strong generalization across various downstream tasks, significantly improving the utility and trustworthiness of factuality evaluation.

Method: Propose a concrete, measurable definition of mergeability. Investigate several potential causes for mergeability differences, focusing on base model knowledge. Develop a simple weighted merging technique based on mergeability definition to better preserve weak knowledge in the base model.

Result: Base model knowledge is identified as a dominant factor: models fine-tuned on instances that the base model knows better are more mergeable than those fine-tuned on instances the base model struggles with. The weighted merging technique effectively preserves weak knowledge.

Conclusion: Understanding mergeability through base model knowledge enables better model merging strategies. The proposed weighted merging technique improves preservation of weak knowledge, advancing practical model merging applications.

Abstract: Model merging has emerged as a promising technique for combining multiple fine-tuned models into a single multitask model without retraining. However, the factors that determine whether merging will succeed or fail remain poorly understood. In this work, we investigate why specific models are merged better than others. To do so, we propose a concrete, measurable definition of mergeability. We investigate several potential causes for high or low mergeability, highlighting the base model knowledge as a dominant factor: Models fine-tuned on instances that the base model knows better are more mergeable than models fine-tuned on instances that the base model struggles with. Based on our mergeability definition, we explore a simple weighted merging technique that better preserves weak knowledge in the base model.

Method: Used translated TOFU benchmarks in seven language/script variants to test major unlearning algorithms, with focus on subspace-projection method.

Result: Most unlearning algorithms fail to remove facts outside the training language, but subspace-projection consistently outperforms others, achieving strong cross-lingual forgetting with minimal degradation.

Conclusion: Multilingual forgetting depends on geometry in weight space, motivating subspace-based approaches for future unlearning systems; analysis reveals shared interlingua structure in learned task subspaces.

Abstract: We present the first comprehensive evaluation of cross-lingual unlearning in multilingual LLMs. Using translated TOFU benchmarks in seven language/script variants, we test major unlearning algorithms and show that most fail to remove facts outside the training language, even when utility remains high. However, subspace-projection consistently outperforms the other methods, achieving strong cross-lingual forgetting with minimal degradation. Analysis of learned task subspaces reveals a shared interlingua structure: removing this shared subspace harms all languages, while removing language-specific components selectively affects one. These results demonstrate that multilingual forgetting depends on geometry in weight space, motivating subspace-based approaches for future unlearning systems.

Method: IDRBench combines: 1) modular multi-agent research framework with on-demand interaction, 2) scalable reference-grounded user simulator, and 3) interaction-aware evaluation suite that jointly measures interaction benefits (quality/alignment) and costs (turns/tokens).

Result: Experiments across seven state-of-the-art LLMs show that interaction consistently improves research quality and robustness, often outweighing differences in model capacity, while revealing substantial trade-offs in interaction efficiency.

Conclusion: Interaction is crucial for deep research systems to handle evolving goals, and IDRBench provides the first comprehensive framework to systematically evaluate interactive capabilities, revealing that interaction benefits often surpass model capacity differences but come with efficiency trade-offs.

Abstract: Deep research agents powered by Large Language Models (LLMs) can perform multi-step reasoning, web exploration, and long-form report generation. However, most existing systems operate in an autonomous manner, assuming fully specified user intent and evaluating only final outputs. In practice, research goals are often underspecified and evolve during exploration, making sustained interaction essential for robust alignment. Despite its importance, interaction remains largely invisible to existing deep research benchmarks, which neither model dynamic user feedback nor quantify its costs. We introduce IDRBench, the first benchmark for systematically evaluating interactive deep research. IDRBench combines a modular multi-agent research framework with on-demand interaction, a scalable reference-grounded user simulator, and an interaction-aware evaluation suite that jointly measures interaction benefits (quality and alignment) and costs (turns and tokens). Experiments across seven state-of-the-art LLMs show that interaction consistently improves research quality and robustness, often outweighing differences in model capacity, while revealing substantial trade-offs in interaction efficiency.

Method: The paper examines LLM responses to prompts, analyzing both lexical and syntactic linguistic factors that influence toxic output generation. Likely involves systematic prompting experiments and linguistic analysis of model responses.

Result: Findings show that LLMs can indeed generate toxic content when prompted, and specific linguistic patterns (both lexical choices and syntactic structures) significantly influence the production of harmful outputs.

Conclusion: Current alignment methods like RLHF are insufficient to prevent toxic output generation, and understanding linguistic triggers is crucial for developing more robust safety mechanisms in language models.

Abstract: In recent years, the advent of the attention mechanism has significantly advanced the field of natural language processing (NLP), revolutionizing text processing and text generation. This has come about through transformer-based decoder-only architectures, which have become ubiquitous in NLP due to their impressive text processing and generation capabilities. Despite these breakthroughs, language models (LMs) remain susceptible to generating undesired outputs: inappropriate, offensive, or otherwise harmful responses. We will collectively refer to these as ``toxic’’ outputs. Although methods like reinforcement learning from human feedback (RLHF) have been developed to align model outputs with human values, these safeguards can often be circumvented through carefully crafted prompts. Therefore, this paper examines the extent to which LLMs generate toxic content when prompted, as well as the linguistic factors – both lexical and syntactic – that influence the production of such outputs in generative models.

Method: GRASP LoRA treats global sparsity as a learnable control variable using a GRPO controller that interleaves with training. It periodically probes candidate prune ratios on a small micro development set and updates a single global prune ratio online based on reward signals. The method operates on merged source and target LoRA adapters on a frozen backbone, replacing grid search with one controller run followed by a single final merge and prune fine-tuning run.

Result: On cross-lingual transfer tasks (English to Arabic and Chinese) including XL-Sum summarization and MLQA question answering with Llama 3 8B, GRASP LoRA improves semantic faithfulness, content coverage, and answer quality over target-only and merge-and-prune baselines. It reduces end-to-end runtime by multiple times relative to grid search and lowers reliance on large development sets.

Conclusion: GRASP LoRA makes adapter reuse practical for low-resource deployment by eliminating the need for expensive grid search, reducing computational requirements, and improving cross-lingual transfer performance through learned optimal sparsity ratios.

Abstract: Parameter efficient fine tuning is a way to adapt LLMs to new languages when compute or data are limited, yet adapter pipelines usually choose a global prune ratio by grid search. This practice is computationally expensive and development set intensive, since it repeats training, freezes sparsity, and misses fractional optima. We introduce GRASP LoRA (GRPO Guided Adapter Sparsity Policy), which treats global sparsity as a learnable control variable. A GRPO controller interleaves with training, periodically probing candidate prune ratios on a small micro development set and updating a single global prune ratio online from its reward signal. It operates on merged source and target LoRA adapters on a frozen backbone and replaces grid search with one controller run that learns a prune ratio, followed by a single final merge and prune fine tuning run with pruning fixed to that ratio. On cross lingual transfer from English into Arabic and Chinese, including XL-Sum summarization and MLQA extractive question answering with Llama 3 8B, GRASP LoRA improves semantic faithfulness, content coverage, and answer quality over strong target only and merge and prune baselines. It reduces end to end runtime by multiple times relative to grid search, lowers reliance on large development sets, and makes adapter reuse practical for low resource deployment.

Method: Defined vertical domain accounting reasoning and proposed evaluation criteria based on GLM training data analysis. Evaluated GLM-6B, GLM-130B, GLM-4, and GPT-4 on accounting reasoning tasks using this framework.

Result: Prompt design significantly affects performance. GPT-4 demonstrated strongest capability among tested models. Despite gains, current models remain insufficient for real-world enterprise accounting applications.

Conclusion: Further optimization needed to unlock full practical value of LLMs for enterprise accounting, as current models still fall short of real-world requirements despite promising performance.

Abstract: Large language models are transforming learning, cognition, and research across many fields. Effectively integrating them into professional domains, such as accounting, is a key challenge for enterprise digital transformation. To address this, we define vertical domain accounting reasoning and propose evaluation criteria derived from an analysis of the training data characteristics of representative GLM models. These criteria support systematic study of accounting reasoning and provide benchmarks for performance improvement. Using this framework, we evaluate GLM-6B, GLM-130B, GLM-4, and OpenAI GPT-4 on accounting reasoning tasks. Results show that prompt design significantly affects performance, with GPT-4 demonstrating the strongest capability. Despite these gains, current models remain insufficient for real-world enterprise accounting, indicating the need for further optimization to unlock their full practical value.

Method: The paper analyzes the field’s methodological pipeline: from visual processing (image restoration, character recognition) through contextual analysis (artifact rejoining, dating) to advanced reasoning for automated decipherment and human-AI collaboration, examining technological shifts from classical computer vision to modern deep learning paradigms.

Result: The analysis maps digital resources for oracle bone, bronze, and bamboo slip scripts, identifies core challenges (data scarcity, disconnect between AI capabilities and humanistic inquiry), and synthesizes the field’s current state.

Conclusion: Advocates for a future research agenda focused on creating multimodal, few-shot, and human-centric AI systems to augment scholarly expertise in Chinese paleography.

Abstract: Chinese paleography, the study of ancient Chinese writing, is undergoing a computational turn powered by artificial intelligence. This position paper charts the trajectory of this emerging field, arguing that it is evolving from automating isolated visual tasks to creating integrated digital ecosystems for scholarly research. We first map the landscape of digital resources, analyzing critical datasets for oracle bone, bronze, and bamboo slip scripts. The core of our analysis follows the field’s methodological pipeline: from foundational visual processing (image restoration, character recognition), through contextual analysis (artifact rejoining, dating), to the advanced reasoning required for automated decipherment and human-AI collaboration. We examine the technological shift from classical computer vision to modern deep learning paradigms, including transformers and large multimodal models. Finally, we synthesize the field’s core challenges – notably data scarcity and a disconnect between current AI capabilities and the holistic nature of humanistic inquiry – and advocate for a future research agenda focused on creating multimodal, few-shot, and human-centric systems to augment scholarly expertise.

Method: Created MTMCS-Bench with over 30k multimodal (image+text) and unimodal (text-only) samples, featuring paired safe/unsafe dialogues. Evaluates two risk settings: escalation-based (gradual risk emergence) and context-switch (same scene supporting different intents). Uses structured metrics for intent recognition, safety-awareness, and helpfulness.

Result: Across 15 MLLMs (8 open-source, 7 proprietary), found persistent trade-offs between contextual safety and utility - models either miss gradual risks or over-refuse benign dialogues. Current guardrails mitigate some failures but don’t fully resolve multi-turn contextual risks.

Conclusion: MTMCS-Bench reveals critical gaps in MLLM safety evaluation, showing that multi-turn contextual risks remain challenging. The benchmark provides a comprehensive framework for assessing and improving multimodal conversational safety.

Abstract: Multimodal large language models (MLLMs) are increasingly deployed as assistants that interact through text and images, making it crucial to evaluate contextual safety when risk depends on both the visual scene and the evolving dialogue. Existing contextual safety benchmarks are mostly single-turn and often miss how malicious intent can emerge gradually or how the same scene can support both benign and exploitative goals. We introduce the Multi-Turn Multimodal Contextual Safety Benchmark (MTMCS-Bench), a benchmark of realistic images and multi-turn conversations that evaluates contextual safety in MLLMs under two complementary settings, escalation-based risk and context-switch risk. MTMCS-Bench offers paired safe and unsafe dialogues with structured evaluation. It contains over 30 thousand multimodal (image+text) and unimodal (text-only) samples, with metrics that separately measure contextual intent recognition, safety-awareness on unsafe cases, and helpfulness on benign ones. Across eight open-source and seven proprietary MLLMs, we observe persistent trade-offs between contextual safety and utility, with models tending to either miss gradual risks or over-refuse benign dialogues. Finally, we evaluate five current guardrails and find that they mitigate some failures but do not fully resolve multi-turn contextual risks.

Method: 1) Construct Ganit dataset: rigorously filtered and decontaminated Bengali math corpus with automatic difficulty tags based on pass@k of a strong evaluator model. 2) Curriculum-GRPO pipeline: combines multi-stage training (SFT + GRPO) with difficulty-aware sampling and verifiable rewards for format, numerical correctness, and Bengali reasoning.

Result: On Bn-MGSM and Bn-MSVAMP benchmarks, GanitLLM-4B improves over Qwen3-4B base by +8 and +7 accuracy points respectively, while increasing Bengali reasoning tokens from 14% to over 88% and reducing average solution length from 943 to 193 words.

Conclusion: The proposed Curriculum-GRPO pipeline with difficulty-aware dataset construction enables effective training of Bengali mathematical reasoning models, addressing the challenge of reward sparsity in low-resource languages while improving both accuracy and reasoning quality.

Abstract: We present a Bengali mathematical reasoning model called GanitLLM (named after the Bangla word for mathematics, “Ganit”), together with a new difficulty-aware Bengali math corpus and a curriculum-based GRPO pipeline. Bengali is one of the world’s most widely spoken languages, yet existing LLMs either reason in English and then translate, or simply fail on multi-step Bengali math, in part because reinforcement learning recipes are tuned for high-resource languages and collapse under reward sparsity in low-resource settings. To address this, we construct Ganit, a rigorously filtered and decontaminated Bengali math dataset with automatic difficulty tags derived from the pass@k of a strong evaluator model. Building on this dataset, we propose Curriculum-GRPO, which combines multi-stage training (SFT + GRPO) with difficulty-aware sampling and verifiable rewards for format, numerical correctness, and Bengali reasoning. On Bn-MGSM and Bn-MSVAMP, GanitLLM-4B improves over its Qwen3-4B base by +8 and +7 accuracy points, respectively, while increasing the percentage of Bengali reasoning tokens from 14% to over 88% and reducing average solution length from 943 to 193 words.

Method: MEM-MCL creates expert models through instruction tuning for specific sentiment tasks, then merges them using evolutionary algorithms to form a unified model. The merging process is optimized with weak data, and curriculum learning provides learning sequences based on task difficulty to improve knowledge extraction from LLMs.

Result: The proposed MEM-MCL model outperforms conventional LLMs in a majority of sentiment analysis tasks, achieving superior results across various subtasks.

Conclusion: The hybrid approach combining evolutionary model merging with meta-data driven curriculum learning effectively enhances sentiment analysis performance in large language models, addressing the need for high accuracy and scalability across multiple subtasks.

Abstract: The emergence of large language models (LLMs) has significantly transformed natural language processing (NLP), enabling more generalized models to perform various tasks with minimal training. However, traditional sentiment analysis methods, which focus on individual tasks such as sentiment classification or aspect-based analysis, are not practical for real-world applications that usually require handling multiple tasks. While offering flexibility, LLMs in sentiment-specific tasks often fall short of the required accuracy. Techniques like fine-tuning and evolutionary model merging help integrate models into a unified framework, which can improve the learning performance while reducing computational costs. The use of task meta-data and curriculum learning to optimize learning processes remains underexplored, while sentiment analysis is a critical task in NLP that requires high accuracy and scalability across multiple subtasks. In this study, we propose a hybrid learning model called Multi-stage Evolutionary Model Merging with Meta data driven Curriculum Learning (MEM-MCL), to enhance the sentiment analysis in large language modeling. In particular, expert models are created through instruction tuning for specific sentiment tasks and then merged using evolutionary algorithms to form a unified model. The merging process is optimized with weak data to enhance performance across tasks. The curriculum learning is incorporated to provide a learning sequence based on task difficulty, improving knowledge extraction from LLMs. Experiment results demonstrate that the proposed MEM-MCL model outperforms conventional LLMs in a majority of sentiment analysis tasks, achieving superior results across various subtasks.

Method: Proposes epistemically-calibrated reasoning (EpiCaR) as a training objective that jointly optimizes reasoning performance and calibration. Uses iterative supervised fine-tuning with explicit self-evaluation signals to teach models both how to reason and when to trust their reasoning.

Result: Achieves Pareto-superiority over baselines in both accuracy and calibration on Llama-3 and Qwen-3 families, especially for models with 3B+ parameters. Generalizes to OOD mathematical reasoning (GSM8K) and code generation (MBPP). Enables 3X reduction in inference compute (matching K=30 performance with only K=10 samples).

Conclusion: Treating reasoning as an epistemic learning problem with joint optimization of performance and calibration addresses model collapse issues, improves uncertainty representation, and significantly reduces inference costs while maintaining accuracy.

Abstract: Improving the reasoning abilities of large language models (LLMs) has largely relied on iterative self-training with model-generated data. While effective at boosting accuracy, existing approaches primarily reinforce successful reasoning paths, incurring a substantial calibration cost: models become overconfident and lose the ability to represent uncertainty. This failure has been characterized as a form of model collapse in alignment, where predictive distributions degenerate toward low-variance point estimates. We address this issue by reframing reasoning training as an epistemic learning problem, in which models must learn not only how to reason, but also when their reasoning should be trusted. We propose epistemically-calibrated reasoning (EpiCaR) as a training objective that jointly optimizes reasoning performance and calibration, and instantiate it within an iterative supervised fine-tuning framework using explicit self-evaluation signals. Experiments on Llama-3 and Qwen-3 families demonstrate that our approach achieves Pareto-superiority over standard baselines in both accuracy and calibration, particularly in models with sufficient reasoning capacity (e.g., 3B+). This framework generalizes effectively to OOD mathematical reasoning (GSM8K) and code generation (MBPP). Ultimately, our approach enables a 3X reduction in inference compute, matching the K=30 performance of STaR with only K=10 samples in capable models.

Method: The authors identify the BOS sink phenomenon where attention heads with high BOS sink scores act as “dumping grounds” for superfluous attention weights. They introduce a pruning strategy that removes high-BOS sink heads, particularly in deeper layers where this phenomenon is most pronounced.

Result: Experiments on Gemma-3, Llama-3.1, and Qwen3 show that BOS sink-based pruning identifies redundant components more reliably than weight- or activation-based criteria, preserves performance close to dense baselines even under aggressive pruning, and remains stable across different sequence lengths.

Conclusion: Structural properties of attention (like BOS sink scores) provide a more intuitive and robust basis for model compression than magnitude-based methods, offering a functional explanation for why certain transformer components are redundant.

Abstract: Large Language Models (LLMs) are known to contain significant redundancy, yet a systematic explanation for why certain components, particularly in higher layers, are more redundant has remained elusive. In this work, we identify the BOS sink phenomenon as a key mechanism driving this layer-wise sensitivity. We show that attention heads with high BOS sink scores are strongly associated with functional redundancy: such heads, especially in deeper layers, contribute little to predictive performance and effectively serve as \emph{dumping grounds} for superfluous attention weights. This provides a concrete functional explanation for the structural redundancy reported in prior studies. Leveraging this insight, we introduce a simple pruning strategy that removes high-BOS sink heads. Experiments on Gemma-3, Llama-3.1, and Qwen3 demonstrate that this approach identifies redundant transformer components more reliably than weight- or activation-based criteria, while preserving performance close to dense baselines even under aggressive pruning. Moreover, we find that the behavior of sink heads remains stable across different sequence lengths. Overall, our results suggest that structural properties of attention offer a more intuitive and robust basis for model compression than magnitude-based methods.

Method: CIRAG introduces: (1) Iterative Construction-Integration module that constructs candidate triples and history-conditionally integrates them to distill core triples and generate next-hop queries, preserving multiple evidence chains; (2) Adaptive Cascaded Multi-Granularity Generation that progressively expands contextual evidence from triples to sentences to full passages based on problem requirements; (3) Trajectory Distillation to distill teacher model’s integration policy into lightweight student for efficient long-horizon reasoning.

Result: Extensive experiments demonstrate that CIRAG achieves superior performance compared to existing iRAG methods.

Conclusion: CIRAG effectively addresses the limitations of existing iRAG approaches by mitigating greedy expansion traps and granularity mismatches through its novel construction-integration and adaptive multi-granularity generation framework.

Abstract: Triple-based Iterative Retrieval-Augmented Generation (iRAG) mitigates document-level noise for multi-hop question answering. However, existing methods still face limitations: (i) greedy single-path expansion, which propagates early errors and fails to capture parallel evidence from different reasoning branches, and (ii) granularity-demand mismatch, where a single evidence representation struggles to balance noise control with contextual sufficiency. In this paper, we propose the Construction-Integration Retrieval and Adaptive Generation model, CIRAG. It introduces an Iterative Construction-Integration module that constructs candidate triples and history-conditionally integrates them to distill core triples and generate the next-hop query. This module mitigates the greedy trap by preserving multiple plausible evidence chains. Besides, we propose an Adaptive Cascaded Multi-Granularity Generation module that progressively expands contextual evidence based on the problem requirements, from triples to supporting sentences and full passages. Moreover, we introduce Trajectory Distillation, which distills the teacher model’s integration policy into a lightweight student, enabling efficient and reliable long-horizon reasoning. Extensive experiments demonstrate that CIRAG achieves superior performance compared to existing iRAG methods.

Method: Two data augmentation strategies: (1) self-training with pseudo-labels from unlabeled speech, and (2) TTS-based augmentation using synthetic speech from Klaam TTS system. Fine-tuning OpenAI Whisper models with these techniques.

Result: Best model (Whisper-Medium with combined self-training and TTS augmentation) achieved 57.1% WER on evaluation set and 51.6% on out-of-domain holdout set, substantially outperforming zero-shot multilingual Whisper (78.8% WER) and MSA-specialized models (73.8-123% WER).

Conclusion: Strategic data augmentation can overcome resource limitations for low-resource dialects, providing a practical roadmap for developing ASR systems for marginalized language varieties. All models, benchmarks, and training pipelines are publicly released.

Abstract: Although many Automatic Speech Recognition (ASR) systems have been developed for Modern Standard Arabic (MSA) and Dialectal Arabic (DA), few studies have focused on dialect-specific implementations, particularly for low-resource Arabic dialects such as Sudanese. This paper presents a comprehensive study of data augmentation techniques for fine-tuning OpenAI Whisper models and establishes the first benchmark for the Sudanese dialect. Two augmentation strategies are investigated: (1) self-training with pseudo-labels generated from unlabeled speech, and (2) TTS-based augmentation using synthetic speech from the Klaam TTS system. The best-performing model, Whisper-Medium fine-tuned with combined self-training and TTS augmentation (28.4 hours), achieves a Word Error Rate (WER) of 57.1% on the evaluation set and 51.6% on an out-of-domain holdout set substantially outperforming zero-shot multilingual Whisper (78.8% WER) and MSA-specialized Arabic models (73.8-123% WER). All experiments used low-cost resources (Kaggle free tier and Lightning.ai trial), demonstrating that strategic data augmentation can overcome resource limitations for low-resource dialects and provide a practical roadmap for developing ASR systems for low-resource Arabic dialects and other marginalized language varieties. The models, evaluation benchmarks, and reproducible training pipelines are publicly released to facilitate future research on low-resource Arabic ASR.

Method: Laser reformulates visual deduction via Dynamic Windowed Alignment Learning (DWAL), aligning latent states with dynamic validity windows of future semantics instead of point-wise prediction. It enforces a “Forest-before-Trees” cognitive hierarchy and maintains interpretability via decodable trajectories while stabilizing learning via Self-Refined Superposition.

Result: Achieves state-of-the-art performance among latent reasoning methods, surpassing Monet baseline by 5.03% on average across 6 benchmarks, with extreme efficiency (97% reduction in inference tokens) and robust generalization to out-of-distribution domains.

Conclusion: Laser provides an effective paradigm for visual reasoning that addresses key limitations of existing methods through dynamic windowed alignment, achieving superior performance with high efficiency while maintaining interpretability.

Abstract: While Chain-of-Thought empowers Large Vision-Language Models with multi-step reasoning, explicit textual rationales suffer from an information bandwidth bottleneck, where continuous visual details are discarded during discrete tokenization. Recent latent reasoning methods attempt to address this challenge, but often fall prey to premature semantic collapse due to rigid autoregressive objectives. In this paper, we propose Laser, a novel paradigm that reformulates visual deduction via Dynamic Windowed Alignment Learning (DWAL). Instead of forcing a point-wise prediction, Laser aligns the latent state with a dynamic validity window of future semantics. This mechanism enforces a “Forest-before-Trees” cognitive hierarchy, enabling the model to maintain a probabilistic superposition of global features before narrowing down to local details. Crucially, Laser maintains interpretability via decodable trajectories while stabilizing unconstrained learning via Self-Refined Superposition. Extensive experiments on 6 benchmarks demonstrate that Laser achieves state-of-the-art performance among latent reasoning methods, surpassing the strong baseline Monet by 5.03% on average. Notably, it achieves these gains with extreme efficiency, reducing inference tokens by more than 97%, while demonstrating robust generalization to out-of-distribution domains.

Method: Proposes a new research task of automated hallucination attribution. Creates AgentHallu benchmark with: 693 high-quality trajectories across 7 agent frameworks and 5 domains, hallucination taxonomy with 5 categories and 14 sub-categories, and multi-level human annotations (binary labels, responsible steps, causal explanations).

Result: Evaluation of 13 leading models shows the task is challenging: best-performing model achieves only 41.1% step localization accuracy. Tool-use hallucinations are most challenging at just 11.6%. GPT-5 and Gemini-2.5-Pro struggle with the task.

Conclusion: AgentHallu will catalyze future research into developing robust, transparent, and reliable agentic systems by providing a comprehensive benchmark for automated hallucination attribution in multi-step LLM agent workflows.

Abstract: As LLM-based agents operate over sequential multi-step reasoning, hallucinations arising at intermediate steps risk propagating along the trajectory, thus degrading overall reliability. Unlike hallucination detection in single-turn responses, diagnosing hallucinations in multi-step workflows requires identifying which step causes the initial divergence. To fill this gap, we propose a new research task, automated hallucination attribution of LLM-based agents, aiming to identify the step responsible for the hallucination and explain why. To support this task, we introduce AgentHallu, a comprehensive benchmark with: (1) 693 high-quality trajectories spanning 7 agent frameworks and 5 domains, (2) a hallucination taxonomy organized into 5 categories (Planning, Retrieval, Reasoning, Human-Interaction, and Tool-Use) and 14 sub-categories, and (3) multi-level annotations curated by humans, covering binary labels, hallucination-responsible steps, and causal explanations. We evaluate 13 leading models, and results show the task is challenging even for top-tier models (like GPT-5, Gemini-2.5-Pro). The best-performing model achieves only 41.1% step localization accuracy, where tool-use hallucinations are the most challenging at just 11.6%. We believe AgentHallu will catalyze future research into developing robust, transparent, and reliable agentic systems.

Method: Introduces Positional Decay Reweighting (PDR), a training-free plug-and-play framework that explicitly reweights token-level scores based on linguistic properties. PDR amplifies distinct signals from early positions (where model uncertainty is highest) while suppressing noise from later positions as context accumulates.

Result: Extensive experiments show PDR acts as a robust prior that can enhance a wide range of advanced memorization detection methods across multiple benchmarks, demonstrating improved performance in detecting pre-training data in LLMs.

Conclusion: PDR effectively leverages the linguistic property that memorization signals are skewed toward high-entropy initial tokens and decay with context accumulation, providing a simple yet effective enhancement to existing memorization detection methods without requiring training.

Abstract: Detecting pre-training data in Large Language Models (LLMs) is crucial for auditing data privacy and copyright compliance, yet it remains challenging in black-box, zero-shot settings where computational resources and training data are scarce. While existing likelihood-based methods have shown promise, they typically aggregate token-level scores using uniform weights, thereby neglecting the inherent information-theoretic dynamics of autoregressive generation. In this paper, we hypothesize and empirically validate that memorization signals are heavily skewed towards the high-entropy initial tokens, where model uncertainty is highest, and decay as context accumulates. To leverage this linguistic property, we introduce Positional Decay Reweighting (PDR), a training-free and plug-and-play framework. PDR explicitly reweights token-level scores to amplify distinct signals from early positions while suppressing noise from later ones. Extensive experiments show that PDR acts as a robust prior and can usually enhance a wide range of advanced methods across multiple benchmarks.

Method: Reformulates MABSA as a generative task using multimodal LLMs with prompt-based generation. Introduces a dependency-syntax-guided sentiment cue strategy that prunes and textualizes aspect-centered dependency syntax trees to enhance aspect-oriented reasoning and explainability. Constructs new datasets with sentiment explanations for fine-tuning.

Result: The approach achieves consistent gains in sentiment classification accuracy and produces faithful, aspect-grounded explanations, demonstrating both improved performance and enhanced explainability.

Conclusion: The proposed generative framework successfully addresses the explainability gap in MABSA by combining sentiment prediction with natural language explanation generation, leveraging multimodal LLMs and syntax-guided reasoning for more interpretable fine-grained sentiment analysis.

Abstract: Multimodal aspect-based sentiment analysis (MABSA) aims to identify aspect-level sentiments by jointly modeling textual and visual information, which is essential for fine-grained opinion understanding in social media. Existing approaches mainly rely on discriminative classification with complex multimodal fusion, yet lacking explicit sentiment explainability. In this paper, we reformulate MABSA as a generative and explainable task, proposing a unified framework that simultaneously predicts aspect-level sentiment and generates natural language explanations. Based on multimodal large language models (MLLMs), our approach employs a prompt-based generative paradigm, jointly producing sentiment and explanation. To further enhance aspect-oriented reasoning capabilities, we propose a dependency-syntax-guided sentiment cue strategy. This strategy prunes and textualizes the aspect-centered dependency syntax tree, guiding the model to distinguish different sentiment aspects and enhancing its explainability. To enable explainability, we use MLLMs to construct new datasets with sentiment explanations to fine-tune. Experiments show that our approach not only achieves consistent gains in sentiment classification accuracy, but also produces faithful, aspect-grounded explanations.

Method: 1. Created DISTRACTMATH-BN benchmark by augmenting existing Bangla math datasets (MGSM and MSVAMP) with semantically coherent but computationally irrelevant information. 2. Evaluated 7 models (3B to 12B parameters) on this benchmark. 3. Proposed DAGGER approach that reformulates math problem solving as executable computational graph generation with explicit modeling of distractor nodes. 4. Fine-tuned Gemma-3 models using supervised fine-tuning followed by Group Relative Policy Optimization.

Result: 1. Standard models dropped by up to 41 points in performance when faced with distractors. 2. Reasoning-specialized models declined by 14-20 points despite using 5x more tokens. 3. DAGGER achieved comparable weighted accuracy on augmented benchmarks while using 89% fewer tokens than reasoning models. 4. Robustness emerged without explicit training on distractor-augmented examples.

Conclusion: Enforcing structured intermediate representations (like computational graphs) improves both robustness and inference efficiency in mathematical reasoning compared to free-form approaches, especially in noisy, low-resource settings. This suggests that structured reasoning approaches are more resilient to irrelevant context.

Abstract: Chain-of-Thought (CoT) prompting is widely adopted for mathematical problem solving, including in low-resource languages, yet its behavior under irrelevant context remains underexplored. To systematically study this challenge, we introduce DISTRACTMATH-BN, a Bangla benchmark that augments MGSM and MSVAMP with semantically coherent but computationally irrelevant information. Evaluating seven models ranging from 3B to 12B parameters, we observe substantial performance degradation under distractors: standard models drop by up to 41 points, while reasoning-specialized models decline by 14 to 20 points despite consuming five times more tokens. We propose †DAGGER, which reformulates mathematical problem solving as executable computational graph generation with explicit modeling of distractor nodes. Fine-tuning Gemma-3 models using supervised fine-tuning followed by Group Relative Policy Optimization achieves comparable weighted accuracy on augmented benchmarks while using 89 percent fewer tokens than reasoning models. Importantly, this robustness emerges without explicit training on distractor-augmented examples. Our results suggest that enforcing structured intermediate representations improves robustness and inference efficiency in mathematical reasoning compared to free-form approaches, particularly in noisy, low-resource settings.

Method: Uses mirrored probe pairs under strict dual-framing scheme with affirmative assertions favoring different targets. Employs randomized instructional wrappers, fixed-choice Likert responses, LLM-based judges for normalization, and produces quantitative bias indicators with effect sizes and neutrality rates.

Result: Provides a standardized methodology for cross-lingual and framing-sensitive bias measurement that complements existing audits, enabling benchmarking of robustness for better deployment decisions.

Conclusion: BiasLab contributes a reproducible framework for output-level bias evaluation across diverse bias axes (demographic, cultural, political, geopolitical) with structured reports and comparative visualizations.

Abstract: Large Language Models (LLMs) are increasingly deployed in high-stakes contexts where their outputs influence real-world decisions. However, evaluating bias in LLM outputs remains methodologically challenging due to sensitivity to prompt wording, limited multilingual coverage, and the lack of standardized metrics that enable reliable comparison across models. This paper introduces BiasLab, an open-source, model-agnostic evaluation framework for quantifying output-level (extrinsic) bias through a multilingual, robustness-oriented experimental design. BiasLab constructs mirrored probe pairs under a strict dual-framing scheme: an affirmative assertion favoring Target A and a reverse assertion obtained by deterministic target substitution favoring Target B, while preserving identical linguistic structure. To reduce dependence on prompt templates, BiasLab performs repeated evaluation under randomized instructional wrappers and enforces a fixed-choice Likert response format to maximize comparability across models and languages. Responses are normalized into agreement labels using an LLM-based judge, aligned for polarity consistency across framings, and aggregated into quantitative bias indicators with descriptive statistics including effect sizes and neutrality rates. The framework supports evaluation across diverse bias axes, including demographic, cultural, political, and geopolitical topics, and produces reproducible artifacts such as structured reports and comparative visualizations. BiasLab contributes a standardized methodology for cross-lingual and framing-sensitive bias measurement that complements intrinsic and dataset-based audits, enabling researchers and institutions to benchmark robustness and make better-informed deployment decisions.

Method: Paraphrasing Adversarial Attack (PAA) - a black-box optimization method that searches for semantically equivalent paraphrases yielding higher review scores while maintaining linguistic naturalness. Uses in-context learning with previous paraphrases and scores to guide candidate generation.

Result: PAA consistently increases review scores across five ML/NLP conferences with three LLM reviewers and five attacking models, without changing paper claims. Human evaluation confirms paraphrases maintain meaning and naturalness. Attacked papers show increased perplexity in reviews (potential detection signal), and paraphrasing submissions can partially mitigate attacks.

Conclusion: PAA demonstrates significant vulnerabilities in LLM-based peer review systems, showing that semantic-preserving paraphrasing can manipulate scores. The method reveals a need for more robust evaluation frameworks and suggests perplexity analysis as a potential detection mechanism.

Abstract: The use of large language models (LLMs) in peer review systems has attracted growing attention, making it essential to examine their potential vulnerabilities. Prior attacks rely on prompt injection, which alters manuscript content and conflates injection susceptibility with evaluation robustness. We propose the Paraphrasing Adversarial Attack (PAA), a black-box optimization method that searches for paraphrased sequences yielding higher review scores while preserving semantic equivalence and linguistic naturalness. PAA leverages in-context learning, using previous paraphrases and their scores to guide candidate generation. Experiments across five ML and NLP conferences with three LLM reviewers and five attacking models show that PAA consistently increases review scores without changing the paper’s claims. Human evaluation confirms that generated paraphrases maintain meaning and naturalness. We also find that attacked papers exhibit increased perplexity in reviews, offering a potential detection signal, and that paraphrasing submissions can partially mitigate attacks.

Method: Proposes a hierarchical attack comment detection dataset with explicit encoding of reply structures and chronological order. Introduces a divide-and-conquer framework that decomposes attack detection into hierarchical subtasks: explicit detection, implicit intent inference, and target identification using specialized lightweight models.

Result: Smaller models using the hierarchical framework significantly outperform larger monolithic models relying on parameter scaling. The approach demonstrates effectiveness on both the proposed dataset and benchmark intention detection datasets.

Conclusion: Structured task decomposition with spatiotemporal modeling effectively addresses limitations in verbal attack detection, particularly for implicit and context-dependent attacks in Chinese social media conversations.

Abstract: In the digital era, effective identification and analysis of verbal attacks are essential for maintaining online civility and ensuring social security. However, existing research is limited by insufficient modeling of conversational structure and contextual dependency, particularly in Chinese social media where implicit attacks are prevalent. Current attack detection studies often emphasize general semantic understanding while overlooking user response relationships, hindering the identification of implicit and context-dependent attacks. To address these challenges, we present the novel “Hierarchical Attack Comment Detection” dataset and propose a divide-and-conquer, fine-grained framework for verbal attack recognition based on spatiotemporal information. The proposed dataset explicitly encodes hierarchical reply structures and chronological order, capturing complex interaction patterns in multi-turn discussions. Building on this dataset, the framework decomposes attack detection into hierarchical subtasks, where specialized lightweight models handle explicit detection, implicit intent inference, and target identification under constrained context. Extensive experiments on the proposed dataset and benchmark intention detection datasets show that smaller models using our framework significantly outperform larger monolithic models relying on parameter scaling, demonstrating the effectiveness of structured task decomposition.

Method: Three-stage analysis: from phenomenon to mechanism to interpretation. They quantify distributional clarity using the Silhouette Coefficient (S) measuring intra-class compactness and inter-class separation in probability assignments to correct vs. incorrect responses. They also introduce a Silhouette-Aware Reweighting strategy that prioritizes low-S samples during training.

Result: High Silhouette Coefficient strongly correlates with RL performance, while low S is associated with severe logic errors and reasoning instability. Experiments across six mathematical benchmarks show consistent improvements across all model families, with gains up to 5.9 points on AIME24.

Conclusion: Distributional clarity is established as a fundamental, trainable property underlying RL-Friendliness. The Silhouette Coefficient serves as a diagnostic tool and the reweighting strategy provides a practical method to improve RL performance across different model families.

Abstract: Language model families exhibit striking disparity in their capacity to benefit from reinforcement learning: under identical training, models like Qwen achieve substantial gains, while others like Llama yield limited improvements. Complementing data-centric approaches, we reveal that this disparity reflects a hidden structural property: \textbf{distributional clarity} in probability space. Through a three-stage analysis-from phenomenon to mechanism to interpretation-we uncover that RL-friendly models exhibit intra-class compactness and inter-class separation in their probability assignments to correct vs. incorrect responses. We quantify this clarity using the \textbf{Silhouette Coefficient} ($S$) and demonstrate that (1) high $S$ correlates strongly with RL performance; (2) low $S$ is associated with severe logic errors and reasoning instability. To confirm this property, we introduce a Silhouette-Aware Reweighting strategy that prioritizes low-$S$ samples during training. Experiments across six mathematical benchmarks show consistent improvements across all model families, with gains up to 5.9 points on AIME24. Our work establishes distributional clarity as a fundamental, trainable property underlying RL-Friendliness.

Method: Models agentic RAG reasoning as a rollout tree where each reasoning step maps to a node. Uses Monte Carlo estimation over descendant outcomes to estimate step utility, enabling fine-grained process advantages without intermediate labels. Introduces efficient online tree construction strategy to preserve exploration diversity under computational constraints.

Result: Experiments on seven multi-hop and general QA benchmarks across multiple model scales show TreePS-RAG consistently and significantly outperforms both outcome-supervised and leading process-supervised RL methods, with rollout cost comparable to strong baselines like Search-R1.

Conclusion: TreePS-RAG provides an effective online RL framework for agentic RAG that enables step-wise credit assignment using only outcome-based rewards, addressing limitations of both sparse reward RL and costly process supervision methods.

Abstract: Agentic retrieval-augmented generation (RAG) formulates question answering as a multi-step interaction between reasoning and information retrieval, and has recently been advanced by reinforcement learning (RL) with outcome-based supervision. While effective, relying solely on sparse final rewards limits step-wise credit assignment and provides weak guidance for intermediate reasoning and actions. Recent efforts explore process-level supervision, but typically depend on offline constructed training data, which risks distribution shift, or require costly intermediate annotations. We present TreePS-RAG, an online, tree-based RL framework for agentic RAG that enables step-wise credit assignment while retaining standard outcome-only rewards. Our key insight is to model agentic RAG reasoning as a rollout tree, where each reasoning step naturally maps to a node. This tree structure allows step utility to be estimated via Monte Carlo estimation over its descendant outcomes, yielding fine-grained process advantages without requiring intermediate labels. To make this paradigm practical, we introduce an efficient online tree construction strategy that preserves exploration diversity under a constrained computational budget. With a rollout cost comparable to strong baselines like Search-R1, experiments on seven multi-hop and general QA benchmarks across multiple model scales show that TreePS-RAG consistently and significantly outperforms both outcome-supervised and leading process-supervised RL methods.

Method: Uses a Teacher-Student architecture: Teacher network trained on articulatory phonetic features produces target embeddings, while Student network learns to approximate these from raw characters. Training uses three-phase curriculum on 57M toponyms with hard negative triplets for discrimination.

Result: Achieves 89.2% Recall@1 on MEHDIE Hebrew-Arabic benchmark, outperforming Levenshtein (81.5%) and Jaro-Winkler (78.5%). Student network achieves 96.6% cosine similarity to Teacher outputs with only 1.7M parameters.

Conclusion: Symphonym enables effective cross-script toponym matching and will support fuzzy phonetic reconciliation across the World Historical Gazetteer’s 67 million place names, with code and models publicly available.

Abstract: Linking place names across languages and writing systems is a fundamental challenge in digital humanities and geographic information retrieval. Existing approaches rely on language-specific phonetic algorithms or transliteration rules that fail when names cross script boundaries – no string metric can determine that “Moscow” when rendered in Cyrillic or Arabic refer to the same city. I present Symphonym, a neural embedding system that maps toponyms from 20 writing systems into a unified 128-dimensional phonetic space. A Teacher network trained on articulatory phonetic features (via Epitran and PanPhon) produces target embeddings, while a Student network learns to approximate these from raw characters. At inference, only the lightweight Student (1.7M parameters) is required, enabling deployment without runtime phonetic conversion. Training uses a three-phase curriculum on 57 million toponyms from GeoNames, Wikidata, and the Getty Thesaurus of Geographic Names. Phase 1 trains the Teacher on 467K phonetically-grounded triplets. Phase 2 aligns the Student to Teacher outputs across 23M samples, achieving 96.6% cosine similarity. Phase 3 fine-tunes on 3.3M hard negative triplets – negatives sharing prefix and script with the anchor but referring to different places – to sharpen discrimination. Evaluation on the MEHDIE Hebrew-Arabic benchmark achieves 89.2% Recall@1, outperforming Levenshtein (81.5%) and Jaro-Winkler (78.5%). The system is optimised for cross-script matching; same-script variants can be handled by complementary string methods. Symphonym will enable fuzzy phonetic reconciliation and search across the World Historical Gazetteer’s 67 million toponyms. Code and models are publicly available.

Method: Proposes SynthSmith, a feature-based synthesis pipeline that generates diverse tasks, verified solutions, and test cases. Uses synthetic data for both supervised fine-tuning and reinforcement learning to train the X-Coder model series.

Result: X-Coder models achieve 62.9 avg@8 on LiveCodeBench v5 and 55.8 on v6, outperforming larger models like DeepCoder-14B-Preview and AReal-boba2-14B despite having only 7B parameters. Scaling laws hold on synthetic data.

Conclusion: Scaling high-quality synthetic data with staged training can advance code reasoning while reducing reliance on real-world coding data. The approach demonstrates strong potential for synthetic data in competitive programming.

Abstract: Competitive programming presents great challenges for Code LLMs due to its intensive reasoning demands and high logical complexity. However, current Code LLMs still rely heavily on real-world data, which limits their scalability. In this paper, we explore a fully synthetic approach: training Code LLMs with entirely generated tasks, solutions, and test cases, to empower code reasoning models without relying on real-world data. To support this, we leverage feature-based synthesis to propose a novel data synthesis pipeline called SynthSmith. SynthSmith shows strong potential in producing diverse and challenging tasks, along with verified solutions and tests, supporting both supervised fine-tuning and reinforcement learning. Based on the proposed synthetic SFT and RL datasets, we introduce the X-Coder model series, which achieves a notable pass rate of 62.9 avg@8 on LiveCodeBench v5 and 55.8 on v6, outperforming DeepCoder-14B-Preview and AReal-boba2-14B despite having only 7B parameters. In-depth analysis reveals that scaling laws hold on our synthetic dataset, and we explore which dimensions are more effective to scale. We further provide insights into code-centric reinforcement learning and highlight the key factors that shape performance through detailed ablations and analysis. Our findings demonstrate that scaling high-quality synthetic data and adopting staged training can greatly advance code reasoning, while mitigating reliance on real-world coding data.

Method: Introduces RealMem benchmark with synthesis pipeline: Project Foundation Construction, Multi-Agent Dialogue Generation, and Memory and Schedule Management to simulate dynamic memory evolution. Uses natural user queries for evaluation across 11 project scenarios.

Result: Experiments show current memory systems struggle significantly with managing long-term project states and dynamic context dependencies inherent in real-world projects.

Conclusion: RealMem addresses a critical gap in LLM memory evaluation by providing a realistic benchmark for project-oriented scenarios, revealing limitations in existing memory systems for long-term consistency.

Abstract: As Large Language Models (LLMs) evolve from static dialogue interfaces to autonomous general agents, effective memory is paramount to ensuring long-term consistency. However, existing benchmarks primarily focus on casual conversation or task-oriented dialogue, failing to capture “long-term project-oriented” interactions where agents must track evolving goals. To bridge this gap, we introduce RealMem, the first benchmark grounded in realistic project scenarios. RealMem comprises over 2,000 cross-session dialogues across eleven scenarios, utilizing natural user queries for evaluation. We propose a synthesis pipeline that integrates Project Foundation Construction, Multi-Agent Dialogue Generation, and Memory and Schedule Management to simulate the dynamic evolution of memory. Experiments reveal that current memory systems face significant challenges in managing the long-term project states and dynamic context dependencies inherent in real-world projects. Our code and datasets are available at https://github.com/AvatarMemory/RealMemBench.

Method: The authors introduce Architectural Fingerprinting, a probing framework that isolates architectural effects on representation. They apply this to a controlled suite of 24 pre-trained speech encoders ranging from 39M to 3.3B parameters, analyzing how different speech features (phonemes, speaker gender, accent, duration) are processed across network depth.

Result: The analysis reveals divergent processing hierarchies: Conformers implement a “Categorize Early” strategy, resolving phoneme categories 29% earlier in depth and speaker gender by 16% depth. In contrast, Transformers use an “Integrate Late” approach, deferring phoneme, accent, and duration encoding to deep layers (49-57% depth).

Conclusion: The architectural fingerprints suggest design heuristics: Conformers’ front-loaded categorization may benefit low-latency streaming applications, while Transformers’ deep integration may favor tasks requiring rich context and cross-utterance normalization. These findings provide guidance for architecture selection based on application requirements.

Abstract: In speech language modeling, two architectures dominate the frontier: the Transformer and the Conformer. However, it remains unknown whether their comparable performance stems from convergent processing strategies or distinct architectural inductive biases. We introduce Architectural Fingerprinting, a probing framework that isolates the effect of architecture on representation, and apply it to a controlled suite of 24 pre-trained encoders (39M-3.3B parameters). Our analysis reveals divergent hierarchies: Conformers implement a “Categorize Early” strategy, resolving phoneme categories 29% earlier in depth and speaker gender by 16% depth. In contrast, Transformers “Integrate Late,” deferring phoneme, accent, and duration encoding to deep layers (49-57%). These fingerprints suggest design heuristics: Conformers’ front-loaded categorization may benefit low-latency streaming, while Transformers’ deep integration may favor tasks requiring rich context and cross-utterance normalization.

Method: The authors define Private State Interactive Tasks (PSITs), prove an impossibility theorem about chat-based agents handling them, develop a self-consistency testing protocol to evaluate agents across forked dialogue branches, and propose a novel architecture with explicit private working memory.

Result: Standard chat-based LLMs and retrieval-based memory baselines fail the self-consistency test regardless of scale, showing semantic retrieval doesn’t enable true state maintenance. The proposed architecture with private working memory successfully restores consistency.

Conclusion: Private state maintenance is a necessary component for interactive language agents, requiring explicit architectural support beyond standard chat interfaces and semantic retrieval mechanisms.

Abstract: As LLMs move from text completion toward autonomous agents, they remain constrained by the standard chat interface, which lacks private working memory. This raises a fundamental question: can agents reliably perform interactive tasks that depend on hidden state? We define Private State Interactive Tasks (PSITs), which require agents to generate and maintain hidden information while producing consistent public responses. We show theoretically that any agent restricted to the public conversation history cannot simultaneously preserve secrecy and consistency in PSITs, yielding an impossibility theorem. To empirically validate this limitation, we introduce a self-consistency testing protocol that evaluates whether agents can maintain a hidden secret across forked dialogue branches. Standard chat-based LLMs and retrieval-based memory baselines fail this test regardless of scale, demonstrating that semantic retrieval does not enable true state maintenance. To address this, we propose a novel architecture incorporating an explicit private working memory; we demonstrate that this mechanism restores consistency, establishing private state as a necessary component for interactive language agents.

Method: Decomposes sequential questions into sub-questions for stepwise reasoning while processing direct questions directly; uses multi-source information retrieval and in-context learning for rich context.

Result: Achieved Exact Match score of 0.84 on BioCreative IX - MedHopQA Shared Task datasets, ranking second on the leaderboard.

Conclusion: The model effectively addresses biomedical QA challenges, offering a versatile solution for advancing medical research and practice through its dual approach to question handling.

Abstract: Biomedical Question Answering systems play a critical role in processing complex medical queries, yet they often struggle with the intricate nature of medical data and the demand for multi-hop reasoning. In this paper, we propose a model designed to effectively address both direct and sequential questions. While sequential questions are decomposed into a chain of sub-questions to perform reasoning across a chain of steps, direct questions are processed directly to ensure efficiency and minimise processing overhead. Additionally, we leverage multi-source information retrieval and in-context learning to provide rich, relevant context for generating answers. We evaluated our model on the BioCreative IX - MedHopQA Shared Task datasets. Our approach achieves an Exact Match score of 0.84, ranking second on the current leaderboard. These results highlight the model’s capability to meet the challenges of Biomedical Question Answering, offering a versatile solution for advancing medical research and practice.

Method: Uses a Retrieval-Augmented Generation (RAG) pipeline with hybrid retrieval from medical textbooks and academic literature (PubMed, Semantic Scholar APIs). Features state-of-the-art reranking to filter/order evidence, then LLM generates final educational content.

Result: Three radiologists assessed outputs as high clinical/educational value. Large-scale LLM-as-a-Judge evaluation showed moderate alignment with human expert judgments, indicating LLMs can help evaluate but still require expert oversight.

Conclusion: MedTutor successfully augments resident training by generating evidence-based educational content from case reports, though expert oversight remains necessary despite LLMs showing promise for evaluation.

Abstract: The learning process for medical residents presents significant challenges, demanding both the ability to interpret complex case reports and the rapid acquisition of accurate medical knowledge from reliable sources. Residents typically study case reports and engage in discussions with peers and mentors, but finding relevant educational materials and evidence to support their learning from these cases is often time-consuming and challenging. To address this, we introduce MedTutor, a novel system designed to augment resident training by automatically generating evidence-based educational content and multiple-choice questions from clinical case reports. MedTutor leverages a Retrieval-Augmented Generation (RAG) pipeline that takes clinical case reports as input and produces targeted educational materials. The system’s architecture features a hybrid retrieval mechanism that synergistically queries a local knowledge base of medical textbooks and academic literature (using PubMed, Semantic Scholar APIs) for the latest related research, ensuring the generated content is both foundationally sound and current. The retrieved evidence is filtered and ordered using a state-of-the-art reranking model and then an LLM generates the final long-form output describing the main educational content regarding the case-report. We conduct a rigorous evaluation of the system. First, three radiologists assessed the quality of outputs, finding them to be of high clinical and educational value. Second, we perform a large scale evaluation using an LLM-as-a Judge to understand if LLMs can be used to evaluate the output of the system. Our analysis using correlation between LLMs outputs and human expert judgments reveals a moderate alignment and highlights the continued necessity of expert oversight.

Method: Adapt transformer-based constituency parser to Middle Dutch. Use joint training with higher-resource auxiliary languages. Evaluate strategies for leveraging newly annotated data from additional domains (fine-tuning, data combination). Explore feature-separation techniques for domain adaptation.

Result: Joint training with auxiliary languages increases F1 scores by up to 0.73, with greatest gains from geographically/temporally closer languages. Fine-tuning and data combination yield comparable improvements. Neural parser consistently outperforms current PCFG-based parser. Approximately 200 examples per domain needed for effective cross-domain performance enhancement.

Conclusion: Transformer-based constituency parsing can be successfully adapted to low-resource historical languages like Middle Dutch. Joint training with related languages and strategic domain adaptation significantly improve performance, with neural approaches outperforming traditional PCFG-based methods.

Abstract: Recent years have seen growing interest in applying neural networks and contextualized word embeddings to the parsing of historical languages. However, most advances have focused on dependency parsing, while constituency parsing for low-resource historical languages like Middle Dutch has received little attention. In this paper, we adapt a transformer-based constituency parser to Middle Dutch, a highly heterogeneous and low-resource language, and investigate methods to improve both its in-domain and cross-domain performance. We show that joint training with higher-resource auxiliary languages increases F1 scores by up to 0.73, with the greatest gains achieved from languages that are geographically and temporally closer to Middle Dutch. We further evaluate strategies for leveraging newly annotated data from additional domains, finding that fine-tuning and data combination yield comparable improvements, and our neural parser consistently outperforms the currently used PCFG-based parser for Middle Dutch. We further explore feature-separation techniques for domain adaptation and demonstrate that a minimum threshold of approximately 200 examples per domain is needed to effectively enhance cross-domain performance.

Method: Developed TurkBench with 8,151 data samples organized into 21 distinct subtasks under six main categories: Knowledge, Language Understanding, Reasoning, Content Moderation, Turkish Grammar and Vocabulary, and Instruction Following.

Result: Created a comprehensive Turkish language benchmark with culturally relevant data that provides researchers and developers with a valuable evaluation tool for Turkish LLMs.

Conclusion: TurkBench addresses the gap in Turkish language model evaluation and is publicly available for online submissions, enabling better assessment and improvement of Turkish LLMs.

Abstract: With the recent surge in the development of large language models, the need for comprehensive and language-specific evaluation benchmarks has become critical. While significant progress has been made in evaluating English language models, benchmarks for other languages, particularly those with unique linguistic characteristics such as Turkish, remain less developed. Our study introduces TurkBench, a comprehensive benchmark designed to assess the capabilities of generative large language models in the Turkish language. TurkBench involves 8,151 data samples across 21 distinct subtasks. These are organized under six main categories of evaluation: Knowledge, Language Understanding, Reasoning, Content Moderation, Turkish Grammar and Vocabulary, and Instruction Following. The diverse range of tasks and the culturally relevant data would provide researchers and developers with a valuable tool for evaluating their models and identifying areas for improvement. We further publish our benchmark for online submissions at https://huggingface.co/turkbench

Method: Three-part methodology: 1) Synthesize 4.5T tokens of high-quality, domain-specific, RL-oriented data; 2) Progressive curriculum coordinating data composition, quality thresholds, and domain coverage across 20T tokens; 3) SnapPO framework for efficient RL optimization to enable reasoning capabilities.

Result: Solar Open achieves competitive performance across benchmarks in English and Korean, demonstrating effectiveness of the methodology for underserved language AI development.

Conclusion: The systematic methodology combining data synthesis, progressive curriculum, and efficient RL optimization enables building competitive LLMs for underserved languages, advancing AI development for languages with limited resources.

Abstract: We introduce Solar Open, a 102B-parameter bilingual Mixture-of-Experts language model for underserved languages. Solar Open demonstrates a systematic methodology for building competitive LLMs by addressing three interconnected challenges. First, to train effectively despite data scarcity for underserved languages, we synthesize 4.5T tokens of high-quality, domain-specific, and RL-oriented data. Second, we coordinate this data through a progressive curriculum jointly optimizing composition, quality thresholds, and domain coverage across 20 trillion tokens. Third, to enable reasoning capabilities through scalable RL, we apply our proposed framework SnapPO for efficient optimization. Across benchmarks in English and Korean, Solar Open achieves competitive performance, demonstrating the effectiveness of this methodology for underserved language AI development.

Method: CFPG reframes narrative quality through payoff realization, transforming narrative continuity into executable causal predicates. It mines and encodes Foreshadow-Trigger-Payoff triples from BookSum corpus to provide structured supervision ensuring foreshadowed commitments are temporally and logically fulfilled.

Result: CFPG significantly outperforms standard prompting baselines in payoff accuracy and narrative alignment. The framework demonstrates improved ability to ensure narrative commitments are properly resolved.

Conclusion: Explicitly codifying narrative mechanics is essential for moving LLMs from surface-level fluency to genuine narrative competence. Structured supervision of narrative dependencies addresses fundamental limitations in current LLM story generation.

Abstract: Foreshadowing and payoff are ubiquitous narrative devices through which authors introduce commitments early in a story and resolve them through concrete, observable outcomes. However, despite advances in story generation, large language models (LLMs) frequently fail to bridge these long-range narrative dependencies, often leaving “Chekhov’s guns” unfired even when the necessary context is present. Existing evaluations largely overlook this structural failure, focusing on surface-level coherence rather than the logical fulfillment of narrative setups. In this paper, we introduce Codified Foreshadowing-Payoff Generation (CFPG), a novel framework that reframes narrative quality through the lens of payoff realization. Recognizing that LLMs struggle to intuitively grasp the “triggering mechanism” of a foreshadowed event, CFPG transforms narrative continuity into a set of executable causal predicates. By mining and encoding Foreshadow-Trigger-Payoff triples from the BookSum corpus, we provide structured supervision that ensures foreshadowed commitments are not only mentioned but also temporally and logically fulfilled. Experiments demonstrate that CFPG significantly outperforms standard prompting baselines in payoff accuracy and narrative alignment. Our findings suggest that explicitly codifying narrative mechanics is essential for moving LLMs from surface-level fluency to genuine narrative competence.

Method: Through attention analysis and controlled prompting experiments, the authors identified specific triggers: a leading “Okay” token induces reasoning behavior, while the newline pattern following “” suppresses it. Based on these findings, they propose Mid-Think, a simple training-free prompting format that combines these triggers to achieve intermediate-budget reasoning.

Result: Mid-Think consistently outperforms fixed-token and prompt-based baselines in accuracy-length trade-off. When applied to RL training after SFT, it reduces training time by approximately 15% while improving Qwen3-8B performance from 69.8% to 72.4% on AIME and from 58.5% to 61.1% on GPQA.

Conclusion: The study demonstrates that hybrid reasoning models are controlled by specific trigger tokens rather than high-level instructions, and the proposed Mid-Think format effectively leverages these triggers for both inference-time control and RL-based reasoning training, improving both performance and efficiency.

Abstract: Hybrid reasoning language models are commonly controlled through high-level Think/No-think instructions to regulate reasoning behavior, yet we found that such mode switching is largely driven by a small set of trigger tokens rather than the instructions themselves. Through attention analysis and controlled prompting experiments, we show that a leading Okay'' token induces reasoning behavior, while the newline pattern following ’’ suppresses it. Based on this observation, we propose Mid-Think, a simple training-free prompting format that combines these triggers to achieve intermediate-budget reasoning, consistently outperforming fixed-token and prompt-based baselines in terms of the accuracy-length trade-off. Furthermore, applying Mid-Think to RL training after SFT reduces training time by approximately 15% while improving final performance of Qwen3-8B on AIME from 69.8% to 72.4% and on GPQA from 58.5% to 61.1%, demonstrating its effectiveness for both inference-time control and RL-based reasoning training.

Method: Treat language training as tasks, generate task vectors by fine-tuning Whisper ASR variants, merge vectors via linear combination optimized on target language’s word error rate.

Result: Consistent performance improvement on target low-resource languages.

Conclusion: Task arithmetic with linear combination of task vectors is effective for low-resource ASR by leveraging related high-resource language models.

Abstract: The development of resource-constrained approaches to automatic speech recognition (ASR) is of great interest due to its broad applicability to many low-resource languages for which there is scant usable data. Existing approaches to many low-resource natural language processing tasks leverage additional data from higher-resource languages that are closely related to a target low-resource language. One increasingly popular approach uses task arithmetic to combine models trained on different tasks to create a model for a task where there is little to no training data. In this paper, we consider training on a particular language to be a task, and we generate task vectors by fine-tuning variants of the Whisper ASR system. For pairings of high- and low-resource languages, we merge task vectors via a linear combination, optimizing the weights of the linear combination on the downstream word error rate on the low-resource target language’s validation set. We find that this approach consistently improves performance on the target languages.

Method: Introduces CLEAR framework that decomposes conflict resolution into four progressive scenarios (multilingual parametric elicitation to competitive multi-source cross-lingual induction). Evaluates six LLMs on multilingual versions of ConflictQA and ConflictingQA covering 10 typologically diverse languages.

Result: Reveals task-dependent decision dichotomy: reasoning-intensive tasks are dominated by language resource abundance (high-resource languages have stronger persuasive power), while entity-centric factual conflicts are decided by linguistic affinity (low-resource but linguistically aligned languages outperform distant high-resource ones).

Conclusion: Cross-lingual knowledge conflict resolution in multilingual LLMs follows distinct patterns based on task type, highlighting the complex interplay between language resources and linguistic relationships in model decision-making.

Abstract: Large Language Models (LLMs) encode vast world knowledge across multiple languages, yet their internal beliefs are often unevenly distributed across linguistic spaces. When external evidence contradicts these language-dependent memories, models encounter \emph{cross-lingual knowledge conflict}, a phenomenon largely unexplored beyond English-centric settings. We introduce \textbf{CLEAR}, a \textbf{C}ross-\textbf{L}ingual knowl\textbf{E}dge conflict ev\textbf{A}luation f\textbf{R}amework that systematically examines how multilingual LLMs reconcile conflicting internal beliefs and multilingual external evidence. CLEAR decomposes conflict resolution into four progressive scenarios, from multilingual parametric elicitation to competitive multi-source cross-lingual induction, and systematically evaluates model behavior across two complementary QA benchmarks with distinct task characteristics. We construct multilingual versions of ConflictQA and ConflictingQA covering 10 typologically diverse languages and evaluate six representative LLMs. Our experiments reveal a task-dependent decision dichotomy. In reasoning-intensive tasks, conflict resolution is dominated by language resource abundance, with high-resource languages exerting stronger persuasive power. In contrast, for entity-centric factual conflicts, linguistic affinity, not resource scale, becomes decisive, allowing low-resource but linguistically aligned languages to outperform distant high-resource ones.

Method: The paper recapitulates simple means of “probability engineering” - techniques that can be used to encourage generative AI to hallucinate to a limited extent, thereby achieving desired results.

Result: The analysis suggests that controlled hallucination can lead to better performance in generative AI systems, challenging the conventional view that hallucination is purely undesirable.

Conclusion: The paper concludes by questioning whether hallucination in generative AI is actually a feature rather than a bug, suggesting that limited, controlled hallucination may be essential for these systems to function effectively.

Abstract: Generative artificial intelligence (AI) is conquering our lives at lightning speed. Large language models such as ChatGPT answer our questions or write texts for us, large computer vision models such as GAIA-1 generate videos on the basis of text descriptions or continue prompted videos. These neural network models are trained using large amounts of text or video data, strictly according to the real data employed in training. However, there is a surprising observation: When we use these models, they only function satisfactorily when they are allowed a certain degree of fantasy (hallucination). While hallucination usually has a negative connotation in generative AI - after all, ChatGPT is expected to give a fact-based answer! - this article recapitulates some simple means of probability engineering that can be used to encourage generative AI to hallucinate to a limited extent and thus lead to the desired results. We have to ask ourselves: Is hallucination in gen-erative AI probably not a bug, but rather a feature?

Method: Systematically compare parametric (unsupervised fine-tuning/continual pretraining, supervised fine-tuning) and non-parametric (retrieval-augmented generation) knowledge injection methods across three 7B-parameter open-source LLMs on two benchmarks: QASC and a new dataset of 10,000+ multi-hop questions from 2024 Wikipedia events.

Result: Unsupervised fine-tuning provides limited gains; RAG yields substantial improvements especially for temporally novel information; supervised fine-tuning achieves highest overall accuracy across models and datasets.

Conclusion: Different knowledge injection mechanisms support multi-hop QA differently: RAG is crucial for external/compositional knowledge, supervised fine-tuning works best overall, and continual pretraining alone is insufficient for improving multi-hop reasoning accuracy.

Abstract: Multi-hop question answering is widely used to evaluate the reasoning capabilities of large language models (LLMs), as it requires integrating multiple pieces of supporting knowledge to arrive at a correct answer. While prior work has explored different mechanisms for providing knowledge to LLMs, such as finetuning and retrieval-augmented generation (RAG), their relative effectiveness for multi-hop question answering remains insufficiently understood, particularly when the required knowledge is temporally novel. In this paper, we systematically compare parametric and non-parametric knowledge injection methods for open-domain multi-hop question answering. We evaluate unsupervised fine-tuning (continual pretraining), supervised fine-tuning, and retrieval-augmented generation across three 7B-parameter open-source LLMs. Experiments are conducted on two benchmarks: QASC, a standard multi-hop science question answering dataset, and a newly constructed dataset of over 10,000 multi-hop questions derived from Wikipedia events in 2024, designed to test knowledge beyond the models’ pretraining cutoff. Our results show that unsupervised fine-tuning provides only limited gains over base models, suggesting that continual pretraining alone is insufficient for improving multi-hop reasoning accuracy. In contrast, retrieval-augmented generation yields substantial and consistent improvements, particularly when answering questions that rely on temporally novel information. Supervised fine-tuning achieves the highest overall accuracy across models and datasets. These findings highlight fundamental differences in how knowledge injection mechanisms support multi-hop question answering and underscore the importance of retrieval-based methods when external or compositional knowledge is required.

Method: Developed SCOPE framework using 141-item, two-hour sociopsychological protocol from 124 U.S. participants. Tested across 7 models, comparing demographic-only personas vs. sociopsychologically enriched personas. Evaluated on SimBench with 441 aligned questions.

Result: Demographic-only personas explain only ~1.5% of variance in human response similarity. Adding sociopsychological facets improves behavioral prediction and reduces over-accentuation. Non-demographic personas based on values and identity achieve strong alignment with lower bias. SCOPE personas outperform default prompting and NVIDIA Nemotron personas.

Conclusion: Persona quality depends on sociopsychological structure rather than demographic templates or summaries. Detailed psychological profiling is essential for creating accurate, less biased synthetic personas for social simulation.

Abstract: Synthetic personas are widely used to condition large language models (LLMs) for social simulation, yet most personas are still constructed from coarse sociodemographic attributes or summaries. We revisit persona creation by introducing SCOPE, a socially grounded framework for persona construction and evaluation, built from a 141-item, two-hour sociopsychological protocol collected from 124 U.S.-based participants. Across seven models, we find that demographic-only personas are a structural bottleneck: demographics explain only ~1.5% of variance in human response similarity. Adding sociopsychological facets improves behavioral prediction and reduces over-accentuation, and non-demographic personas based on values and identity achieve strong alignment with substantially lower bias. These trends generalize to SimBench (441 aligned questions), where SCOPE personas outperform default prompting and NVIDIA Nemotron personas, and SCOPE augmentation improves Nemotron-based personas. Our results indicate that persona quality depends on sociopsychological structure rather than demographic templates or summaries.

Method: Four-stage modular framework: 1) Wake (low-temperature baseline), 2) Dream (high-temperature exploration), 3) Judge (coarse filtering), 4) Re-wake (re-articulation into coherent outputs). Each stage uses independent LLMs for functional separation.

Result: ReMIND reliably induces semantic exploration while preserving stability; dream phase shows substantial semantic displacement; high-quality ideas emerge sporadically rather than as extremes along single metrics.

Conclusion: Serendipitous ideation in LLMs is a rare-event process best approached through system-level modular design that shapes conditions for valuable ideas to emerge and be stabilized, providing a framework for studying computational serendipity.

Abstract: Large language models (LLMs) are used not only for problem solving but also for creative ideation; however, eliciting serendipitous insights that are both novel and internally coherent remains difficult. While stochastic sampling promotes novelty, it often degrades consistency. Here, we propose ReMIND, a REM-inspired modular framework for ideation. ReMIND consists of four stages: wake, which generates a stable low-temperature semantic baseline; dream, which performs high-temperature exploratory generation; judge, which applies coarse evaluation to filter incoherent outputs and extract candidate ideas; and re-wake, which re-articulates selected ideas into coherent final outputs. By instantiating each stage as an independent LLM, ReMIND enables functional separation between exploration and consolidation. Parameter sweeps show that ReMIND reliably induces semantic exploration while preserving downstream stability. Embedding-based analyses confirm substantial semantic displacement during the dream phase, whereas external evaluations reveal that high-quality ideas emerge sporadically rather than as extrema along any single metric. These results suggest that serendipitous ideation in LLMs is a rare-event process best approached through system level design that shapes the conditions under which valuable ideas can emerge and be stabilized. ReMIND provides a general framework for studying the computational basis of serendipity and illustrates how modular LLM orchestration can bridge exploration and stabilization.

Method: Algorithmically generated puzzles combining factual temporal anchors with cross-cultural calendar relations, each admitting one or multiple valid solution dates. Evaluated 13 diverse LLMs with and without tools (web search, code interpreter).

Result: TimePuzzles effectively distinguishes iterative temporal reasoning capabilities: GPT-5 reaches only 49.3% accuracy, all other models below 31% without tools. Web search yields substantial gains, code interpreter shows mixed effects. Models perform much better when constraints are rewritten with explicit dates.

Conclusion: TimePuzzles provides a simple, cost-effective diagnostic for tool-augmented iterative temporal reasoning, revealing significant gaps in LLMs’ ability to reliably use tools for temporal reasoning tasks.

Abstract: We introduce TimePuzzles, a constraint-based date inference task for evaluating iterative temporal reasoning. Each puzzle combines factual temporal anchors with (cross-cultural) calendar relations, admits one or multiple valid solution dates, and is algorithmically generated for controlled, dynamic, and continual evaluation. Across 13 diverse LLMs, TimePuzzles well distinguishes their iterative temporal reasoning capabilities and remains challenging without tools: GPT-5 reaches only 49.3% accuracy and all other models stay below 31%, despite the dataset’s simplicity. Web search consistently yields substantial gains and using code interpreter shows mixed effects, but all models perform much better when constraints are rewritten with explicit dates, revealing a gap in reliable tool use. Overall, TimePuzzles presents a simple, cost-effective diagnostic for tool-augmented iterative temporal reasoning.

Method: Created CodeMixQA benchmark with 16 parallel code-switched language-pair variants spanning multiple regions and patterns, including original scripts and transliterations. Used this to analyze LLM reasoning behavior on QA tasks and systematically evaluated LLM-generated synthetic code-switched text for naturalness and semantic fidelity.

Result: Revealed persistent challenges in both reasoning and generation under code-switching conditions, uncovering limitations in current LLM capabilities for processing mixed-language inputs.

Conclusion: Provides actionable insights for building more robust multilingual LLMs and releases the dataset and code as open source to advance research in this area.

Abstract: Code-switching is a pervasive phenomenon in multilingual communication, yet the robustness of large language models (LLMs) in mixed-language settings remains insufficiently understood. In this work, we present a comprehensive evaluation of LLM capabilities in understanding, reasoning over, and generating code-switched text. We introduce CodeMixQA a novel benchmark with high-quality human annotations, comprising 16 diverse parallel code-switched language-pair variants that span multiple geographic regions and code-switching patterns, and include both original scripts and their transliterated forms. Using this benchmark, we analyze the reasoning behavior of LLMs on code-switched question-answering tasks, shedding light on how models process and reason over mixed-language inputs. We further conduct a systematic evaluation of LLM-generated synthetic code-switched text, focusing on both naturalness and semantic fidelity, and uncover key limitations in current generation capabilities. Our findings reveal persistent challenges in both reasoning and generation under code-switching conditions and provide actionable insights for building more robust multilingual LLMs. We release the dataset and code as open source.

Method: Structured Reasoning (SCR) framework decouples reasoning trajectories into explicit, evaluable components using Generate-Verify-Revise paradigm. Uses structured training data with Dynamic Termination Supervision and progressive two-stage reinforcement learning: first stage for generation/verification, second for revision.

Result: SCR substantially improves reasoning efficiency and self-verification across three backbone models, reducing output token length by up to 50% compared to existing reasoning paradigms.

Conclusion: SCR effectively addresses LLM reasoning inefficiency by structuring the reasoning process, enabling targeted supervision of critical reasoning abilities, and significantly reducing redundant computations while maintaining performance.

Abstract: Large language models (LLMs) achieve strong performance by generating long chains of thought, but longer traces always introduce redundant or ineffective reasoning steps. One typical behavior is that they often perform unnecessary verification and revisions even if they have reached the correct answers. This limitation stems from the unstructured nature of reasoning trajectories and the lack of targeted supervision for critical reasoning abilities. To address this, we propose Structured Reasoning (SCR), a framework that decouples reasoning trajectories into explicit, evaluable, and trainable components. We mainly implement SCR using a Generate-Verify-Revise paradigm. Specifically, we construct structured training data and apply Dynamic Termination Supervision to guide the model in deciding when to terminate reasoning. To avoid interference between learning signals for different reasoning abilities, we adopt a progressive two-stage reinforcement learning strategy: the first stage targets initial generation and self-verification, and the second stage focuses on revision. Extensive experiments on three backbone models show that SCR substantially improves reasoning efficiency and self-verification. Besides, compared with existing reasoning paradigms, it reduces output token length by up to 50%.

Method: Relink dynamically constructs query-specific evidence graphs by instantiating required facts from a latent relation pool to repair broken paths, and uses query-aware evaluation to jointly consider KG and latent relation candidates, actively discarding distractors.

Result: On five Open-Domain Question Answering benchmarks, Relink achieves average improvements of 5.4% in EM and 5.2% in F1 over leading GraphRAG baselines.

Conclusion: The ‘reason-and-construct’ paradigm with dynamic evidence graph construction is superior to static ‘build-then-reason’ approaches, effectively addressing incompleteness and distractor issues in GraphRAG.

Abstract: Graph-based Retrieval-Augmented Generation (GraphRAG) mitigates hallucinations in Large Language Models (LLMs) by grounding them in structured knowledge. However, current GraphRAG methods are constrained by a prevailing \textit{build-then-reason} paradigm, which relies on a static, pre-constructed Knowledge Graph (KG). This paradigm faces two critical challenges. First, the KG’s inherent incompleteness often breaks reasoning paths. Second, the graph’s low signal-to-noise ratio introduces distractor facts, presenting query-relevant but misleading knowledge that disrupts the reasoning process. To address these challenges, we argue for a \textit{reason-and-construct} paradigm and propose Relink, a framework that dynamically builds a query-specific evidence graph. To tackle incompleteness, \textbf{Relink} instantiates required facts from a latent relation pool derived from the original text corpus, repairing broken paths on the fly. To handle misleading or distractor facts, Relink employs a unified, query-aware evaluation strategy that jointly considers candidates from both the KG and latent relations, selecting those most useful for answering the query rather than relying on their pre-existence. This empowers Relink to actively discard distractor facts and construct the most faithful and precise evidence path for each query. Extensive experiments on five Open-Domain Question Answering benchmarks show that Relink achieves significant average improvements of 5.4% in EM and 5.2% in F1 over leading GraphRAG baselines, demonstrating the superiority of our proposed framework.

Method: Proposes MI-PRUN which uses mutual information to evaluate hidden state transitions and identify redundant blocks. Incorporates Data Processing Inequality to analyze relationships between contiguous blocks and individual blocks. Develops Fast-Block-Select algorithm for iterative block combination updates to achieve global optimality efficiently.

Result: Extensive experiments across various models and datasets demonstrate the method’s stability and effectiveness in compressing LLMs while maintaining performance.

Conclusion: MI-PRUN provides a stable and effective block pruning approach for LLMs that achieves globally optimal solutions with improved efficiency, addressing limitations of existing methods.

Abstract: Large Language Models (LLMs) have become indispensable across various domains, but this comes at the cost of substantial computational and memory resources. Model pruning addresses this by removing redundant components from models. In particular, block pruning can achieve significant compression and inference acceleration. However, existing block pruning methods are often unstable and struggle to attain globally optimal solutions. In this paper, we propose a mutual information based pruning method MI-PRUN for LLMs. Specifically, we leverages mutual information to identify redundant blocks by evaluating transitions in hidden states. Additionally, we incorporate the Data Processing Inequality (DPI) to reveal the relationship between the importance of entire contiguous blocks and that of individual blocks. Moreover, we develop the Fast-Block-Select algorithm, which iteratively updates block combinations to achieve a globally optimal solution while significantly improving the efficiency. Extensive experiments across various models and datasets demonstrate the stability and effectiveness of our method.

Method: The authors conduct a comprehensive literature survey organized around two key questions: 1) whether linguistic disparities arise from representation and allocation choices rather than inherent complexity, and 2) which design choices can reduce inequities. They examine linguistic features (orthography, morphology, lexical diversity, syntax, information density, typological distance) and link them to concrete modeling mechanisms.

Result: The survey finds that performance gaps often shrink when segmentation, encoding, and data exposure are normalized, suggesting that much apparent difficulty stems from current modeling choices rather than intrinsic linguistic complexity. The analysis reveals how different linguistic features interact with specific modeling mechanisms.

Conclusion: The paper synthesizes insights into design recommendations for tokenization, sampling, architectures, and evaluation to support more balanced multilingual language models. The findings suggest that with appropriate modeling choices, multilingual LMs can achieve more equitable performance across diverse languages.

Abstract: Multilingual language models (LMs) promise broader NLP access, yet current systems deliver uneven performance across the world’s languages. This survey examines why these gaps persist and whether they reflect intrinsic linguistic difficulty or modeling artifacts. We organize the literature around two questions: do linguistic disparities arise from representation and allocation choices (e.g., tokenization, encoding, data exposure, parameter sharing) rather than inherent complexity; and which design choices mitigate inequities across typologically diverse languages. We review linguistic features, such as orthography, morphology, lexical diversity, syntax, information density, and typological distance, linking each to concrete modeling mechanisms. Gaps often shrink when segmentation, encoding, and data exposure are normalized, suggesting much apparent difficulty stems from current modeling choices. We synthesize these insights into design recommendations for tokenization, sampling, architectures, and evaluation to support more balanced multilingual LMs.

Method: ActiShade iteratively: 1) detects overshadowed keyphrases in queries, 2) retrieves documents relevant to both query and overshadowed keyphrase, 3) generates new queries based on retrieved documents to guide next-round iteration, supplementing overshadowed knowledge while minimizing irrelevant noise.

Result: Extensive experiments show ActiShade outperforms existing methods across multiple datasets and LLMs.

Conclusion: ActiShade effectively addresses knowledge overshadowing in multi-hop reasoning by activating overshadowed knowledge, reducing error accumulation and improving retrieval quality in iterative RAG systems.

Abstract: In multi-hop reasoning, multi-round retrieval-augmented generation (RAG) methods typically rely on LLM-generated content as the retrieval query. However, these approaches are inherently vulnerable to knowledge overshadowing - a phenomenon where critical information is overshadowed during generation. As a result, the LLM-generated content may be incomplete or inaccurate, leading to irrelevant retrieval and causing error accumulation during the iteration process. To address this challenge, we propose ActiShade, which detects and activates overshadowed knowledge to guide large language models (LLMs) in multi-hop reasoning. Specifically, ActiShade iteratively detects the overshadowed keyphrase in the given query, retrieves documents relevant to both the query and the overshadowed keyphrase, and generates a new query based on the retrieved documents to guide the next-round iteration. By supplementing the overshadowed knowledge during the formulation of next-round queries while minimizing the introduction of irrelevant noise, ActiShade reduces the error accumulation caused by knowledge overshadowing. Extensive experiments show that ActiShade outperforms existing methods across multiple datasets and LLMs.

Method: Systematic investigation of verbalized calibration in tool-use agents, identifying confidence dichotomy by tool type. Proposed reinforcement learning fine-tuning framework that jointly optimizes task accuracy and calibration, supported by holistic benchmark of reward designs. Conducted pilot study and trained agents to evaluate generalization from local training to noisy web settings and distinct domains.

Result: Evidence tools (e.g., web search) systematically induce severe overconfidence due to inherent noise in retrieved information, while verification tools (e.g., code interpreters) can ground reasoning through deterministic feedback and mitigate miscalibration. RL-trained agents achieve superior calibration and exhibit robust generalization from local training environments to noisy web settings and distinct domains like mathematical reasoning.

Conclusion: Domain-specific calibration strategies are necessary for tool-use agents. The work establishes a foundation for building self-aware agents that can reliably communicate uncertainty in high-stakes, real-world deployments, highlighting the importance of tailored approaches for different tool types in agentic workflows.

Abstract: Autonomous agents based on large language models (LLMs) are rapidly evolving to handle multi-turn tasks, but ensuring their trustworthiness remains a critical challenge. A fundamental pillar of this trustworthiness is calibration, which refers to an agent’s ability to express confidence that reliably reflects its actual performance. While calibration is well-established for static models, its dynamics in tool-integrated agentic workflows remain underexplored. In this work, we systematically investigate verbalized calibration in tool-use agents, revealing a fundamental confidence dichotomy driven by tool type. Specifically, our pilot study identifies that evidence tools (e.g., web search) systematically induce severe overconfidence due to inherent noise in retrieved information, while verification tools (e.g., code interpreters) can ground reasoning through deterministic feedback and mitigate miscalibration. To robustly improve calibration across tool types, we propose a reinforcement learning (RL) fine-tuning framework that jointly optimizes task accuracy and calibration, supported by a holistic benchmark of reward designs. We demonstrate that our trained agents not only achieve superior calibration but also exhibit robust generalization from local training environments to noisy web settings and to distinct domains such as mathematical reasoning. Our results highlight the necessity of domain-specific calibration strategies for tool-use agents. More broadly, this work establishes a foundation for building self-aware agents that can reliably communicate uncertainty in high-stakes, real-world deployments.

Method: Proposes DocZSRE-SI framework that leverages Entity Side Information (Entity Mention Descriptions and Entity Mention Hypernyms) to perform zero-shot relation extraction without depending on LLM-generated synthetic data, using a low-complexity model.

Result: Achieves an average improvement of 11.6% in macro F1-Score compared to baseline models and existing benchmarks, demonstrating better performance for low-resource languages like Malaysian English.

Conclusion: DocZSRE-SI provides a robust, efficient, and scalable alternative to error-prone LLM-based methods, advancing relation extraction for low-resource languages and linguistic diversity, particularly in contexts like Malaysian English news articles.

Abstract: Document-Level Zero-Shot Relation Extraction (DocZSRE) aims to predict unseen relation labels in text documents without prior training on specific relations. Existing approaches rely on Large Language Models (LLMs) to generate synthetic data for unseen labels, which poses challenges for low-resource languages like Malaysian English. These challenges include the incorporation of local linguistic nuances and the risk of factual inaccuracies in LLM-generated data. This paper introduces Document-Level Zero-Shot Relation Extraction with Entity Side Information (DocZSRE-SI) to address limitations in the existing DocZSRE approach. The DocZSRE-SI framework leverages Entity Side Information, such as Entity Mention Descriptions and Entity Mention Hypernyms, to perform ZSRE without depending on LLM-generated synthetic data. The proposed low-complexity model achieves an average improvement of 11.6% in the macro F1-Score compared to baseline models and existing benchmarks. By utilizing Entity Side Information, DocZSRE-SI offers a robust and efficient alternative to error-prone, LLM-based methods, demonstrating significant advancements in handling low-resource languages and linguistic diversity in relation extraction tasks. This research provides a scalable and reliable solution for ZSRE, particularly in contexts like Malaysian English news articles, where traditional LLM-based approaches fall short.

Method: Train a speech encoder using only ASR (automatic speech recognition) data to achieve cross-dialect semantic alignment between Chinese dialects and Mandarin. Create a new benchmark for spoken Chinese varieties and evaluate using speech-to-speech retrieval.

Result: The speech encoder demonstrates cross-dialect semantic alignment through speech-to-speech retrieval on the new Chinese dialect benchmark. It also achieves state-of-the-art ASR performance on Chinese dialects.

Conclusion: The work provides three key contributions: a Chinese dialect benchmark, semantically aligned speech representations, and speech-to-speech retrieval evaluation, laying the groundwork for future Chinese dialect speech-LLMs. The benchmark is publicly released.

Abstract: Despite having hundreds of millions of speakers, Chinese dialects lag behind Mandarin in speech and language technologies. Most varieties are primarily spoken, making dialect-to-Mandarin speech-LLMs (large language models) more practical than dialect LLMs. Building dialect-to-Mandarin speech-LLMs requires speech representations with cross-dialect semantic alignment between Chinese dialects and Mandarin. In this paper, we achieve such a cross-dialect semantic alignment by training a speech encoder with ASR (automatic speech recognition)-only data, as demonstrated by speech-to-speech retrieval on a new benchmark of spoken Chinese varieties that we contribute. Our speech encoder further demonstrates state-of-the-art ASR performance on Chinese dialects. Together, our Chinese dialect benchmark, semantically aligned speech representations, and speech-to-speech retrieval evaluation lay the groundwork for future Chinese dialect speech-LLMs. We release the benchmark at https://github.com/kalvinchang/yubao.

Method: 1) Created ReasonTabQA benchmark with 1,932 tables across 30 industry domains, featuring bilingual annotations for answers and reasoning chains. 2) Developed TabCodeRL, a reinforcement learning method using table-aware verifiable rewards to guide logical reasoning path generation.

Result: TabCodeRL shows substantial performance gains on open-source LLMs across ReasonTabQA and 4 other TableQA datasets. However, there’s still a persistent performance gap on ReasonTabQA, highlighting the inherent complexity of real-world industrial TableQA.

Conclusion: Industrial TableQA presents unique challenges that current benchmarks and methods don’t adequately address. The ReasonTabQA benchmark reveals significant gaps in handling real-world complexity, suggesting need for more sophisticated reasoning approaches.

Abstract: Recent advancements in Large Language Models (LLMs) have significantly catalyzed table-based question answering (TableQA). However, existing TableQA benchmarks often overlook the intricacies of industrial scenarios, which are characterized by multi-table structures, nested headers, and massive scales. These environments demand robust table reasoning through deep structured inference, presenting a significant challenge that remains inadequately addressed by current methodologies. To bridge this gap, we present ReasonTabQA, a large-scale bilingual benchmark encompassing 1,932 tables across 30 industry domains such as energy and automotive. ReasonTabQA provides high-quality annotations for both final answers and explicit reasoning chains, supporting both thinking and no-thinking paradigms. Furthermore, we introduce TabCodeRL, a reinforcement learning method that leverages table-aware verifiable rewards to guide the generation of logical reasoning paths. Extensive experiments on ReasonTabQA and 4 TableQA datasets demonstrate that while TabCodeRL yields substantial performance gains on open-source LLMs, the persistent performance gap on ReasonTabQA underscores the inherent complexity of real-world industrial TableQA.

Method: Propose PsyCLIENT framework grounded in conversational trajectory modeling, conditioning LLM generation on predefined real-world trajectories with explicit behavior labels and content constraints. Also introduce PsyCLIENT-CP, the first open-source Chinese client profile dataset covering 60 distinct counseling topics.

Result: Significantly outperforms baselines in authenticity and training effectiveness. Simulated clients are nearly indistinguishable from human clients, achieving ~95% expert confusion rate in discrimination tasks. Comprehensive evaluations by licensed professional counselors validate effectiveness.

Conclusion: Conversational trajectory modeling effectively bridges gap between theoretical client profiles and dynamic, realistic simulations, offering robust solution for mental health education and research. Code and data will be released to facilitate future research.

Abstract: LLM-based client simulation has emerged as a promising tool for training novice counselors and evaluating automated counseling systems. However, existing client simulation approaches face three key challenges: (1) limited diversity and realism in client profiles, (2) the lack of a principled framework for modeling realistic client behaviors, and (3) a scarcity in Chinese-language settings. To address these limitations, we propose PsyCLIENT, a novel simulation framework grounded in conversational trajectory modeling. By conditioning LLM generation on predefined real-world trajectories that incorporate explicit behavior labels and content constraints, our approach ensures diverse and realistic interactions. We further introduce PsyCLIENT-CP, the first open-source Chinese client profile dataset, covering 60 distinct counseling topics. Comprehensive evaluations involving licensed professional counselors demonstrate that PsyCLIENT significantly outperforms baselines in terms of authenticity and training effectiveness. Notably, the simulated clients are nearly indistinguishable from human clients, achieving an about 95% expert confusion rate in discrimination tasks. These findings indicate that conversational trajectory modeling effectively bridges the gap between theoretical client profiles and dynamic, realistic simulations, offering a robust solution for mental health education and research. Code and data will be released to facilitate future research in mental health counseling.

Method: Created Mitrasamgraha dataset with 391,548 Sanskrit-English bitext pairs (4x larger than previous largest), covering 3+ millennia and broad historical domains, with temporal/domain annotations. Also released validation (5,587) and test (5,552) sets with post-correction. Benchmarked commercial/open models, fine-tuned NLLB and Gemma models, and analyzed in-context learning effects.

Result: Dataset enables fine-grained study of domain/time period effects on MT performance. Fine-tuning NLLB and Gemma models showed significant improvements, though challenges remain in translating complex compounds, philosophical concepts, and multi-layered metaphors. In-context learning analysis revealed performance impacts on commercial models.

Conclusion: Mitrasamgraha addresses the resource gap for Sanskrit MT and enables detailed analysis of temporal/domain effects, showing that while progress is possible through fine-tuning, significant challenges remain for complex Sanskrit content, highlighting that Sanskrit MT is far from a “solved problem.”

Abstract: While machine translation is regarded as a “solved problem” for many high-resource languages, close analysis quickly reveals that this is not the case for content that shows challenges such as poetic language, philosophical concepts, multi-layered metaphorical expressions, and more. Sanskrit literature is a prime example of this, as it combines a large number of such challenges in addition to inherent linguistic features like sandhi, compounding, and heavy morphology, which further complicate NLP downstream tasks. It spans multiple millennia of text production time as well as a large breadth of different domains, ranging from ritual formulas via epic narratives, philosophical treatises, poetic verses up to scientific material. As of now, there is a strong lack of publicly available resources that cover these different domains and temporal layers of Sanskrit. We therefore introduce Mitrasamgraha, a high-quality Sanskrit-to-English machine translation dataset consisting of 391,548 bitext pairs, more than four times larger than the largest previously available Sanskrit dataset Itih=asa. It covers a time period of more than three millennia and a broad range of historical Sanskrit domains. In contrast to web-crawled datasets, the temporal and domain annotation of this dataset enables fine-grained study of domain and time period effects on MT performance. We also release a validation set consisting of 5,587 and a test set consisting of 5,552 post-corrected bitext pairs. We conduct experiments benchmarking commercial and open models on this dataset and fine-tune NLLB and Gemma models on the dataset, showing significant improvements, while still recognizing significant challenges in the translation of complex compounds, philosophical concepts, and multi-layered metaphors. We also analyze how in-context learning on this dataset impacts the performance of commercial models

Method: Step-by-step workflow using 1029 short stories: text preprocessing, network construction (co-occurrence vs TFMNs), feature extraction (structural measures, spreading-activation indices, emotion scores), and regression modeling to predict creativity ratings.

Result: TFMNs consistently outperformed co-occurrence networks with lower prediction errors (best MAE = 0.581 for TFMN vs 0.592 for co-occurrence). Network-structural features performed best (MAE = 0.591), emotion features worse (MAE = 0.711), and spreading-activation measures contributed little (MAE = 0.788).

Conclusion: The paper provides an open, reproducible workflow for network-based creativity analysis, showing when syntactic networks (TFMNs) are preferable to surface co-occurrence models, offering guidance for both newcomers and experienced researchers in cognitive network analysis.

Abstract: This tutorial paper provides a step-by-step workflow for building and analysing semantic networks from short creative texts. We introduce and compare two widely used text-to-network approaches: word co-occurrence networks and textual forma mentis networks (TFMNs). We also demonstrate how they can be used in machine learning to predict human creativity ratings. Using a corpus of 1029 short stories, we guide readers through text preprocessing, network construction, feature extraction (structural measures, spreading-activation indices, and emotion scores), and application of regression models. We evaluate how network-construction choices influence both network topology and predictive performance. Across all modelling settings, TFMNs consistently outperformed co-occurrence networks through lower prediction errors (best MAE = 0.581 for TFMN, vs 0.592 for co-occurrence with window size 3). Network-structural features dominated predictive performance (MAE = 0.591 for TFMN), whereas emotion features performed worse (MAE = 0.711 for TFMN) and spreading-activation measures contributed little (MAE = 0.788 for TFMN). This paper offers practical guidance for researchers interested in applying network-based methods for cognitive fields like creativity research. we show when syntactic networks are preferable to surface co-occurrence models, and provide an open, reproducible workflow accessible to newcomers in the field, while also offering deeper methodological insight for experienced researchers.

Method: BayesRAG uses Bayesian inference and Dempster-Shafer evidence theory to model intrinsic consistency of retrieved candidates across modalities as probabilistic evidence. It computes posterior association probability for multimodal retrieval result combinations, prioritizing text-image pairs that mutually corroborate in terms of semantics and layout.

Result: Extensive experiments show BayesRAG significantly outperforms state-of-the-art methods on challenging multimodal benchmarks. The framework establishes a new paradigm for multimodal retrieval fusion that effectively resolves the isolation of heterogeneous modalities.

Conclusion: BayesRAG enhances retrieval robustness through evidence fusion mechanism, providing a novel approach to multimodal RAG that better handles visually rich documents by leveraging cross-modal consistency and layout coherence.

Abstract: Retrieval-Augmented Generation (RAG) has become a pivotal paradigm for Large Language Models (LLMs), yet current approaches struggle with visually rich documents by treating text and images as isolated retrieval targets. Existing methods relying solely on cosine similarity often fail to capture the semantic reinforcement provided by cross-modal alignment and layout-induced coherence. To address these limitations, we propose BayesRAG, a novel multimodal retrieval framework grounded in Bayesian inference and Dempster-Shafer evidence theory. Unlike traditional approaches that rank candidates strictly by similarity, BayesRAG models the intrinsic consistency of retrieved candidates across modalities as probabilistic evidence to refine retrieval confidence. Specifically, our method computes the posterior association probability for combinations of multimodal retrieval results, prioritizing text-image pairs that mutually corroborate each other in terms of both semantics and layout. Extensive experiments demonstrate that BayesRAG significantly outperforms state-of-the-art (SOTA) methods on challenging multimodal benchmarks. This study establishes a new paradigm for multimodal retrieval fusion that effectively resolves the isolation of heterogeneous modalities through an evidence fusion mechanism and enhances the robustness of retrieval outcomes. Our code is available at https://github.com/TioeAre/BayesRAG.

Method: 1) Created MENT dataset covering four non-literal translation domains with 7,530 human-annotated scores; 2) Proposed RATE framework with a reflective Core Agent that dynamically invokes specialized sub-agents for translation evaluation; 3) Conducted experiments comparing RATE against traditional MT metrics and LLM-as-a-Judge approaches.

Result: Traditional MT metrics show inaccuracies, and LLM-as-a-Judge has limitations including knowledge cutoff and score inconsistency. RATE achieves at least 3.2 meta score improvement over current metrics and demonstrates robustness for general-domain MT evaluation.

Conclusion: The paper addresses critical limitations in MT evaluation for non-literal translations through a novel dataset and agentic framework. RATE provides a more reliable approach to translation quality assessment, with demonstrated effectiveness across both specialized and general translation domains.

Abstract: Large Language Models (LLMs) have significantly advanced Machine Translation (MT), applying them to linguistically complex domains-such as Social Network Services, literature etc. In these scenarios, translations often require handling non-literal expressions, leading to the inaccuracy of MT metrics. To systematically investigate the reliability of MT metrics, we first curate a meta-evaluation dataset focused on non-literal translations, namely MENT. MENT encompasses four non-literal translation domains and features source sentences paired with translations from diverse MT systems, with 7,530 human-annotated scores on translation quality. Experimental results reveal the inaccuracies of traditional MT metrics and the limitations of LLM-as-a-Judge, particularly the knowledge cutoff and score inconsistency problem. To mitigate these limitations, we propose RATE, a novel agentic translation evaluation framework, centered by a reflective Core Agent that dynamically invokes specialized sub-agents. Experimental results indicate the efficacy of RATE, achieving an improvement of at least 3.2 meta score compared with current metrics. Further experiments demonstrate the robustness of RATE to general-domain MT evaluation. Code and dataset are available at: https://github.com/BITHLP/RATE.

Method: Proposed Diffusion Entity-Relation Modeling (DiffER) with three key components: 1) whole-entity masking to prevent entity fragmentation, 2) distribution-symmetric data construction to address data asymmetry, and 3) relation-enhanced data construction to handle missing entity relations.

Result: Extensive experiments show that DiffER effectively alleviates the reversal curse in Diffusion LLMs, demonstrating improved bidirectional reasoning capabilities compared to baseline approaches.

Conclusion: The reversal curse has deeper causes than just training directionality. DiffER provides a solution through entity-aware training and balanced data construction, offering new perspectives for future research on bidirectional reasoning in language models.

Abstract: The “reversal curse” refers to the phenomenon where large language models (LLMs) exhibit predominantly unidirectional behavior when processing logically bidirectional relationships. Prior work attributed this to autoregressive training – predicting the next token inherently favors left-to-right information flow over genuine bidirectional knowledge associations. However, we observe that Diffusion LLMs (DLLMs), despite being trained bidirectionally, also suffer from the reversal curse. To investigate the root causes, we conduct systematic experiments on DLLMs and identify three key reasons: 1) entity fragmentation during training, 2) data asymmetry, and 3) missing entity relations. Motivated by the analysis of these reasons, we propose Diffusion Entity-Relation Modeling (DiffER), which addresses the reversal curse through entity-aware training and balanced data construction. Specifically, DiffER introduces whole-entity masking, which mitigates entity fragmentation by predicting complete entities in a single step. DiffER further employs distribution-symmetric and relation-enhanced data construction strategies to alleviate data asymmetry and missing relations. Extensive experiments demonstrate that DiffER effectively alleviates the reversal curse in Diffusion LLMs, offering new perspectives for future research.

Method: CSE consists of three components: Diversified Planning Initialization for broad solution space coverage, Genetic Evolution with feedback-guided mutation and compositional crossover, and Hierarchical Evolution Memory capturing both successful and failed experiences at inter-task and intra-task levels.

Result: CSE consistently outperforms all baselines across various LLM backbones on EffiBench-X, achieves higher efficiency from early generations, and maintains continuous improvement throughout evolution.

Conclusion: CSE effectively addresses exploration inefficiency in self-evolution methods through controlled mechanisms, demonstrating superior performance and efficiency in code generation tasks.

Abstract: Self-evolution methods enhance code generation through iterative “generate-verify-refine” cycles, yet existing approaches suffer from low exploration efficiency, failing to discover solutions with superior complexity within limited budgets. This inefficiency stems from initialization bias trapping evolution in poor solution regions, uncontrolled stochastic operations lacking feedback guidance, and insufficient experience utilization across tasks.To address these bottlenecks, we propose Controlled Self-Evolution (CSE), which consists of three key components. Diversified Planning Initialization generates structurally distinct algorithmic strategies for broad solution space coverage. Genetic Evolution replaces stochastic operations with feedback-guided mechanisms, enabling targeted mutation and compositional crossover. Hierarchical Evolution Memory captures both successful and failed experiences at inter-task and intra-task levels.Experiments on EffiBench-X demonstrate that CSE consistently outperforms all baselines across various LLM backbones. Furthermore, CSE achieves higher efficiency from early generations and maintains continuous improvement throughout evolution. Our code is publicly available at https://github.com/QuantaAlpha/EvoControl.

Method: Propose RM-NLHF using natural language feedback for process rewards, computing similarity between GRM-generated and human critiques. Introduce MetaRM to learn process reward prediction from datasets with human critiques and generalize to data without them.

Result: Experiments on multiple benchmarks show consistent outperformance over state-of-the-art GRMs trained with outcome-only reward, confirming superiority of natural language feedback over binary supervision.

Conclusion: Natural language feedback provides more accurate reward signals than binary outcomes, and MetaRM enables scalable process reward modeling by learning from limited human critiques.

Abstract: Reinforcement Learning with Verifiable reward (RLVR) on preference data has become the mainstream approach for training Generative Reward Models (GRMs). Typically in pairwise rewarding tasks, GRMs generate reasoning chains ending with critiques and preference labels, and RLVR then relies on the correctness of the preference labels as the training reward. However, in this paper, we demonstrate that such binary classification tasks make GRMs susceptible to guessing correct outcomes without sound critiques. Consequently, these spurious successes introduce substantial noise into the reward signal, thereby impairing the effectiveness of reinforcement learning. To address this issue, we propose Reward Modeling from Natural Language Human Feedback (RM-NLHF), which leverages natural language feedback to obtain process reward signals, thereby mitigating the problem of limited solution space inherent in binary tasks. Specifically, we compute the similarity between GRM-generated and human critiques as the training reward, which provides more accurate reward signals than outcome-only supervision. Additionally, considering that human critiques are difficult to scale up, we introduce Meta Reward Model (MetaRM) which learns to predict process reward from datasets with human critiques and then generalizes to data without human critiques. Experiments on multiple benchmarks demonstrate that our method consistently outperforms state-of-the-art GRMs trained with outcome-only reward, confirming the superiority of integrating natural language over binary human feedback as supervision.

Method: Proposes EvoToken-DLM that replaces hard binary masks with evolving soft token distributions, enabling progressive transition from masked states to discrete outputs with continuous trajectory supervision for training alignment.

Result: Extensive experiments across multiple benchmarks show EvoToken-DLM consistently achieves superior performance, outperforming strong diffusion-based and masked DLM baselines.

Conclusion: EvoToken-DLM offers a more effective diffusion-based language modeling approach through soft token evolution and continuous supervision, enabling revisable decoding and better utilization of intermediate representations.

Abstract: Diffusion Language Models (DLMs) offer a promising alternative for language modeling by enabling parallel decoding through iterative refinement. However, most DLMs rely on hard binary masking and discrete token assignments, which hinder the revision of early decisions and underutilize intermediate probabilistic representations. In this paper, we propose EvoToken-DLM, a novel diffusion-based language modeling approach that replaces hard binary masks with evolving soft token distributions. EvoToken-DLM enables a progressive transition from masked states to discrete outputs, supporting revisable decoding. To effectively support this evolution, we introduce continuous trajectory supervision, which aligns training objectives with iterative probabilistic updates. Extensive experiments across multiple benchmarks show that EvoToken-DLM consistently achieves superior performance, outperforming strong diffusion-based and masked DLM baselines. Project webpage: https://aim-uofa.github.io/EvoTokenDLM.

Method: TALON constructs draft trees iteratively until a fixed token budget is met, using a hybrid expansion strategy that adaptively allocates node budget to each layer, creating deep-and-narrow trees for deterministic contexts and shallow-and-wide trees for uncertain branches.

Result: Extensive experiments across 5 models and 6 datasets show TALON consistently outperforms state-of-the-art EAGLE-3, achieving up to 5.16x end-to-end speedup over auto-regressive decoding.

Conclusion: TALON provides an effective plug-and-play framework for adaptive tree expansion in speculative decoding that optimizes the trade-off between exploration width and generation depth under budget constraints.

Abstract: Speculative decoding (SD) has become a standard technique for accelerating LLM inference without sacrificing output quality. Recent advances in speculative decoding have shifted from sequential chain-based drafting to tree-structured generation, where the draft model constructs a tree of candidate tokens to explore multiple possible drafts in parallel. However, existing tree-based SD methods typically build a fixed-width, fixed-depth draft tree, which fails to adapt to the varying difficulty of tokens and contexts. As a result, the draft model cannot dynamically adjust the tree structure to early stop on difficult tokens and extend generation for simple ones. To address these challenges, we introduce TALON, a training-free, budget-driven adaptive tree expansion framework that can be plugged into existing tree-based methods. Unlike static methods, TALON constructs the draft tree iteratively until a fixed token budget is met, using a hybrid expansion strategy that adaptively allocates the node budget to each layer of the draft tree. This framework naturally shapes the draft tree into a “deep-and-narrow” form for deterministic contexts and a “shallow-and-wide” form for uncertain branches, effectively optimizing the trade-off between exploration width and generation depth under a given budget. Extensive experiments across 5 models and 6 datasets demonstrate that TALON consistently outperforms state-of-the-art EAGLE-3, achieving up to 5.16x end-to-end speedup over auto-regressive decoding.

Method: MetaGlyph encodes instructions as mathematical symbols (∈ for membership, ⇒ for implication) instead of prose, leveraging models’ pre-existing understanding from training data without requiring explicit decoding rules.

Result: Achieved 62-81% token reduction across all task types. Performance varied by model: Gemini 2.5 Flash (75% semantic equivalence, 49.9% membership fidelity), Kimi K2 (98.1% implication fidelity, 100% accuracy), GPT-5.2 Chat (91.3% membership fidelity), Claude Haiku 4.5 (100% parse success, 26% membership fidelity), Qwen 2.5 7B (62% equivalence). Mid-sized models (7B-12B) showed near-zero operator fidelity.

Conclusion: MetaGlyph effectively compresses prompts using mathematical symbols, with performance showing a U-shaped relationship with model scale - sufficient scale overcomes instruction-tuning biases, making symbolic prompts viable for large models but challenging for mid-sized ones.

Abstract: We introduce MetaGlyph, a symbolic language for compressing prompts by encoding instructions as mathematical symbols rather than prose. Unlike systems requiring explicit decoding rules, MetaGlyph uses symbols like $\in$ (membership) and $\Rightarrow$ (implication) that models already understand from their training data. We test whether these symbols work as ‘‘instruction shortcuts’’ that models can interpret without additional teaching. We evaluate eight models across two dimensions relevant to practitioners: scale (3B-1T parameters) and accessibility (open-source for local deployment vs. proprietary APIs). MetaGlyph achieves 62-81% token reduction across all task types. For API-based deployments, this translates directly to cost savings; for local deployments, it reduces latency and memory pressure. Results vary by model. Gemini 2.5 Flash achieves 75% semantic equivalence between symbolic and prose instructions on selection tasks, with 49.9% membership operator fidelity. Kimi K2 reaches 98.1% fidelity for implication ($\Rightarrow$) and achieves perfect (100%) accuracy on selection tasks with symbolic prompts. GPT-5.2 Chat shows the highest membership fidelity observed (91.3%), though with variable parse success across task types. Claude Haiku 4.5 achieves 100% parse success with 26% membership fidelity. Among mid-sized models, Qwen 2.5 7B shows 62% equivalence on extraction tasks. Mid-sized open-source models (7B-12B) show near-zero operator fidelity, suggesting a U-shaped relationship where sufficient scale overcomes instruction-tuning biases.

Method: The researchers used various machine learning models on a binary classification task comparing human-written novel excerpts with similar LLM-generated text. They employed simple features (single-token unigrams) and short text samples. They specifically used an inherently interpretable linear classifier to analyze feature importance and understand the underlying patterns.

Result: Machine learning models achieved 0.93-0.98 accuracy on unseen test data, significantly outperforming human observers who performed near chance levels. The interpretable linear classifier (98% accuracy) revealed that LLMs tend to use a larger variety of synonyms, creating detectable probability distributions. Four additional explanatory categories were identified: temporal drift, Americanisms, foreign language usage, and colloquialisms.

Conclusion: The classification is robust because it depends on multiple linguistic features working together, making it difficult for malicious actors to circumvent. The findings demonstrate that while humans struggle to distinguish AI-generated fiction, machine learning can reliably detect it using simple features, with LLMs’ tendency to use more synonyms being a key distinguishing factor.

Abstract: We consider the problem of distinguishing human-written creative fiction (excerpts from novels) from similar text generated by an LLM. Our results show that, while human observers perform poorly (near chance levels) on this binary classification task, a variety of machine-learning models achieve accuracy in the range 0.93 - 0.98 over a previously unseen test set, even using only short samples and single-token (unigram) features. We therefore employ an inherently interpretable (linear) classifier (with a test accuracy of 0.98), in order to elucidate the underlying reasons for this high accuracy. In our analysis, we identify specific unigram features indicative of LLM-generated text, one of the most important being that the LLM tends to use a larger variety of synonyms, thereby skewing the probability distributions in a manner that is easy to detect for a machine learning classifier, yet very difficult for a human observer. Four additional explanation categories were also identified, namely, temporal drift, Americanisms, foreign language usage, and colloquialisms. As identification of the AI-generated text depends on a constellation of such features, the classification appears robust, and therefore not easy to circumvent by malicious actors intent on misrepresenting AI-generated text as human work.

Method: Introduces Engram, a conditional memory module based on modernized N-gram embeddings for O(1) lookup. Formulates the Sparsity Allocation problem and discovers a U-shaped scaling law to optimize trade-off between neural computation (MoE) and static memory (Engram).

Result: Scales Engram to 27B parameters, outperforming iso-parameter/iso-FLOPs MoE baselines. Shows gains not only in knowledge retrieval (MMLU +3.4, CMMLU +4.0) but also in general reasoning (BBH +5.0, ARC-Challenge +3.7) and code/math (HumanEval +3.0, MATH +2.4). Improves long-context retrieval (Multi-Query NIAH: 84.2 to 97.0) with infrastructure-aware efficiency via deterministic addressing for prefetching.

Conclusion: Conditional memory (via Engram) is an indispensable modeling primitive for next-generation sparse models, relieving early layers from static reconstruction, deepening networks for complex reasoning, and freeing attention capacity for global context.

Abstract: While Mixture-of-Experts (MoE) scales capacity via conditional computation, Transformers lack a native primitive for knowledge lookup, forcing them to inefficiently simulate retrieval through computation. To address this, we introduce conditional memory as a complementary sparsity axis, instantiated via Engram, a module that modernizes classic $N$-gram embedding for O(1) lookup. By formulating the Sparsity Allocation problem, we uncover a U-shaped scaling law that optimizes the trade-off between neural computation (MoE) and static memory (Engram). Guided by this law, we scale Engram to 27B parameters, achieving superior performance over a strictly iso-parameter and iso-FLOPs MoE baseline. Most notably, while the memory module is expected to aid knowledge retrieval (e.g., MMLU +3.4; CMMLU +4.0), we observe even larger gains in general reasoning (e.g., BBH +5.0; ARC-Challenge +3.7) and code/math domains~(HumanEval +3.0; MATH +2.4). Mechanistic analyses reveal that Engram relieves the backbone’s early layers from static reconstruction, effectively deepening the network for complex reasoning. Furthermore, by delegating local dependencies to lookups, it frees up attention capacity for global context, substantially boosting long-context retrieval (e.g., Multi-Query NIAH: 84.2 to 97.0). Finally, Engram establishes infrastructure-aware efficiency: its deterministic addressing enables runtime prefetching from host memory, incurring negligible overhead. We envision conditional memory as an indispensable modeling primitive for next-generation sparse models.

Method: GROKE uses OpenStreetMap data in structured JSON/textual formats, employing a hierarchical architecture combining sub-instruction planning with topological graph navigation, without requiring visual simulators or training.

Result: Structured spatial formats outperform grid/visual graph representations; hierarchical architecture reduces navigation error by 68.5% compared to baselines on Map2Seq dataset; provides scalable, interpretable evaluation without visual dependencies.

Conclusion: GROKE establishes a scalable, interpretable evaluation paradigm for navigation instructions using OSM data, eliminating visual dependencies while capturing functional navigability through execution success, trajectory fidelity, and decision patterns.

Abstract: The evaluation of navigation instructions remains a persistent challenge in Vision-and-Language Navigation (VLN) research. Traditional reference-based metrics such as BLEU and ROUGE fail to capture the functional utility of spatial directives, specifically whether an instruction successfully guides a navigator to the intended destination. Although existing VLN agents could serve as evaluators, their reliance on high-fidelity visual simulators introduces licensing constraints and computational costs, and perception errors further confound linguistic quality assessment. This paper introduces GROKE(Graph-based Reasoning over OSM Knowledge for instruction Evaluation), a vision-free training-free hierarchical LLM-based framework for evaluating navigation instructions using OpenStreetMap data. Through systematic ablation studies, we demonstrate that structured JSON and textual formats for spatial information substantially outperform grid-based and visual graph representations. Our hierarchical architecture combines sub-instruction planning with topological graph navigation, reducing navigation error by 68.5% compared to heuristic and sampling baselines on the Map2Seq dataset. The agent’s execution success, trajectory fidelity, and decision patterns serve as proxy metrics for functional navigability given OSM-visible landmarks and topology, establishing a scalable and interpretable evaluation paradigm without visual dependencies. Code and data are available at https://anonymous.4open.science/r/groke.

Method: OAR uses two strategies: OAR-P estimates outcome sensitivity through counterfactual token perturbations for high-fidelity attribution; OAR-G uses input-gradient sensitivity as a proxy with single backward pass. Both integrate with conservative Bi-Level advantage reshaping that suppresses low-impact tokens and boosts pivotal ones while preserving overall advantage mass.

Result: On extensive mathematical reasoning benchmarks, OAR-P sets the performance upper bound, while OAR-G achieves comparable gains with negligible computational overhead. Both significantly outperform strong GRPO baselines.

Conclusion: OAR advances critic-free LLM reasoning by enabling fine-grained credit assignment, addressing limitations of uniform reward propagation in GRPO and pushing boundaries of critic-free reinforcement learning for reasoning tasks.

Abstract: Group Relative Policy Optimization (GRPO) has emerged as a promising critic-free reinforcement learning paradigm for reasoning tasks. However, standard GRPO employs a coarse-grained credit assignment mechanism that propagates group-level rewards uniformly to to every token in a sequence, neglecting the varying contribution of individual reasoning steps. We address this limitation by introducing Outcome-grounded Advantage Reshaping (OAR), a fine-grained credit assignment mechanism that redistributes advantages based on how much each token influences the model’s final answer. We instantiate OAR via two complementary strategies: (1) OAR-P, which estimates outcome sensitivity through counterfactual token perturbations, serving as a high-fidelity attribution signal; (2) OAR-G, which uses an input-gradient sensitivity proxy to approximate the influence signal with a single backward pass. These importance signals are integrated with a conservative Bi-Level advantage reshaping scheme that suppresses low-impact tokens and boosts pivotal ones while preserving the overall advantage mass. Empirical results on extensive mathematical reasoning benchmarks demonstrate that while OAR-P sets the performance upper bound, OAR-G achieves comparable gains with negligible computational overhead, both significantly outperforming a strong GRPO baseline, pushing the boundaries of critic-free LLM reasoning.

Method: Used attention knockout and token patching to validate and disentangle two truthfulness pathways. Conducted experiments to uncover properties of these mechanisms and their association with LLM knowledge boundaries.

Result: Identified two distinct information pathways for truthfulness encoding: Question-Anchored (depends on question-answer information flow) and Answer-Anchored (derives self-contained evidence from generated answer). Found these mechanisms are closely associated with LLM knowledge boundaries and internal representations are aware of their distinctions.

Conclusion: The work provides new insight into how LLMs internally encode truthfulness, offering directions for more reliable and self-aware generative systems, with proposed applications to enhance hallucination detection performance.

Abstract: Despite their impressive capabilities, large language models (LLMs) frequently generate hallucinations. Previous work shows that their internal states encode rich signals of truthfulness, yet the origins and mechanisms of these signals remain unclear. In this paper, we demonstrate that truthfulness cues arise from two distinct information pathways: (1) a Question-Anchored pathway that depends on question-answer information flow, and (2) an Answer-Anchored pathway that derives self-contained evidence from the generated answer itself. First, we validate and disentangle these pathways through attention knockout and token patching. Afterwards, we uncover notable and intriguing properties of these two mechanisms. Further experiments reveal that (1) the two mechanisms are closely associated with LLM knowledge boundaries; and (2) internal representations are aware of their distinctions. Finally, building on these insightful findings, two applications are proposed to enhance hallucination detection performance. Overall, our work provides new insight into how LLMs internally encode truthfulness, offering directions for more reliable and self-aware generative systems.

Method: Created SAD dataset with 392,822 examples grounded in argumentation theories. Annotated each utterance with five strategy types (allowing multiple strategies per utterance). Dataset requires models to generate contextually appropriate arguments conditioned on dialogue history, specified stance on topic, and targeted argumentation strategies.

Result: Presented first large-scale strategic argumentative dialogue dataset. Benchmarked range of pretrained generative models on SAD. Provided in-depth analysis of strategy usage patterns in argumentation.

Conclusion: SAD enables deeper modeling of argumentation dialogue by moving beyond single-turn settings to multi-turn strategic dialogues, supporting research in generating contextually appropriate arguments with targeted persuasive strategies.

Abstract: Argumentation generation has attracted substantial research interest due to its central role in human reasoning and decision-making. However, most existing argumentative corpora focus on non-interactive, single-turn settings, either generating arguments from a given topic or refuting an existing argument. In practice, however, argumentation is often realized as multi-turn dialogue, where speakers defend their stances and employ diverse argumentative strategies to strengthen persuasiveness. To support deeper modeling of argumentation dialogue, we present the first large-scale \textbf{S}trategic \textbf{A}rgumentative \textbf{D}ialogue dataset, SAD, consisting of 392,822 examples. Grounded in argumentation theories, we annotate each utterance with five strategy types, allowing multiple strategies per utterance. Unlike prior datasets, SAD requires models to generate contextually appropriate arguments conditioned on the dialogue history, a specified stance on the topic, and targeted argumentation strategies. We further benchmark a range of pretrained generative models on SAD and present in-depth analysis of strategy usage patterns in argumentation.

Method: KALE has two components: 1) Knowledge-Induced data synthesis extracts multi-hop reasoning paths from knowledge graphs to generate high-quality rationales for QA pairs; 2) Knowledge-Aware fine-tuning enhances knowledge manipulation by minimizing KL divergence between predictions with and without rationales.

Result: Extensive experiments on eight benchmarks across six LLMs show KALE achieves accuracy improvements up to 11.72% and average of 4.18%.

Conclusion: KALE effectively enhances LLMs’ knowledge manipulation ability by leveraging knowledge graphs to generate rationales and internalize reasoning processes, addressing the known&incorrect problem.

Abstract: Despite the impressive performance of large language models (LLMs) pretrained on vast knowledge corpora, advancing their knowledge manipulation-the ability to effectively recall, reason, and transfer relevant knowledge-remains challenging. Existing methods mainly leverage Supervised Fine-Tuning (SFT) on labeled datasets to enhance LLMs’ knowledge manipulation ability. However, we observe that SFT models still exhibit the known&incorrect phenomenon, where they explicitly possess relevant knowledge for a given question but fail to leverage it for correct answers. To address this challenge, we propose KALE (Knowledge-Aware LEarning)-a post-training framework that leverages knowledge graphs (KGs) to generate high-quality rationales and enhance LLMs’ knowledge manipulation ability. Specifically, KALE first introduces a Knowledge-Induced (KI) data synthesis method that efficiently extracts multi-hop reasoning paths from KGs to generate high-quality rationales for question-answer pairs. Then, KALE employs a Knowledge-Aware (KA) fine-tuning paradigm that enhances knowledge manipulation by internalizing rationale-guided reasoning through minimizing the KL divergence between predictions with and without rationales. Extensive experiments on eight popular benchmarks across six different LLMs demonstrate the effectiveness of KALE, achieving accuracy improvements of up to 11.72% and an average of 4.18%.

Method: Introduces a controlled swapped-reference QA framework that induces reference-belief conflicts by replacing reference answers with incorrect entities, constructing diverse pairings of original/swapped references with corresponding candidate answers.

Result: Grading reliability drops sharply under swapped references across various judge models; vulnerability is driven by judges’ over-reliance on parametric knowledge, causing them to disregard given references under conflict.

Conclusion: This failure persists despite common prompt-based mitigation strategies, highlighting a fundamental limitation of LLM-as-a-judge evaluation and motivating reference-based protocols that enforce stronger adherence to provided references.

Abstract: While large language models (LLMs) are increasingly used as automatic judges for question answering (QA) and other reference-conditioned evaluation tasks, little is known about their ability to adhere to a provided reference. We identify a critical failure mode of such reference-based LLM QA evaluation: when the provided reference conflicts with the judge model’s parametric knowledge, the resulting scores become unreliable, substantially degrading evaluation fidelity. To study this phenomenon systematically, we introduce a controlled swapped-reference QA framework that induces reference-belief conflicts. Specifically, we replace the reference answer with an incorrect entity and construct diverse pairings of original and swapped references with correspondingly aligned candidate answers. Surprisingly, grading reliability drops sharply under swapped references across a broad set of judge models. We empirically show that this vulnerability is driven by judges’ over-reliance on parametric knowledge, leading judges to disregard the given reference under conflict. Finally, we find that this failure persists under common prompt-based mitigation strategies, highlighting a fundamental limitation of LLM-as-a-judge evaluation and motivating reference-based protocols that enforce stronger adherence to the provided reference.

Method: SMoA uses structured modulation to selectively amplify/suppress important features across multiple subspaces while freezing pretrained weights, maintaining higher rank with fewer parameters.

Result: SMoA outperforms LoRA and its variants on 10 tasks, with ablation studies validating its effectiveness in improving model capacity and performance.

Conclusion: SMoA provides an efficient high-rank adaptation method that enhances representational capacity while maintaining parameter efficiency, offering better performance than existing PEFT approaches.

Abstract: As the number of model parameters increases, parameter-efficient fine-tuning (PEFT) has become the go-to choice for tailoring pre-trained large language models. Low-rank Adaptation (LoRA) uses a low-rank update method to simulate full parameter fine-tuning, which is widely used to reduce resource requirements. However, decreasing the rank encounters challenges with limited representational capacity when compared to full parameter fine-tuning. We present \textbf{SMoA}, a high-rank \textbf{S}tructured \textbf{MO}dulation \textbf{A}dapter that uses fewer trainable parameters while maintaining a higher rank, thereby improving the model’s representational capacity and offering improved performance potential. The core idea is to freeze the original pretrained weights and selectively amplify or suppress important features of the original weights across multiple subspaces. The subspace mechanism provides an efficient way to increase the capacity and complexity of a model. We conduct both theoretical analyses and empirical studies on various tasks. Experiment results show that SMoA outperforms LoRA and its variants on 10 tasks, with extensive ablation studies validating its effectiveness.

Method: Learn compact latent action space using learning from observation mechanism with codebook construction. Use both paired image-text and text-only data with cross-modal projector (initialized on paired data, trained on text-only with cycle consistency loss).

Result: Latent action based method outperforms competitive baselines on two conversation tasks across various RL algorithms.

Conclusion: Learning compact latent action space with cross-modal projection from both multimodal and text-only data effectively addresses large token space challenge in RL fine-tuning of multimodal conversational agents.

Abstract: Vision-language models are increasingly employed as multimodal conversational agents (MCAs) for diverse conversational tasks. Recently, reinforcement learning (RL) has been widely explored for adapting MCAs to various human-AI interaction scenarios. Despite showing great enhancement in generalization performance, fine-tuning MCAs via RL still faces challenges in handling the extremely large text token space. To address this, we learn a compact latent action space for RL fine-tuning instead. Specifically, we adopt the learning from observation mechanism to construct the codebook for the latent action space, where future observations are leveraged to estimate current latent actions that could further be used to reconstruct future observations. However, the scarcity of paired image-text data hinders learning a codebook with sufficient coverage. Thus, we leverage both paired image-text data and text-only data to construct the latent action space, using a cross-modal projector for transforming text embeddings into image-text embeddings. We initialize the cross-modal projector on paired image-text data, and further train it on massive text-only data with a novel cycle consistency loss to enhance its robustness. We show that our latent action based method outperforms competitive baselines on two conversation tasks across various RL algorithms.

Method: Allow LLMs to reason freely in natural language until specific trigger tokens are generated, then switch to structured generation mode to produce standardized, guaranteed-parsable outputs.

Result: Achieves up to 27% accuracy improvement over natural generation on classification and reasoning tasks, with only 10-20 extra token overhead.

Conclusion: The hybrid approach successfully combines the benefits of both natural and structured generation, preserving expressive reasoning while ensuring reliable structured outputs with minimal overhead.

Abstract: Natural generation allows Language Models (LMs) to produce free-form responses with rich reasoning, but the lack of guaranteed structure makes outputs difficult to parse or verify. Structured generation, or constrained decoding, addresses this drawback by producing content in standardized formats such as JSON, ensuring consistency and guaranteed-parsable outputs, but it can inadvertently restrict the model’s reasoning capabilities. In this work, we propose a simple approach that combines the advantages of both natural and structured generation. By allowing LLMs to reason freely until specific trigger tokens are generated, and then switching to structured generation, our method preserves the expressive power of natural language reasoning while ensuring the reliability of structured outputs. We further evaluate our approach on several datasets, covering both classification and reasoning tasks, to demonstrate its effectiveness, achieving a substantial gain of up to 27% in accuracy compared to natural generation, while requiring only a small overhead of 10-20 extra tokens.

Method: Created ISLAMICFAITHQA benchmark (3,810 bilingual items), developed Islamic modelling suite (25K SFT pairs, 5K preference samples, 6K verse retrieval corpus), and built agentic Quran-grounding framework using structured tool calls for iterative evidence seeking and answer revision.

Result: Retrieval improves correctness, and agentic RAG yields largest gains beyond standard RAG, achieving SOTA performance and strong Arabic-English robustness even with small models like Qwen3 4B.

Conclusion: The paper introduces comprehensive resources for grounded Islamic QA evaluation and demonstrates effectiveness of agentic RAG framework for improving LLM reliability in religious contexts.

Abstract: LLMs are increasingly used for Islamic question answering, where ungrounded responses may carry serious religious consequences. Yet standard MCQ/MRC-style evaluations do not capture key real-world failure modes, notably free-form hallucinations and whether models appropriately abstain when evidence is lacking. To shed a light on this aspect we introduce ISLAMICFAITHQA, a 3,810-item bilingual (Arabic/English) generative benchmark with atomic single-gold answers, which enables direct measurement of hallucination and abstention. We additionally developed an end-to-end grounded Islamic modelling suite consisting of (i) 25K Arabic text-grounded SFT reasoning pairs, (ii) 5K bilingual preference samples for reward-guided alignment, and (iii) a verse-level Qur’an retrieval corpus of $\sim$6k atomic verses (ayat). Building on these resources, we develop an agentic Quran-grounding framework (agentic RAG) that uses structured tool calls for iterative evidence seeking and answer revision. Experiments across Arabic-centric and multilingual LLMs show that retrieval improves correctness and that agentic RAG yields the largest gains beyond standard RAG, achieving state-of-the-art performance and stronger Arabic-English robustness even with a small model (i.e., Qwen3 4B). We will make the experimental resources and datasets publicly available for the community.

Method: EGMF combines expert-guided multimodal fusion with LLMs using three specialized expert networks (fine-grained local, semantic correlation, global context) integrated via hierarchical dynamic gating. Enhanced representations are integrated with LLMs through pseudo token injection and prompt-based conditioning, with LoRA fine-tuning for efficiency.

Result: Experiments on bilingual benchmarks (MELD, CHERMA, MOSEI, SIMS-V2) show consistent improvements over state-of-the-art methods, with superior cross-lingual robustness revealing universal patterns in multimodal emotional expressions across English and Chinese.

Conclusion: EGMF provides a unified generative framework for multimodal emotion understanding that effectively handles both classification and regression tasks while demonstrating strong cross-lingual generalization capabilities.

Abstract: Multimodal emotion understanding requires effective integration of text, audio, and visual modalities for both discrete emotion recognition and continuous sentiment analysis. We present EGMF, a unified framework combining expert-guided multimodal fusion with large language models. Our approach features three specialized expert networks–a fine-grained local expert for subtle emotional nuances, a semantic correlation expert for cross-modal relationships, and a global context expert for long-range dependencies–adaptively integrated through hierarchical dynamic gating for context-aware feature selection. Enhanced multimodal representations are integrated with LLMs via pseudo token injection and prompt-based conditioning, enabling a single generative framework to handle both classification and regression through natural language generation. We employ LoRA fine-tuning for computational efficiency. Experiments on bilingual benchmarks (MELD, CHERMA, MOSEI, SIMS-V2) demonstrate consistent improvements over state-of-the-art methods, with superior cross-lingual robustness revealing universal patterns in multimodal emotional expressions across English and Chinese. We will release the source code publicly.

Method: ES-Mem incorporates two core components: (1) dynamic event segmentation module that partitions long-term interactions into semantically coherent events with distinct boundaries, and (2) hierarchical memory architecture that constructs multi-layered memories and leverages boundary semantics to anchor specific episodic memory for precise context localization.

Result: Evaluations on two memory benchmarks show consistent performance gains over baseline methods. The event segmentation module also exhibits robust applicability on dialogue segmentation datasets.

Conclusion: ES-Mem effectively addresses limitations of existing memory mechanisms by incorporating event segmentation theory, enabling better coherence and precise context localization for dialogue agents in long-term interactions.

Abstract: Memory is critical for dialogue agents to maintain coherence and enable continuous adaptation in long-term interactions. While existing memory mechanisms offer basic storage and retrieval capabilities, they are hindered by two primary limitations: (1) rigid memory granularity often disrupts semantic integrity, resulting in fragmented and incoherent memory units; (2) prevalent flat retrieval paradigms rely solely on surface-level semantic similarity, neglecting the structural cues of discourse required to navigate and locate specific episodic contexts. To mitigate these limitations, drawing inspiration from Event Segmentation Theory, we propose ES-Mem, a framework incorporating two core components: (1) a dynamic event segmentation module that partitions long-term interactions into semantically coherent events with distinct boundaries; (2) a hierarchical memory architecture that constructs multi-layered memories and leverages boundary semantics to anchor specific episodic memory for precise context localization. Evaluations on two memory benchmarks demonstrate that ES-Mem yields consistent performance gains over baseline methods. Furthermore, the proposed event segmentation module exhibits robust applicability on dialogue segmentation datasets.

Method: PoT freezes a pre-cutoff snapshot of evidence in an offline sandbox and asks models to forecast post-cutoff outcomes (citations, research agenda shifts). It provides a controlled testbed for agent-based research judgments with tool-using agents vs. non-agent baselines.

Result: Across 30,000+ instances in four benchmark domains: higher interaction budgets generally improve agent performance, while the benefit of tool use is strongly task-dependent.

Conclusion: PoT supports scalable evaluation of agents on future-facing scientific idea judgment tasks by combining time-partitioned, future-verifiable targets with an offline sandbox for tool use.

Abstract: Large language models are increasingly being used to assess and forecast research ideas, yet we lack scalable ways to evaluate the quality of models’ judgments about these scientific ideas. Towards this goal, we introduce PoT, a semi-verifiable benchmarking framework that links scientific idea judgments to downstream signals that become observable later (e.g., citations and shifts in researchers’ agendas). PoT freezes a pre-cutoff snapshot of evidence in an offline sandbox and asks models to forecast post-cutoff outcomes, enabling verifiable evaluation when ground truth arrives, scalable benchmarking without exhaustive expert annotation, and analysis of human-model misalignment against signals such as peer-review awards. In addition, PoT provides a controlled testbed for agent-based research judgments that evaluate scientific ideas, comparing tool-using agents to non-agent baselines under prompt ablations and budget scaling. Across 30,000+ instances spanning four benchmark domains, we find that, compared with non-agent baselines, higher interaction budgets generally improve agent performance, while the benefit of tool use is strongly task-dependent. By combining time-partitioned, future-verifiable targets with an offline sandbox for tool use, PoT supports scalable evaluation of agents on future-facing scientific idea judgment tasks.

Method: Pilot deployment of Descartes multilingual model in Wikipedia Android app, offering AI-generated short description suggestions to editors across 12 languages with 3,900+ articles and 375+ editors.

Result: 90% of accepted Descartes descriptions rated at least 3/5 in quality, comparable to human-written ones; low revert/report rates; editors adopted suggestions both directly and with modifications.

Conclusion: Descartes can effectively support editors in reducing content gaps, but requires technical, design, and community safeguards including addressing latency, language-specific gaps, and sensitive topics.

Abstract: Short descriptions are a key part of the Wikipedia user experience, but their coverage remains uneven across languages and topics. In previous work, we introduced Descartes, a multilingual model for generating short descriptions. In this report, we present the results of a pilot deployment of Descartes in the Wikipedia Android app, where editors were offered suggestions based on outputs from Descartes while editing short descriptions. The experiment spanned 12 languages, with over 3,900 articles and 375 editors participating. Overall, 90% of accepted Descartes descriptions were rated at least 3 out of 5 in quality, and their average ratings were comparable to human-written ones. Editors adopted machine suggestions both directly and with modifications, while the rate of reverts and reports remained low. The pilot also revealed practical considerations for deployment, including latency, language-specific gaps, and the need for safeguards around sensitive topics. These results indicate that Descartes’s short descriptions can support editors in reducing content gaps, provided that technical, design, and community guardrails are in place.

Method: 1) Layer-wise vision token masking reveals three-stage pattern in MLLMs; 2) Plateau-guided model merging selectively injects base language model parameters into MLLMs; 3) Training-free framework.

Result: Experimental results on five MLLMs across nine benchmarks demonstrate effectiveness. Attention analysis shows merging shifts attention from diffuse patterns to focused localization on task-relevant visual regions.

Conclusion: Proposed plateau-guided model merging effectively mitigates degradation of text reasoning in MLLMs without additional training, improving multimodal performance through selective parameter injection.

Abstract: Multimodal Large Language Models (MLLMs) rely on strong linguistic reasoning inherited from their base language models. However, multimodal instruction fine-tuning paradoxically degrades this text’s reasoning capability, undermining multimodal performance. To address this issue, we propose a training-free framework to mitigate this degradation. Through layer-wise vision token masking, we reveal a common three-stage pattern in multimodal large language models: early-modal separation, mid-modal alignment, and late-modal degradation. By analyzing the behavior of MLLMs at different stages, we propose a plateau-guided model merging method that selectively injects base language model parameters into MLLMs. Experimental results based on five MLLMs on nine benchmarks demonstrate the effectiveness of our method. Attention-based analysis further reveals that merging shifts attention from diffuse, scattered patterns to focused localization on task-relevant visual regions. Our repository is on https://github.com/wzj1718/PlaM.

Method: Automatic information extraction scheme to gather key evaluation information from 14,171 papers across four major conferences (ACL, EMNLP, NAACL, INLG) over six years. Analysis focuses on different evaluation methods and their usage patterns across tasks.

Result: Three key findings: (1) Task Divergence - Dialogue Generation rapidly shifts to LLM-as-a-judge (>40% in 2025), Machine Translation sticks to n-gram metrics, Question Answering shows decline in human evaluation. (2) Metric Inertia - General-purpose metrics (BLEU, ROUGE) remain widely used without justification despite new semantic metrics. (3) Human-LaaJ Divergence - LLM-as-a-judge and human evaluations prioritize different signals with only moderate to low correlation; explicit validation is scarce (<8% of papers).

Conclusion: Current NLG evaluation practices show problematic patterns: task-specific evaluation biases, persistence of outdated metrics, and insufficient validation of LLM-as-a-judge methods. Practical recommendations are derived to improve rigor in future NLG evaluation.

Abstract: Despite advances in Natural Language Generation (NLG), evaluation remains challenging. Although various new metrics and LLM-as-a-judge (LaaJ) methods are proposed, human judgment persists as the gold standard. To systematically review how NLG evaluation has evolved, we employ an automatic information extraction scheme to gather key information from NLG papers, focusing on different evaluation methods (metrics, LaaJ and human evaluation). With extracted metadata from 14,171 papers across four major conferences (ACL, EMNLP, NAACL, and INLG) over the past six years, we reveal several critical findings: (1) Task Divergence: While Dialogue Generation demonstrates a rapid shift toward LaaJ (>40% in 2025), Machine Translation remains locked into n-gram metrics, and Question Answering exhibits a substantial decline in the proportion of studies conducting human evaluation. (2) Metric Inertia: Despite the development of semantic metrics, general-purpose metrics (e.g., BLEU, ROUGE) continue to be widely used across tasks without empirical justification, often lacking the discriminative power to distinguish between specific quality criteria. (3) Human-LaaJ Divergence: Our association analysis challenges the assumption that LLMs act as mere proxies for humans; LaaJ and human evaluations prioritize very different signals, and explicit validation is scarce (<8% of papers comparing the two), with only moderate to low correlation. Based on these observations, we derive practical recommendations to improve the rigor of future NLG evaluation.

Method: ASL adaptively chooses selection layers for KV cache reduction by exploiting the variance of token ranks ordered by attention score. It operates during prefilling stage and can be combined with existing methods like SnapKV for decoding optimization. Uses one-shot token selection where tokens are selected at a layer and propagated to deeper layers.

Result: ASL outperforms state-of-the-art layer-wise token selection methods in accuracy while maintaining decoding speed and KV cache reduction, as demonstrated on InfiniteBench, RULER, and NIAH benchmarks.

Conclusion: ASL provides a flexible, training-free approach to KV cache reduction that balances performance across different tasks while meeting user-specified KV budget requirements, and can be integrated with existing methods for further optimization.

Abstract: Due to the prevalence of large language models (LLMs), key-value (KV) cache reduction for LLM inference has received remarkable attention. Among numerous works that have been proposed in recent years, layer-wise token pruning approaches, which select a subset of tokens at particular layers to retain in KV cache and prune others, are one of the most popular schemes. They primarily adopt a set of pre-defined layers, at which tokens are selected. Such design is inflexible in the sense that the accuracy significantly varies across tasks and deteriorates in harder tasks such as KV retrieval. In this paper, we propose ASL, a training-free method that adaptively chooses the selection layer for KV cache reduction, exploiting the variance of token ranks ordered by attention score. The proposed method balances the performance across different tasks while meeting the user-specified KV budget requirement. ASL operates during the prefilling stage and can be jointly used with existing KV cache reduction methods such as SnapKV to optimize the decoding stage. By evaluations on the InfiniteBench, RULER, and NIAH benchmarks, we show that equipped with one-shot token selection, where tokens are selected at a layer and propagated to deeper layers, ASL outperforms state-of-the-art layer-wise token selection methods in accuracy while maintaining decoding speed and KV cache reduction.

Method: Created a novel QA task based on geopolitical indicators across countries and years requiring step breakdown, data retrieval, and mathematical operations. Analyzed meta-level reasoning by examining LLMs’ tool selection for answering questions, comparing against predefined ’essential actions’ rather than just final answer accuracy.

Result: LLMs demonstrate good meta-level reasoning but have flaws in task understanding. N-shot prompting has little effect on accuracy, error messages don’t significantly deteriorate performance, and LLMs show poor numeracy. The analysis provides deeper insights beyond simple accuracy metrics.

Conclusion: The structured framework for analyzing reasoning provides valuable insights into LLM capabilities, revealing strengths in planning but weaknesses in execution and numeracy. The approach offers a more nuanced evaluation method that can generalize to other task domains.

Abstract: Recent advancements in Large Language Models (LLMs) are increasingly focused on “reasoning” ability, a concept with many overlapping definitions in the LLM discourse. We take a more structured approach, distinguishing meta-level reasoning (denoting the process of reasoning about intermediate steps required to solve a task) from object-level reasoning (which concerns the low-level execution of the aforementioned steps.) We design a novel question answering task, which is based around the values of geopolitical indicators for various countries over various years. Questions require breaking down into intermediate steps, retrieval of data, and mathematical operations over that data. The meta-level reasoning ability of LLMs is analysed by examining the selection of appropriate tools for answering questions. To bring greater depth to the analysis of LLMs beyond final answer accuracy, our task contains ’essential actions’ against which we can compare the tool call output of LLMs to infer the strength of reasoning ability. We find that LLMs demonstrate good meta-level reasoning on our task, yet are flawed in some aspects of task understanding. We find that n-shot prompting has little effect on accuracy; error messages encountered do not often deteriorate performance; and provide additional evidence for the poor numeracy of LLMs. Finally, we discuss the generalisation and limitation of our findings to other task domains.

Method: Developed a controllable seeker simulator driven by nine psychological and linguistic features underpinning seeker behavior. Trained the model using authentic Reddit conversations via a Mixture-of-Experts (MoE) architecture that differentiates diverse seeker behaviors into specialized parameter subspaces for enhanced fine-grained controllability.

Result: The simulator achieves superior profile adherence and behavioral diversity compared to existing approaches. Evaluation of 7 prominent supporter models with this system uncovers previously obscured performance degradations.

Conclusion: The framework provides a more faithful and stress-tested evaluation for emotional support chatbots, demonstrating its utility in revealing performance issues that were previously hidden by less sophisticated evaluation methods.

Abstract: As emotional support chatbots have recently gained significant traction across both research and industry, a common evaluation strategy has emerged: use help-seeker simulators to interact with supporter chatbots. However, current simulators suffer from two critical limitations: (1) they fail to capture the behavioral diversity of real-world seekers, often portraying them as overly cooperative, and (2) they lack the controllability required to simulate specific seeker profiles. To address these challenges, we present a controllable seeker simulator driven by nine psychological and linguistic features that underpin seeker behavior. Using authentic Reddit conversations, we train our model via a Mixture-of-Experts (MoE) architecture, which effectively differentiates diverse seeker behaviors into specialized parameter subspaces, thereby enhancing fine-grained controllability. Our simulator achieves superior profile adherence and behavioral diversity compared to existing approaches. Furthermore, evaluating 7 prominent supporter models with our system uncovers previously obscured performance degradations. These findings underscore the utility of our framework in providing a more faithful and stress-tested evaluation for emotional support chatbots.

Method: The authors conduct an extensive, empirically driven evaluation of Enhanced and Agentic RAG across multiple scenarios and dimensions. They systematically compare both paradigms to understand their trade-offs.

Result: The evaluation provides practical insights into the trade-offs between Enhanced and Agentic RAG paradigms. The results offer guidance on selecting the most effective RAG design for real-world applications, considering both costs and performance.

Conclusion: The paper provides empirical evidence to help practitioners choose between Enhanced RAG (with engineered modules) and Agentic RAG (with LLM orchestration) based on specific conditions, costs, and performance requirements in real-world applications.

Abstract: Retrieval-Augmented Generation (RAG) systems are usually defined by the combination of a generator and a retrieval component that extracts textual context from a knowledge base to answer user queries. However, such basic implementations exhibit several limitations, including noisy or suboptimal retrieval, misuse of retrieval for out-of-scope queries, weak query-document matching, and variability or cost associated with the generator. These shortcomings have motivated the development of “Enhanced” RAG, where dedicated modules are introduced to address specific weaknesses in the workflow. More recently, the growing self-reflective capabilities of Large Language Models (LLMs) have enabled a new paradigm, which we refer to as “Agentic” RAG. In this approach, the LLM orchestrates the entire process-deciding which actions to perform, when to perform them, and whether to iterate-thereby reducing reliance on fixed, manually engineered modules. Despite the rapid adoption of both paradigms, it remains unclear which approach is preferable under which conditions. In this work, we conduct an extensive, empirically driven evaluation of Enhanced and Agentic RAG across multiple scenarios and dimensions. Our results provide practical insights into the trade-offs between the two paradigms, offering guidance on selecting the most effective RAG design for real-world applications, considering both costs and performance.

Method: Proposes a framework that incorporates structured information using Knowledge Graphs (KGs) extracted from documents using a proposed schema, combined with LLM predictions for numerical reasoning tasks.

Result: Evaluated on FinQA benchmark using Llama 3.1 8B Instruct, the framework improved execution accuracy by approximately 12% relative to the vanilla LLM.

Conclusion: Incorporating structured information through Knowledge Graphs significantly enhances LLM performance for numerical reasoning in financial documents, addressing the bottleneck of accurate numerical processing in financial analytics.

Abstract: Numerical reasoning is an important task in the analysis of financial documents. It helps in understanding and performing numerical predictions with logical conclusions for the given query seeking answers from financial texts. Recently, Large Language Models (LLMs) have shown promising results in multiple Question-Answering (Q-A) systems with the capability of logical reasoning. As documents related to finance often consist of long and complex financial contexts, LLMs appear well-suited for building high-quality automated financial question-answering systems. However, LLMs often face challenges in accurately processing the various numbers within financial reports. Extracting numerical data from unstructured text and semi-structured tables, and reliably performing accurate calculations, remains a significant bottleneck for numerical reasoning in most state-of-the-art LLMs. Recent studies have shown that structured data augmentations, such as Knowledge Graphs (KGs), have notably improved the predictions of LLMs along with logical explanations. Thus, it is an important requirement to consider inherent structured information in financial reports while using LLMs for various financial analytics. This paper proposes a framework to incorporate structured information using KGs along with LLM predictions for numerical reasoning tasks. The KGs are extracted using a proposed schema inherently from the document under processing. We evaluated our proposed framework over the benchmark data FinQA, using an open-source LLM, namely Llama 3.1 8B Instruct. We observed that the proposed framework improved execution accuracy by approximately 12% relative to the vanilla LLM.

Method: Contrastive learning framework that learns story embeddings from narrative twins (stories sharing same plot but differing in surface form). Model distinguishes a story from both its narrative twin and a distractor with similar surface features but different plot. Evaluates four narratological operations: deletion, shifting, disruption, and summarization.

Result: Contrastively learned story embeddings outperform masked-language-model baseline. Summarization is the most reliable operation for identifying salient sentences. When narrative twins are unavailable, random dropout can generate twins from a single story. Effective distractors can be obtained by prompting LLMs or using different parts of the same story in long-form narratives.

Conclusion: The contrastive learning framework with narrative twins effectively models narrative salience, with summarization being the most effective operation. The approach works even when narrative twins are not available through techniques like random dropout and LLM-generated distractors.

Abstract: Understanding narratives requires identifying which events are most salient for a story’s progression. We present a contrastive learning framework for modeling narrative salience that learns story embeddings from narrative twins: stories that share the same plot but differ in surface form. Our model is trained to distinguish a story from both its narrative twin and a distractor with similar surface features but different plot. Using the resulting embeddings, we evaluate four narratologically motivated operations for inferring salience (deletion, shifting, disruption, and summarization). Experiments on short narratives from the ROCStories corpus and longer Wikipedia plot summaries show that contrastively learned story embeddings outperform a masked-language-model baseline, and that summarization is the most reliable operation for identifying salient sentences. If narrative twins are not available, random dropout can be used to generate the twins from a single story. Effective distractors can be obtained either by prompting LLMs or, in long-form narratives, by using different parts of the same story.

Method: PR-CoT employs structured multi-perspective reflection after initial CoT, guiding LLMs to self-assess reasoning across four angles: logical consistency, information completeness, biases/ethics, and alternative solutions, implemented purely via prompt engineering.

Result: Experiments across arithmetic, commonsense, ethical decision-making, and logical puzzles show PR-CoT significantly outperforms traditional CoT and existing reflection methods in logical consistency and error correction, with notable gains in ethical decision-making.

Conclusion: The poly-reflective paradigm fosters more reliable LLM reasoning, with ablation studies, human evaluations, and qualitative analyses validating the contribution of each reflection perspective and overall efficacy.

Abstract: While Chain-of-Thought (CoT) prompting advances LLM reasoning, challenges persist in consistency, accuracy, and self-correction, especially for complex or ethically sensitive tasks. Existing single-dimensional reflection methods offer insufficient improvements. We propose MyGO Poly-Reflective Chain-of-Thought (PR-CoT), a novel methodology employing structured multi-perspective reflection. After initial CoT, PR-CoT guides the LLM to self-assess its reasoning across multiple predefined angles: logical consistency, information completeness, biases/ethics, and alternative solutions. Implemented purely via prompt engineering, this process refines the initial CoT into a more robust and accurate final answer without model retraining. Experiments across arithmetic, commonsense, ethical decision-making, and logical puzzles, using GPT-three point five and GPT-four models, demonstrate PR-CoT’s superior performance. It significantly outperforms traditional CoT and existing reflection methods in logical consistency and error correction, with notable gains in nuanced domains like ethical decision-making. Ablation studies, human evaluations, and qualitative analyses further validate the contribution of each reflection perspective and the overall efficacy of our poly-reflective paradigm in fostering more reliable LLM reasoning.

Method: TOOLQP decomposes instructions into sub-tasks and dynamically generates queries to interact with the retriever, bridging the semantic gap. It’s trained using synthetic query trajectories followed by optimization via Reinforcement Learning with Verifiable Rewards (RLVR).

Result: TOOLQP achieves state-of-the-art performance, exhibiting superior zero-shot generalization, robustness across diverse retrievers, and significant improvements in downstream agentic execution.

Conclusion: Modeling retrieval as iterative query planning effectively addresses the limitations of single-shot dense retrievers for complex tool composition tasks, enabling better agent performance.

Abstract: LLM agents operating over massive, dynamic tool libraries rely on effective retrieval, yet standard single-shot dense retrievers struggle with complex requests. These failures primarily stem from the disconnect between abstract user goals and technical documentation, and the limited capacity of fixed-size embeddings to model combinatorial tool compositions. To address these challenges, we propose TOOLQP, a lightweight framework that models retrieval as iterative query planning. Instead of single-shot matching, TOOLQP decomposes instructions into sub-tasks and dynamically generates queries to interact with the retriever, effectively bridging the semantic gap by targeting the specific sub-tasks required for composition. We train TOOLQP using synthetic query trajectories followed by optimization via Reinforcement Learning with Verifiable Rewards (RLVR). Experiments demonstrate that TOOLQP achieves state-of-the-art performance, exhibiting superior zero-shot generalization, robustness across diverse retrievers, and significant improvements in downstream agentic execution.

Method: Developed a generative pipeline that produces large-scale, realistic, culture-specific genealogical data (family trees) with explicit marriage constraints. From these genealogies, created textual inference tasks requiring reasoning over implicit relational chains. Evaluated six state-of-the-art LLMs (open and closed-source) using zero-shot protocol with deterministic decoding.

Result: KinshipQA yields a wide spread of outcomes and exposes systematic differences in multi-hop reasoning across models and cultural settings. Performance was measured using exact-match and set-based metrics.

Conclusion: The benchmark successfully probes LLMs’ multi-hop reasoning capabilities and reveals important differences in how models handle reasoning across different cultural kinship systems and relational complexities.

Abstract: Large language models (LLMs) are increasingly evaluated on their ability to perform multi-hop reasoning, i.e., to combine multiple pieces of information into a coherent inference. We introduce KinshipQA, a benchmark designed to probe this capability through reasoning over kinship relations. The central contribution of our work is a generative pipeline that produces, on demand, large-scale, realistic, and culture-specific genealogical data: collections of interconnected family trees that satisfy explicit marriage constraints associated with different kinship systems. This allows task difficulty, cultural assumptions, and relational depth to be systematically controlled and varied. From these genealogies, we derive textual inference tasks that require reasoning over implicit relational chains. We evaluate the resulting benchmark using six state-of-the-art LLMs, spanning both open-source and closed-source models, under a uniform zero-shot protocol with deterministic decoding. Performance is measured using exact-match and set-based metrics. Our results demonstrate that KinshipQA yields a wide spread of outcomes and exposes systematic differences in multi-hop reasoning across models and cultural settings.

Method: Analyzed 397 human-LLM conversations about socio-political issues using linguistic and interactional feature analysis, mediation analysis to identify mechanisms, and moderation analysis to examine conditional effects by political efficacy.

Result: LLM explanatory richness partially supports confidence through fostering reflective insight, while its effect on knowledge gain operates entirely through cognitive engagement. Effects vary by political efficacy: confidence gains depend on how high-efficacy users handle uncertainty, and knowledge gains depend on high-efficacy users’ ability to leverage extended interaction.

Conclusion: Learning from LLMs is an interactional achievement, not just a result of better explanations. LLM explanatory behavior must align with users’ engagement states to support effective learning in Human-AI interactive systems.

Abstract: Large language models (LLMs) are increasingly used as conversational partners for learning, yet the interactional dynamics supporting users’ learning and engagement are understudied. We analyze the linguistic and interactional features from both LLM and participant chats across 397 human-LLM conversations about socio-political issues to identify the mechanisms and conditions under which LLM explanations shape changes in political knowledge and confidence. Mediation analyses reveal that LLM explanatory richness partially supports confidence by fostering users’ reflective insight, whereas its effect on knowledge gain operates entirely through users’ cognitive engagement. Moderation analyses show that these effects are highly conditional and vary by political efficacy. Confidence gains depend on how high-efficacy users experience and resolve uncertainty. Knowledge gains depend on high-efficacy users’ ability to leverage extended interaction, with longer conversations benefiting primarily reflective users. In summary, we find that learning from LLMs is an interactional achievement, not a uniform outcome of better explanations. The findings underscore the importance of aligning LLM explanatory behavior with users’ engagement states to support effective learning in designing Human-AI interactive systems.

Method: The research focuses on gender bias in gendered pronoun resolution tasks. It examines probability confidence calibration across six state-of-the-art LLMs and introduces a new calibration metric called Gender-ECE specifically designed to measure gender disparities in resolution tasks.

Result: Among the six state-of-the-art models tested, Gemma-2 demonstrated the worst calibration according to the gender bias benchmark. The study shows that calibration metrics can reveal fairness-related disparities in LLMs.

Conclusion: The primary contribution is a fairness-aware evaluation of LLMs’ confidence calibration, providing guidance for ethical deployment. The new Gender-ECE metric offers a specialized tool for measuring gender disparities in resolution tasks.

Abstract: The increased use of Large Language Models (LLMs) in sensitive domains leads to growing interest in how their confidence scores correspond to fairness and bias. This study examines the alignment between LLM-predicted confidence and human-annotated bias judgments. Focusing on gender bias, the research investigates probability confidence calibration in contexts involving gendered pronoun resolution. The goal is to evaluate if calibration metrics based on predicted confidence scores effectively capture fairness-related disparities in LLMs. The results show that, among the six state-of-the-art models, Gemma-2 demonstrates the worst calibration according to the gender bias benchmark. The primary contribution of this work is a fairness-aware evaluation of LLMs’ confidence calibration, offering guidance for ethical deployment. In addition, we introduce a new calibration metric, Gender-ECE, designed to measure gender disparities in resolution tasks.

Method: The researchers use reference games as testbeds because they are controlled, self-contained, and make clarification needs explicit and measurable. They evaluate three vision-language models on two tasks: baseline reference resolution vs. an experiment where models are instructed to request clarification when uncertain.

Result: The results show that even in simple reference game tasks, models often fail to recognize their internal uncertainty and translate it into adequate clarification behavior. This highlights limitations in models’ interactive conversational capabilities.

Conclusion: Reference games are valuable testbeds for evaluating interaction qualities of vision and language models, and current models demonstrate significant limitations in assuming active addressee roles with appropriate uncertainty recognition and clarification behavior.

Abstract: In human conversation, both interlocutors play an active role in maintaining mutual understanding. When addressees are uncertain about what speakers mean, for example, they can request clarification. It is an open question for language models whether they can assume a similar addressee role, recognizing and expressing their own uncertainty through clarification. We argue that reference games are a good testbed to approach this question as they are controlled, self-contained, and make clarification needs explicit and measurable. To test this, we evaluate three vision-language models comparing a baseline reference resolution task to an experiment where the models are instructed to request clarification when uncertain. The results suggest that even in such simple tasks, models often struggle to recognize internal uncertainty and translate it into adequate clarification behavior. This demonstrates the value of reference games as testbeds for interaction qualities of (vision and) language models.

Method: Domain-adaptive continual pre-training on financial domain data, with exploration of simple but effective data selection strategies to optimize training efficiency.

Result: FinPythia-6.9B shows consistent improvements on financial tasks over the original foundational model. Data selection strategies outperform vanilla continual pre-training with just 10% of corpus size and cost, without degradation on open-domain tasks.

Conclusion: Continual pre-training with data selection provides a cost-effective alternative solution for building domain-specific LLMs, maintaining open-domain capabilities while improving domain performance.

Abstract: Large language models (LLMs) have demonstrated remarkable open-domain capabilities. LLMs tailored for a domain are typically trained entirely on domain corpus to excel at handling domain-specific tasks. In this work, we explore an alternative strategy of continual pre-training as a means to develop domain-specific LLMs over an existing open-domain LLM. We introduce FinPythia-6.9B, developed through domain-adaptive continual pre-training on the financial domain. Continual pre-trained FinPythia showcases consistent improvements on financial tasks over the original foundational model. We further explore simple but effective data selection strategies for continual pre-training. Our data selection strategies outperform vanilla continual pre-training’s performance with just 10% of corpus size and cost, without any degradation on open-domain standard tasks. Our work proposes an alternative solution to building domain-specific LLMs cost-effectively.

Method: Proposes a two-phase “browse-and-concentrate” paradigm: 1) “Browse” through inputs to extract essential insights, 2) “Concentrate” by revisiting inputs to focus on crucial details guided by those insights. Also develops specialized training strategies to enhance multi-image understanding.

Result: Method significantly boosts performance on 7 multi-image scenarios, achieving average accuracy improvements of 2.13% and 7.60% against strong MLLM baselines with 3B and 11B LLMs respectively.

Conclusion: The browse-and-concentrate paradigm effectively addresses prior-LLM modality isolation in MLLMs, enabling deeper multimodal context fusion and significantly improving multi-image understanding capabilities.

Abstract: With the bloom of Large Language Models (LLMs), Multimodal Large Language Models (MLLMs) that incorporate LLMs with pre-trained vision models have recently demonstrated impressive performance across diverse vision-language tasks. However, they fall short to comprehend context involving multiple images. A primary reason for this shortcoming is that the visual features for each images are encoded individually by frozen encoders before feeding into the LLM backbone, lacking awareness of other images and the multimodal instructions. We term this issue as prior-LLM modality isolation and propose a two phase paradigm, browse-and-concentrate, to enable in-depth multimodal context fusion prior to feeding the features into LLMs. This paradigm initially “browses” through the inputs for essential insights, and then revisits the inputs to “concentrate” on crucial details, guided by these insights, to achieve a more comprehensive understanding of the multimodal inputs. Additionally, we develop training strategies specifically to enhance the understanding of multi-image inputs. Our method markedly boosts the performance on 7 multi-image scenarios, contributing to increments on average accuracy by 2.13% and 7.60% against strong MLLMs baselines with 3B and 11B LLMs, respectively.

Method: MUSE augments LLMs with vision-language modeling and web retrieval over relevant, credible sources to identify misinformation, explain why it’s misleading, and provide grounded references.

Result: MUSE consistently produces high-quality outputs across diverse social media content, outperforming GPT-4 by 37% and high-quality human responses by 29% on comprehensive rubrics.

Conclusion: The work provides a general methodological and evaluative framework for correcting misinformation at scale, demonstrating effectiveness even on content not previously fact-checked online.

Abstract: Real-world information, often multimodal, can be misinformed or potentially misleading due to factual errors, outdated claims, missing context, misinterpretation, and more. Such “misinformation” is understudied, challenging to address, and harms many social domains – particularly on social media, where it can spread rapidly. Manual correction that identifies and explains its (in)accuracies is widely accepted but difficult to scale. While large language models (LLMs) can generate human-like language that could accelerate misinformation correction, they struggle with outdated information, hallucinations, and limited multimodal capabilities. We propose MUSE, an LLM augmented with vision-language modeling and web retrieval over relevant, credible sources to generate responses that determine whether and which part(s) of the given content can be misinformed or potentially misleading, and to explain why with grounded references. We further define a comprehensive set of rubrics to measure response quality, ranging from the accuracy of identifications and factuality of explanations to the relevance and credibility of references. Results show that MUSE consistently produces high-quality outputs across diverse social media content (e.g., modalities, domains, political leanings), including content that has not previously been fact-checked online. Overall, MUSE outperforms GPT-4 by 37% and even high-quality responses from social media users by 29%. Our work provides a general methodological and evaluative framework for correcting misinformation at scale.

Method: For each instruction, GRAPE gathers responses from various LLMs and selects the one with the highest probability measured by the target model, then proceeds with standard SFT training.

Result: GRAPE significantly outperforms strong baselines: up to 13.8% absolute gain over distillation, 17.3% improvement over 3x more data, and 3.5% better than Tulu3-SFT with 1/3 data and half epochs.

Conclusion: GRAPE effectively selects training data aligned with target model distribution, achieving superior performance with less data and training time compared to existing approaches.

Abstract: High-quality supervised fine-tuning (SFT) data are crucial for eliciting strong capabilities from pretrained large language models (LLMs). Typically, instructions are paired with multiple responses sampled from other LLMs, which are often out of the distribution of the target model to be fine-tuned. This, at scale, can lead to diminishing returns and even hurt the models’ performance and robustness. We propose GRAPE, a novel SFT framework that accounts for the unique characteristics of the target model. For each instruction, it gathers responses from various LLMs and selects the one with the highest probability measured by the target model, indicating that it aligns most closely with the target model’s pretrained distribution; it then proceeds with standard SFT training. We first evaluate GRAPE with a controlled experiment, where we sample various solutions for each question in UltraInteract from multiple models and fine-tune commonly used LMs like LLaMA3.1-8B, Mistral-7B, and Qwen2.5-7B on GRAPE-selected data. GRAPE significantly outperforms strong baselines, including distilling from the strongest model with an absolute gain of up to 13.8%, averaged across benchmarks, and training on 3x more data with a maximum performance improvement of 17.3%. GRAPE’s strong performance generalizes to realistic settings. We experiment with the post-training data used for Tulu3 and Olmo-2. GRAPE outperforms strong baselines trained on 4.5 times more data by 6.1% and a state-of-the-art data selection approach by 3% on average performance. Remarkably, using 1/3 of the data and half the number of epochs, GRAPE enables LLaMA3.1-8B to surpass the performance of Tulu3-SFT by 3.5%.

Method: Uses nine collaborative LLM agents for entity discovery, relation extraction, schema alignment, and conflict resolution that iteratively parse documents, verify extracted knowledge, and integrate it into existing graph structures while adhering to domain-specific schemas.

Result: Experiments on 1,200 PubMed articles across three domains identified up to 38,230 new entities with 83.1% LLM-verified correctness and reduced conflict edges by 18.6% through multi-layer assessments.

Conclusion: KARMA demonstrates effective automated knowledge graph enrichment through multi-agent LLM collaboration, addressing scalability challenges in maintaining comprehensive KGs from scientific literature.

Abstract: Maintaining comprehensive and up-to-date knowledge graphs (KGs) is critical for modern AI systems, but manual curation struggles to scale with the rapid growth of scientific literature. This paper presents KARMA, a novel framework employing multi-agent large language models (LLMs) to automate KG enrichment through structured analysis of unstructured text. Our approach employs nine collaborative agents, spanning entity discovery, relation extraction, schema alignment, and conflict resolution that iteratively parse documents, verify extracted knowledge, and integrate it into existing graph structures while adhering to domain-specific schema. Experiments on 1,200 PubMed articles from three different domains demonstrate the effectiveness of KARMA in knowledge graph enrichment, with the identification of up to 38,230 new entities while achieving 83.1% LLM-verified correctness and reducing conflict edges by 18.6% through multi-layer assessments.

Method: Introduces Normalized Probability Shift by the Question (NPSQ) scoring method to isolate question impact. Experiments with various input formats (cloze, symbols, hybrid) and compares against traditional methods like log-likelihood and length-normalized variants.

Result: Traditional scoring methods are vulnerable to superficial characteristics of answer choices, while NPSQ remains stable even when answer options are modified, providing more reliable assessment of comprehension.

Conclusion: NPSQ offers a more reliable way to evaluate LLM comprehension in MCQA by focusing on question impact rather than being influenced by answer choice characteristics.

Abstract: Recent findings raise concerns about whether the evaluation of Multiple-Choice Question Answering (MCQA) accurately reflects the comprehension abilities of large language models. This paper explores the concept of choice sensitivity, which refers to the tendency for model decisions to be more influenced by the answer options than by a genuine understanding of the question. We introduce a new scoring method called Normalized Probability Shift by the Question (NPSQ), designed to isolate the impact of the question itself and provide a more reliable assessment of comprehension. Through experiments involving various input formats, including cloze, symbols, and hybrid formats, we find that traditional scoring methods - such as those based on log-likelihood or its length-normalized variant - are vulnerable to superficial characteristics of the answer choices. In contrast, NPSQ remains stable even when modifications are made to the answer options.

Method: LoRA-Null initializes LoRA parameters in the null space of input activations rather than pre-trained weights. This approach leverages that: 1) input activations incorporate information from all previous layers and input data, not just current layer weights; 2) input activations have much smaller effective ranks than weights, making their null space more accurate and containing less pre-trained knowledge information.

Result: Experimental results show LoRA-Null effectively preserves pre-trained world knowledge of LLMs while achieving good fine-tuning performance, outperforming existing methods like MiLoRA that use the null space of weights.

Conclusion: The null space of input activations is superior to that of weights for LoRA initialization to prevent catastrophic forgetting. LoRA-Null successfully balances knowledge preservation with fine-tuning effectiveness, demonstrating the importance of considering activation spaces rather than just weight spaces for parameter-efficient fine-tuning.

Abstract: Low-Rank Adaptation (LoRA) is the leading parameter-efficient fine-tuning method for Large Language Models (LLMs), but it still suffers from catastrophic forgetting. Recent work has shown that specialized LoRA initialization can alleviate catastrophic forgetting. There are currently two approaches to LoRA initialization aimed at preventing knowledge forgetting during fine-tuning: (1) making residual weights close to pre-trained weights, and (2) ensuring the space of LoRA initialization is orthogonal to pre-trained knowledge. The former is what current methods strive to achieve, while the importance of the latter is not sufficiently recognized. We find that the space of LoRA initialization is the key to preserving pre-trained knowledge rather than the residual weights. Existing methods like MiLoRA propose making the LoRA initialization space orthogonal to pre-trained weights. However, MiLoRA utilizes the null space of pre-trained weights. Compared to pre-trained weights, the input activations of pre-trained knowledge take into account the parameters of all previous layers as well as the input data, while pre-trained weights only contain information from the current layer. Moreover, we find that the effective ranks of input activations are much smaller than those of pre-trained weights. Thus, the null space of activations is more accurate and contains less pre-trained knowledge information compared to that of weights. Based on these, we introduce LoRA-Null, our proposed method that initializes LoRA in the null space of activations. Experimental results show that LoRA-Null effectively preserves the pre-trained world knowledge of LLMs while achieving good fine-tuning performance, as evidenced by extensive experiments. Code is available at {https://github.com/HungerPWAY/LoRA-Null}.

Method: Adds one identity layer before initial LM layers, performs preference learning only on this layer to map unaligned embeddings to aligned space, and uses interpolation coefficient during inference to control alignment strength.

Result: Efficient fine-tuning method performs comparable to full fine-tuning, with clear interpolation and extrapolation effects when controlling alignment parameter.

Conclusion: CLM enables flexible alignment control within a single model through simple architectural modification and interpolation mechanism, offering efficient preference adaptation.

Abstract: Post-training alignment has increasingly become a crucial factor in enhancing the usability of language models (LMs). However, the strength of alignment varies depending on individual preferences. This paper proposes a method to incorporate alignment control into a single model, referred to as CLM. This approach adds one identity layer preceding the initial layers and performs preference learning only on this layer to map unaligned input token embeddings into the aligned space. Experimental results demonstrate that this efficient fine-tuning method performs comparable to full fine-tuning. During inference, the input embeddings are processed through the aligned and unaligned layers, which are then merged through the interpolation coefficient. By controlling this parameter, the alignment exhibits a clear interpolation and extrapolation phenomenon.

Method: AdaSpec introduces: 1) A theoretical model to analyze and predict speculative strategy efficiency across diverse scenarios, 2) Intelligent drafting and verification algorithms that dynamically adjust speculative strategies based on real-time request loads and system configurations.

Result: Experimental results on real-world LLM service traces show AdaSpec consistently meets SLOs and achieves substantial performance improvements, delivering up to 66% speedup compared to state-of-the-art speculative inference systems.

Conclusion: AdaSpec provides an efficient adaptive solution for LLM inference that dynamically optimizes speculative decoding strategies to handle dynamic workloads while ensuring SLO compliance and significant performance gains.

Abstract: Cloud-based Large Language Model (LLM) services often face challenges in achieving low inference latency and meeting Service Level Objectives (SLOs) under dynamic request patterns. Speculative decoding, which exploits lightweight models for drafting and LLMs for verification, has emerged as a compelling technique to accelerate LLM inference. However, existing speculative decoding solutions often fail to adapt to fluctuating workloads and dynamic system environments, resulting in impaired performance and SLO violations. In this paper, we introduce AdaSpec, an efficient LLM inference system that dynamically adjusts speculative strategies according to real-time request loads and system configurations. AdaSpec proposes a theoretical model to analyze and predict the efficiency of speculative strategies across diverse scenarios. Additionally, it implements intelligent drafting and verification algorithms to maximize performance while ensuring high SLO attainment. Experimental results on real-world LLM service traces demonstrate that AdaSpec consistently meets SLOs and achieves substantial performance improvements, delivering up to 66% speedup compared to state-of-the-art speculative inference systems. The source code is publicly available at https://github.com/cerebellumking/AdaSpec

Method: Used LLMs with prompt engineering to classify spider species epithets from a dataset compiled by Mammola et al., comparing the LLM-based labeling results with human annotations across different categories (Morphology, Geography, People, Ecology & Behavior, Modern & Past Culture).

Result: LLM-based classification achieved high accuracy in Morphology, Geography, and People categories, but showed lower accuracy in Ecology & Behavior and Modern & Past Culture categories, indicating challenges in interpreting animal behavior and cultural contexts.

Conclusion: LLMs show promise for automating species name labeling but need improvement for complex categories. Future work will focus on optimizing few-shot learning and retrieval-augmented generation techniques, and expanding to diverse biological taxa.

Abstract: Scientific names of organisms consist of a genus name and a species epithet, with the latter often reflecting aspects such as morphology, ecology, distribution, and cultural background. Traditionally, researchers have manually labeled species names by carefully examining taxonomic descriptions, a process that demands substantial time and effort when dealing with large datasets. This study evaluates the feasibility of automatic species name labeling using large language model (LLM) by leveraging their text classification and semantic extraction capabilities. Using the spider name dataset compiled by Mammola et al., we compared LLM-based labeling results-enhanced through prompt engineering-with human annotations. The results indicate that LLM-based classification achieved high accuracy in Morphology, Geography, and People categories. However, classification accuracy was lower in Ecology & Behavior and Modern & Past Culture, revealing challenges in interpreting animal behavior and cultural contexts. Future research will focus on improving accuracy through optimized few-shot learning and retrieval-augmented generation techniques, while also expanding the applicability of LLM-based labeling to diverse biological taxa.

Method: Evaluated eight widely used LLMs on two datasets (POLIGEN and ECONOLEX) covering political and economic discourse. Used Socratic method to analyze LLMs’ feedback on their own outputs, examining inconsistencies in reasoning through binary comparisons.

Result: Most LLMs can accurately annotate ideologically framed text, with GPT-4o achieving human-level accuracy and high agreement with human annotators. However, Socratic probing reveals that when confronted with binary comparisons, LLMs often exhibit preference toward one perspective or perceive certain viewpoints as less biased.

Conclusion: While LLMs show promising capability in detecting ideological framing bias, their reasoning reveals inconsistencies and preference biases, highlighting the need for careful evaluation of LLM-as-a-judge systems and their potential to reinforce certain perspectives.

Abstract: Large Language Models (LLMs) are widely used for text generation, making it crucial to address potential bias. This study investigates ideological framing bias in LLM-generated articles, focusing on the subtle and subjective nature of such bias in journalistic contexts. We evaluate eight widely used LLMs on two datasets-POLIGEN and ECONOLEX-covering political and economic discourse where framing bias is most pronounced. Beyond text generation, LLMs are increasingly used as evaluators (LLM-as-a-judge), providing feedback that can shape human judgment or inform newer model versions. Inspired by the Socratic method, we further analyze LLMs’ feedback on their own outputs to identify inconsistencies in their reasoning. Our results show that most LLMs can accurately annotate ideologically framed text, with GPT-4o achieving human-level accuracy and high agreement with human annotators. However, Socratic probing reveals that when confronted with binary comparisons, LLMs often exhibit preference toward one perspective or perceive certain viewpoints as less biased.

Method: ToolACE-R introduces: 1) Model-aware iterative training that progressively adjusts training samples based on the model’s evolving capabilities; 2) Self-refinement training corpus emphasizing LLMs’ ability to iteratively refine tool calls without external feedback; 3) Adaptive self-refinement mechanism for test-time scaling where the trained model autonomously determines when to stop iterative refinement.

Result: Extensive experiments across several benchmark datasets show ToolACE-R achieves competitive performance compared to advanced API-based models. Tool invocation performance can be further improved efficiently through adaptive self-refinement, demonstrating effectiveness and generalizability.

Conclusion: ToolACE-R offers a promising direction for more efficient and scalable tool learning by maximizing model potential through iterative training and adaptive refinement, enabling LLMs to better leverage external tools for complex tasks.

Abstract: Tool learning, which allows Large Language Models (LLMs) to leverage external tools for solving complex user tasks, has emerged as a promising avenue for extending model capabilities. However, existing approaches primarily focus on data synthesis for fine-tuning LLMs to invoke tools effectively, largely ignoring how to fully stimulate the potential of the model. In this paper, we propose ToolACE-R, a novel framework that includes both model-aware iterative training and adaptive refinement for tool learning. ToolACE-R features a model-aware iterative training procedure that progressively adjust training samples based on the model’s evolving capabilities to maximize its potential. Additionally, it incorporates self-refinement training corpus which emphasizes LLM’s ability to iteratively refine their tool calls, optimizing performance without requiring external feedback. Furthermore, we introduce adaptive self-refinement mechanism for efficient test-time scaling, where the trained model can autonomously determine when to stop the process based on iterative self-refinement. We conduct extensive experiments across several benchmark datasets, showing that ToolACE-R achieves competitive performance compared to advanced API-based models. The performance of tool invocation can be further improved efficiently through adaptive self-refinement. These results highlight the effectiveness and generalizability of ToolACE-R, offering a promising direction for more efficient and scalable tool learning.

Method: Systematic analysis of prompting-based methods, parameter-efficient and full fine-tuning strategies, synthetic data generation, preference-based optimization, reinforcement learning with human/weakly supervised feedback, Mixture-of-Experts models, MT-focused LLMs, multilingual alignment, and document-level/discourse-aware approaches.

Result: LLM-based MT represents an evolution where gains increasingly depend on data quality, preference alignment, and context utilization rather than scale alone, with emerging approaches in low-resource translation, document-level processing, and specialized models.

Conclusion: LLM-based MT is positioned as an evolution of traditional systems, with open challenges remaining for building robust, inclusive, and controllable translation systems that effectively leverage data quality, preference alignment, and context utilization beyond mere scale.

Abstract: Large Language Models (LLMs) are rapidly reshaping machine translation (MT), particularly by introducing instruction-following, in-context learning, and preference-based alignment into what has traditionally been a supervised encoder-decoder paradigm. This survey provides a comprehensive and up-to-date overview of how LLMs are being leveraged for MT across data regimes, languages, and application settings. We systematically analyze prompting-based methods, parameter-efficient and full fine-tuning strategies, synthetic data generation, preference-based optimization, and reinforcement learning with human and weakly supervised feedback. Special attention is given to low-resource translation, where we examine the roles of synthetic data quality, diversity, and preference signals, as well as the limitations of current RLHF pipelines. We further review recent advances in Mixture-of-Experts models, MT-focused LLMs, and multilingual alignment, highlighting trade-offs between scalability, specialization, and accessibility. Beyond sentence-level translation, we survey emerging document-level and discourse-aware MT methods with LLMs, showing that most approaches extend sentence-level pipelines through structured context selection, post-editing, or reranking rather than requiring fundamentally new data regimes or architectures. Finally, we discuss LLM-based evaluation, its strengths and biases, and its role alongside learned metrics. Overall, this survey positions LLM-based MT as an evolution of traditional MT systems, where gains increasingly depend on data quality, preference alignment, and context utilization rather than scale alone, and outlines open challenges for building robust, inclusive, and controllable translation systems.

Method: Three-stage PDCA-inspired framework: (1) Complexity-aware exemplar selection guides plan generation by aligning decomposition granularity with question difficulty; (2) Structured retrieve-then-read execution; (3) Dual verification that identifies/corrects errors and dynamically adjusts verification strength based on question complexity.

Result: PAR-RAG consistently outperforms competitive baselines across diverse benchmarks. Ablation studies confirm the complementary roles of complexity-aware planning and dual verification.

Conclusion: PAR-RAG establishes a robust and generalizable framework for reliable multi-hop reasoning by integrating theoretical grounding with practical robustness through semantic complexity as a unifying principle.

Abstract: Retrieval-augmented generation (RAG) has demonstrated strong performance in single-hop question answering (QA) by integrating external knowledge into large language models (LLMs). However, its effectiveness remains limited in multi-hop QA, which demands both stable reasoning and factual consistency. Existing approaches often provide partial solutions, addressing either reasoning trajectory stability or factual verification, but rarely achieving both simultaneously. To bridge this gap, we propose PAR-RAG, a three-stage Plan-then-Act-and-Review framework inspired by the PDCA cycle. PAR-RAG incorporates semantic complexity as a unifying principle through three key components: (i) complexity-aware exemplar selection guides plan generation by aligning decomposition granularity with question difficulty, thereby stabilizing reasoning trajectories; (ii) execution follows a structured retrieve-then-read process; and (iii) dual verification identifies and corrects intermediate errors while dynamically adjusting verification strength based on question complexity: emphasizing accuracy for simple queries and multi-evidence consistency for complex ones. This cognitively inspired framework integrates theoretical grounding with practical robustness. Experiments across diverse benchmarks demonstrate that PAR-RAG consistently outperforms competitive baselines, while ablation studies confirm the complementary roles of complexity-aware planning and dual verification. Collectively, these results establish PAR-RAG as a robust and generalizable framework for reliable multi-hop reasoning.

Method: Created WiserUI-Bench with 300 real-world UI image pairs from industry A/B tests, each with empirically validated winners that induced more user actions. Includes expert-curated key interpretations for each instance. Evaluated multiple MLLMs on two tasks: predicting more effective UI and explaining it post-hoc.

Result: Experiments show MLLMs exhibit limited understanding of the behavioral impact of UI/UX design. Models struggle to predict which UI designs are more effective and explain why they succeed in alignment with expert interpretations.

Conclusion: The work introduces a benchmark to foster research on leveraging MLLMs for visual design in user behavior contexts, highlighting current limitations in understanding design-behavior relationships.

Abstract: User interface (UI) design goes beyond visuals to shape user experience (UX), underscoring the shift toward UI/UX as a unified concept. While recent studies have explored UI evaluation using Multimodal Large Language Models (MLLMs), they largely focus on surface-level features, overlooking how design choices influence user behavior at scale. To fill this gap, we introduce WiserUI-Bench, a novel benchmark for multimodal understanding of how UI/UX design affects user behavior, built on 300 real-world UI image pairs from industry A/B tests, with empirically validated winners that induced more user actions. For future design progress in practice, post-hoc understanding of why such winners succeed with mass users is also required; we support this via expert-curated key interpretations for each instance. Experiments across multiple MLLMs on WiserUI-Bench for two main tasks, (1) predicting the more effective UI image between an A/B-tested pair, and (2) explaining it post-hoc in alignment with expert interpretations, show that models exhibit limited understanding of the behavioral impact of UI/UX design. We believe our work will foster research on leveraging MLLMs for visual design in user behavior contexts.

Method: Unsupervised framework using clustering for aspect category extraction, aspect-aware embedding vectors via negative sampling, and validation through multi-aspect labeling with pretrained language models.

Result: Models achieve high performance with automatically generated labels, showing superior consistency/scalability vs. LLMs, and human evaluation confirms label quality comparable to manual labeling.

Conclusion: Demonstrates robust multi-aspect labeling approach overcoming supervised method limitations, adaptable to multilingual, multi-domain environments, with future work on automatic summarization and AI agent integration.

Abstract: Effectively analyzing online review data is essential across industries. However, many existing studies are limited to specific domains and languages or depend on supervised learning approaches that require large-scale labeled datasets. To address these limitations, we propose a multilingual, scalable, and unsupervised framework for cross-domain aspect detection. This framework is designed for multi-aspect labeling of multilingual and multi-domain review data. In this study, we apply automatic labeling to Korean and English review datasets spanning various domains and assess the quality of the generated labels through extensive experiments. Aspect category candidates are first extracted through clustering, and each review is then represented as an aspect-aware embedding vector using negative sampling. To evaluate the framework, we conduct multi-aspect labeling and fine-tune several pretrained language models to measure the effectiveness of the automatically generated labels. Results show that these models achieve high performance, demonstrating that the labels are suitable for training. Furthermore, comparisons with publicly available large language models highlight the framework’s superior consistency and scalability when processing large-scale data. A human evaluation also confirms that the quality of the automatic labels is comparable to those created manually. This study demonstrates the potential of a robust multi-aspect labeling approach that overcomes limitations of supervised methods and is adaptable to multilingual, multi-domain environments. Future research will explore automatic review summarization and the integration of artificial intelligence agents to further improve the efficiency and depth of review analysis.

Method: Identify model-agnostic prompt quality merits, validate them empirically, then train MePO (merit-guided prompt optimizer) on a merit-guided prompt preference dataset generated by a lightweight LLM for local deployment.

Result: MePO achieves better results across diverse tasks and model types, offering scalable and robust solutions for real-world deployment while reducing privacy concerns.

Conclusion: MePO provides an explicit, interpretable, and locally deployable approach to prompt optimization that generalizes effectively across different model scales, addressing the limitations of existing self-generation methods.

Abstract: Prompt optimization (PO) provides a practical way to improve response quality when users lack the time or expertise to manually craft effective prompts. Existing methods typically rely on LLMs’ self-generation ability to optimize prompts. However, due to limited downward compatibility, the instruction-heavy prompts generated by advanced LLMs can overwhelm lightweight inference models and degrade response quality, while also lacking interpretability due to implicit optimization. In this work, we rethink prompt optimization through the lens of explicit and interpretable design. We first identify a set of model-agnostic prompt quality merits and empirically validate their effectiveness in enhancing prompt and response quality. We then introduce MePO, a merit-guided, locally deployable prompt optimizer trained on our merit-guided prompt preference dataset generated by a lightweight LLM. MePO avoids online optimization, reduces privacy concerns, and, by learning clear, interpretable merits, generalizes effectively to both large-scale and lightweight inference models. Experiments demonstrate that MePO achieves better results across diverse tasks and model types, offering a scalable and robust solution for real-world deployment. The code, model and dataset can be found in https://github.com/MidiyaZhu/MePO

Method: Leverages knowledge-graph-based document representations to extract QA pairs at multiple complexity levels along three key dimensions: multi-hop retrieval, set operations, and answer plurality.

Result: Constructed a dataset of 20,139 QA pairs from financial credit agreements and evaluated 16 proprietary and open-weight LLMs, finding that even best-performing models struggle with set-based comparisons and multi-hop reasoning over long contexts.

Conclusion: The framework reveals systematic failure modes in LLMs tied to semantic misinterpretation and inability to handle implicit relations, highlighting limitations in complex reasoning capabilities despite overall strong performance.

Abstract: We introduce KG-MuLQA (Knowledge-Graph-based Multi-Level Question-Answer Extraction): a framework that (1) extracts QA pairs at multiple complexity levels (2) along three key dimensions – multi-hop retrieval, set operations, and answer plurality, (3) by leveraging knowledge-graph-based document representations. This approach enables fine-grained assessment of model performance across controlled difficulty levels. Using this framework, we construct a dataset of 20,139 QA pairs based on financial credit agreements and evaluate 16 proprietary and open-weight Large Language Models, observing that even the best-performing models struggle with set-based comparisons and multi-hop reasoning over long contexts. Our analysis reveals systematic failure modes tied to semantic misinterpretation and inability to handle implicit relations.

Method: Uses the Min-Max theory of communicative interaction and analyzes a massive dataset of 1,942 language corpora tagged for parts of speech (Ring 2025), examining correlations between average word class lengths and word order.

Result: Average lengths of particular word classes correlate with word order, allowing prediction of basic word order from diverse corpora. Word class length provides stronger explanation than genealogical or areal factors.

Conclusion: The Min-Max theory offers a general universal explanation for word order change, unifying efficiency and surprisal approaches, and highlights the importance of language corpora for investigating linguistic universals.

Abstract: A fundamental concern in linguistics has been to understand how languages change, such as in relation to word order. Since the order of words in a sentence (i.e. the relative placement of Subject, Object, and Verb) is readily identifiable in most languages, this has been a productive field of study for decades (see Greenberg 1963; Dryer 2007; Hawkins 2014). However, a language’s word order can change over time, with competing explanations for such changes (Carnie and Guilfoyle 2000; Crisma and Longobardi 2009; Martins and Cardoso 2018; Dunn et al. 2011; Jager and Wahle 2021). This paper proposes a general universal explanation for word order change based on a theory of communicative interaction (the Min-Max theory of language behavior) in which agents seek to minimize effort while maximizing information. Such an account unifies opposing findings from language processing (Piantadosi et al. 2011; Wasow 2022; Levy 2008) that make different predictions about how word order should be realized crosslinguistically. The marriage of both “efficiency” and “surprisal” approaches under the Min-Max theory is justified with evidence from a massive dataset of 1,942 language corpora tagged for parts of speech (Ring 2025), in which average lengths of particular word classes correlates with word order, allowing for prediction of basic word order from diverse corpora. The general universal pressure of word class length in corpora is shown to give a stronger explanation for word order realization than either genealogical or areal factors, highlighting the importance of language corpora for investigating such questions.

Method: Two-stage approach: 1) Use small curated data to develop initial judgment thinking capabilities, 2) Optimize judgment thinking traces using reinforcement learning with two methods: offline (training critic model to construct examples) and online (using rule-based rewards as feedback).

Result: Think-J significantly enhances generative LLM-Judge evaluation capabilities, surpassing both generative and classifier-based LLM-Judge approaches without requiring additional human annotations.

Conclusion: Teaching generative LLMs how to think through reinforcement learning optimization effectively improves their performance as LLM-Judge, offering a promising approach for automatic preference modeling in LLM evaluation and reward modeling.

Abstract: LLM-as-a-Judge refers to the automatic modeling of preferences for responses generated by Large Language Models (LLMs), which is of significant importance for both LLM evaluation and reward modeling. Although generative LLMs have made substantial progress in various tasks, their performance as LLM-Judge still falls short of expectations. In this work, we propose Think-J, which improves generative LLM-as-a-Judge by learning how to think. We first utilized a small amount of curated data to develop the model with initial judgment thinking capabilities. Subsequently, we optimize the judgment thinking traces based on reinforcement learning (RL). We propose two methods for judgment thinking optimization, based on offline and online RL, respectively. The offline method requires training a critic model to construct positive and negative examples for learning. The online method defines rule-based reward as feedback for optimization. Experimental results showed that our approach can significantly enhance the evaluation capability of generative LLM-Judge, surpassing both generative and classifier-based LLM-Judge without requiring extra human annotations.

Method: Created ALTPRAG dataset grounded in pragmatic concept of alternatives, where each instance pairs two equally plausible but pragmatically divergent continuations. Models must: (1) infer speaker’s intended meaning, and (2) explain why a speaker would choose one utterance over its alternative. Evaluated 22 LLMs across 3 training stages: post-pretraining, supervised fine-tuning (SFT), and preference optimization.

Result: Base models already show notable sensitivity to pragmatic cues, improving with model and data scale. SFT and RLHF provide additional gains, especially in cognitive-pragmatic scenarios. Pragmatic competence emerges as a compositional property of LLM training.

Conclusion: Pragmatic competence is an emergent property that develops throughout LLM training, with improvements at each stage. The findings offer insights for aligning models with human communicative norms and understanding how social intelligence capabilities emerge in language models.

Abstract: Current large language models (LLMs) have demonstrated emerging capabilities in social intelligence tasks, including implicature resolution and theory-of-mind reasoning, both of which require substantial pragmatic understanding. However, how LLMs acquire this pragmatic competence throughout the training process remains poorly understood. In this work, we introduce ALTPRAG, a dataset grounded in the pragmatic concept of alternatives, to evaluate whether LLMs at different training stages can accurately infer nuanced speaker intentions. Each instance pairs two equally plausible yet pragmatically divergent continuations and requires the model to (i) infer the speaker’s intended meaning and (ii) explain when and why a speaker would choose one utterance over its alternative, thus directly probing pragmatic competence through contrastive reasoning. We systematically evaluate 22 LLMs across 3 key training stages: after pre-training, supervised fine-tuning (SFT), and preference optimization, to examine the development of pragmatic competence. Our results show that even base models exhibit notable sensitivity to pragmatic cues, which improves consistently with increases in model and data scale. Additionally, SFT and RLHF contribute further gains, particularly in cognitive-pragmatic scenarios. These findings highlight pragmatic competence as an emergent and compositional property of LLM training and offer new insights for aligning models with human communicative norms.

Method: Created BnMMLU benchmark spanning 41 domains across STEM, humanities, social sciences, and general knowledge with 134,375 multiple-choice questions. Includes MathML for mathematical content and BnMMLU-HARD subset of frequently missed questions. Evaluated 24 model variants across 11 LLM families using standardized protocols with two prompting styles (Direct vs. Chain-of-Thought) and two context regimes (0-shot vs. 5-shot).

Result: Benchmarking revealed persistent gaps in reasoning and application skills across models. Analysis showed sublinear returns to scale across different model sizes. The dataset is the most extensive Bengali evaluation suite to date.

Conclusion: BnMMLU enables rigorous, reproducible assessment of Bengali language understanding and aims to catalyze progress in multilingual NLP. The released dataset and evaluation templates support standardized benchmarking and highlight areas needing improvement in Bengali language models.

Abstract: Large-scale multitask benchmarks have driven rapid progress in language modeling, yet most emphasize high-resource languages such as English, leaving Bengali underrepresented. We present BnMMLU, a comprehensive benchmark for measuring massive multitask language understanding in Bengali. BnMMLU spans 41 domains across STEM, humanities, social sciences, and general knowledge, and contains 134,375 multiple-choice question-option pairs–the most extensive Bengali evaluation suite to date. The dataset preserves mathematical content via MathML, and includes BnMMLU-HARD, a compact subset constructed from questions most frequently missed by top systems to stress difficult cases. We benchmark 24 model variants across 11 LLM families, spanning open-weights general/multilingual, Bengali-centric open-weights, and proprietary models, covering multiple parameter scales and instruction-tuned settings. We evaluate models under standardized protocols covering two prompting styles (Direct vs. Chain-of-Thought) and two context regimes (0-shot vs. 5-shot), reporting accuracy consistently across families. Our analysis highlights persistent gaps in reasoning and application skills and indicates sublinear returns to scale across model sizes. We release the dataset and evaluation templates to support rigorous, reproducible assessment of Bengali language understanding and to catalyze progress in multilingual NLP.

Method: PLI inserts premature layers formed through mathematical interpolation with adjacent layers, inspired by stable diffusion and sampling steps, extending the depth of information processing and transmission in LLMs without training.

Result: Experiments on four publicly available datasets show PLI effectively reduces hallucinations and outperforms existing baselines in most cases. Analysis suggests layer interpolation success is linked to LLMs’ internal mechanisms.

Conclusion: PLI offers a novel, training-free intervention for enhancing LLM factuality by leveraging layer interpolation to improve information refinement, demonstrating effectiveness across multiple datasets with available open-source implementation.

Abstract: Large Language Models (LLMs) demonstrate remarkable capabilities in text understanding and generation. However, their tendency to produce factually inconsistent outputs, commonly referred to as ‘‘hallucinations’’, remains a critical challenge. Existing approaches, such as retrieval-based and inference-time correction methods, primarily address this issue at the input or output level, often overlooking the intrinsic information refinement process and the role of premature layers. Meanwhile, alignment- and fine-tuning-based methods are resource-intensive. In this paper, we propose PLI (Premature Layers Interpolation), a novel, training-free, and plug-and-play intervention designed to enhance factuality. PLI mitigates hallucinations by inserting premature layers formed through mathematical interpolation with adjacent layers. Inspired by stable diffusion and sampling steps, PLI extends the depth of information processing and transmission in LLMs, improving factual coherence. Experiments on four publicly available datasets demonstrate that PLI effectively reduces hallucinations while outperforming existing baselines in most cases. Further analysis suggests that the success of layer interpolation is closely linked to LLMs’ internal mechanisms. Our dataset and code are available at https://github.com/CuSO4-Chen/PLI.

Method: Controlled evaluation for QA tasks under systematically scaled token budgets, comparing two recent multi-stage pipelines (ReadAgent and RAPTOR) against three baselines including DOS RAG (Document’s Original Structure RAG) - a simple retrieve-then-read method that preserves original passage order.

Result: DOS RAG consistently matches or outperforms more intricate methods on multiple long-context QA benchmarks. Its strength comes from maintaining source fidelity and document structure, prioritizing recall within effective context windows, and favoring simplicity over added pipeline complexity.

Conclusion: DOS RAG should be established as a simple yet strong baseline for future RAG evaluations, paired with state-of-the-art embedding and language models, and benchmarked under matched token budgets to ensure added pipeline complexity is justified by clear performance gains.

Abstract: With the rise of long-context language models (LMs) capable of processing tens of thousands of tokens in a single context window, do multi-stage retrieval-augmented generation (RAG) pipelines still offer measurable benefits over simpler, single-stage approaches? To assess this question, we conduct a controlled evaluation for QA tasks under systematically scaled token budgets, comparing two recent multi-stage pipelines, ReadAgent and RAPTOR, against three baselines, including DOS RAG (Document’s Original Structure RAG), a simple retrieve-then-read method that preserves original passage order. Despite its straightforward design, DOS RAG consistently matches or outperforms more intricate methods on multiple long-context QA benchmarks. We trace this strength to a combination of maintaining source fidelity and document structure, prioritizing recall within effective context windows, and favoring simplicity over added pipeline complexity. We recommend establishing DOS RAG as a simple yet strong baseline for future RAG evaluations, paired with state-of-the-art embedding and language models, and benchmarked under matched token budgets, to ensure that added pipeline complexity is justified by clear performance gains as models continue to improve.

Method: Two-component framework: 1) Training-time Realignment (TrRa) uses controllable fusion of logits from reference and aligned models for efficient realignment. 2) Inference-time Realignment (InRa) uses a layer adapter initialized for identity transformation, inserted before original layers, with controllable interpolation at logit level.

Result: TrRa reduces token usage by 54.63% on DeepSeek-R1-Distill-Qwen-1.5B without performance degradation (outperforming DeepScaleR-1.5B’s 33.86%). Upgraded DeepSeek-R1-Distill-Qwen-7B to support both fast and slow thinking with flexible alignment control during inference, even surpassing original performance through deeper reasoning.

Conclusion: The proposed realignment framework provides flexible quantitative control over alignment degree during both training and inference, enabling efficient model realignment with significant token usage reduction and performance improvements, including upgrading models to support multiple thinking modes.

Abstract: Realignment becomes necessary when a language model (LM) fails to meet expected performance. We propose a flexible realignment framework that supports quantitative control of alignment degree during training and inference. This framework incorporates Training-time Realignment (TrRa), which efficiently realigns the reference model by leveraging the controllable fusion of logits from both the reference and already aligned models. For example, TrRa reduces token usage by 54.63% on DeepSeek-R1-Distill-Qwen-1.5B without any performance degradation, outperforming DeepScaleR-1.5B’s 33.86%. To complement TrRa during inference, we introduce a layer adapter that enables smooth Inference-time Realignment (InRa). This adapter is initialized to perform an identity transformation at the bottom layer and is inserted preceding the original layers. During inference, input embeddings are simultaneously processed by the adapter and the original layer, followed by the remaining layers, and then controllably interpolated at the logit level. We upgraded DeepSeek-R1-Distill-Qwen-7B from a slow-thinking model to one that supports both fast and slow thinking, allowing flexible alignment control even during inference. By encouraging deeper reasoning, it even surpassed its original performance.

Method: Systematically evaluated different reasoning settings across model sizes, distribution expression methods, and steering methods, resulting in 60 experimental setups across 3 tasks. Compared RLVR-style reasoning with naive Chain-of-Thought reasoning for disagreement modeling.

Result: Surprisingly, RLVR-style reasoning degrades performance in disagreement modeling, while naive Chain-of-Thought reasoning improves the performance of RLHF LLMs. The findings were consistent across the experimental setups.

Conclusion: There is potential risk in replacing human annotators with reasoning LLMs, especially when disagreements are important. RLVR-style reasoning, despite improving performance on many tasks, actually harms disagreement modeling capability, while simpler CoT reasoning benefits RLHF LLMs in this specific context.

Abstract: Variation in human annotation (i.e., disagreements) is common in NLP, often reflecting important information like task subjectivity and sample ambiguity. Modeling this variation is important for applications that are sensitive to such information. Although RLVR-style reasoning (Reinforcement Learning with Verifiable Rewards) has improved Large Language Model (LLM) performance on many tasks, it remains unclear whether such reasoning enables LLMs to capture informative variation in human annotation. In this work, we evaluate the influence of different reasoning settings on LLM disagreement modeling. We systematically evaluate each reasoning setting across model sizes, distribution expression methods, and steering methods, resulting in 60 experimental setups across 3 tasks. Surprisingly, our results show that RLVR-style reasoning degrades performance in disagreement modeling, while naive Chain-of-Thought (CoT) reasoning improves the performance of RLHF LLMs (RL from human feedback). These findings underscore the potential risk of replacing human annotators with reasoning LLMs, especially when disagreements are important.

Method: Proposes a multi-perspective approach using soft labels to develop perspective-aware models. Tests across diverse subjective text classification tasks (hate speech, irony, abusive language, stance detection) and uses Jensen-Shannon Divergence (JSD) to measure approximation to human label distributions.

Result: The multi-perspective approach better approximates human label distributions (lower JSD) and achieves superior classification performance (higher F1 scores) compared to traditional approaches. However, shows lower confidence in highly subjective tasks like irony and stance detection. Explainable AI (XAI) reveals meaningful insights into model uncertainty and predictions.

Conclusion: Capturing human disagreement through multi-perspective modeling leads to more inclusive, pluralistic models that better represent diverse viewpoints, especially in subjective NLP tasks where traditional aggregation methods overlook important minority perspectives.

Abstract: In the realm of Natural Language Processing (NLP), common approaches for handling human disagreement consist of aggregating annotators’ viewpoints to establish a single ground truth. However, prior studies show that disregarding individual opinions can lead can lead to the side effect of underrepresenting minority perspectives, especially in subjective tasks, where annotators may systematically disagree because of their preferences. Recognizing that labels reflect the diverse backgrounds, life experiences, and values of individuals, this study proposes a new multi-perspective approach using soft labels to encourage the development of the next generation of perspective aware models, more inclusive and pluralistic. We conduct an extensive analysis across diverse subjective text classification tasks, including hate speech, irony, abusive language, and stance detection, to highlight the importance of capturing human disagreements, often overlooked by traditional aggregation methods. Results show that the multi-perspective approach not only better approximates human label distributions, as measured by Jensen-Shannon Divergence (JSD), but also achieves superior classification performance (higher F1 scores), outperforming traditional approaches. However, our approach exhibits lower confidence in tasks like irony and stance detection, likely due to the inherent subjectivity present in the texts. Lastly, leveraging Explainable AI (XAI), we explore model uncertainty and uncover meaningful insights into model predictions.

Method: Two-step approach: (1) Use the logit of the last token to speculate the next next token, (2) Retrieve relevant references for both the next token and the speculated next next token to generate draft tokens. Training-free and plug-and-play.

Result: Achieves up to 2.61× speedup and 3.28 mean accepted tokens per decoding step across various text generation benchmarks. Outperforms existing retrieval-based speculative decoding methods.

Conclusion: LogitSpec effectively addresses the limitation of retrieval-based speculative decoding by leveraging logit information to expand retrieval range, resulting in significant inference acceleration without requiring training or draft models.

Abstract: Speculative decoding (SD), where a small draft model is employed to propose draft tokens in advance and then the target model validates them in parallel, has emerged as a promising technique for LLM inference acceleration. Many endeavors to improve SD are to eliminate the need for a draft model and generate draft tokens in a retrieval-based manner in order to further alleviate the drafting overhead and significantly reduce the difficulty in deployment and applications. However, retrieval-based SD relies on a matching paradigm to retrieval the most relevant reference as the draft tokens, where these methods often fail to find matched and accurate draft tokens. To address this challenge, we propose LogitSpec to effectively expand the retrieval range and find the most relevant reference as drafts. Our LogitSpec is motivated by the observation that the logit of the last token can not only predict the next token, but also speculate the next next token. Specifically, LogitSpec generates draft tokens in two steps: (1) utilizing the last logit to speculate the next next token; (2) retrieving relevant reference for both the next token and the next next token. LogitSpec is training-free and plug-and-play, which can be easily integrated into existing LLM inference frameworks. Extensive experiments on a wide range of text generation benchmarks demonstrate that LogitSpec can achieve up to 2.61 $\times$ speedup and 3.28 mean accepted tokens per decoding step. Our code is available at https://github.com/smart-lty/LogitSpec.

Method: Used annotated corpora from three domains: healthcare (I2B2), legal (TAB), and biography (Wikipedia). Evaluated models across four dimensions: in-domain performance, cross-domain transferability, data fusion, and few-shot learning.

Result: Legal-domain data transfers well to biographical texts, while medical domains resist incoming transfer. Fusion benefits are domain-specific. High-quality PII recognition is achievable with only 10% of training data in low-specialization domains.

Conclusion: Cross-domain transfer for PII recognition shows promising results with legal-to-biography transfer, but medical data requires domain-specific approaches. Efficient learning is possible with minimal data in less specialized domains.

Abstract: Accurate recognition of personally identifiable information (PII) is central to automated text anonymization. This paper investigates the effectiveness of cross-domain model transfer, multi-domain data fusion, and sample-efficient learning for PII recognition. Using annotated corpora from healthcare (I2B2), legal (TAB), and biography (Wikipedia), we evaluate models across four dimensions: in-domain performance, cross-domain transferability, fusion, and few-shot learning. Results show legal-domain data transfers well to biographical texts, while medical domains resist incoming transfer. Fusion benefits are domain-specific, and high-quality recognition is achievable with only 10% of training data in low-specialization domains.

Method: Used XLM-RoBERTa-large transformer model to predict four punctuation marks (period, comma, question mark, exclamation mark). Built large training corpus with data augmentation (alpha = 0.20%) to address resource scarcity.

Result: Achieved 97.1% accuracy on News test set, 91.2% on Reference set, and 90.2% on ASR set. Model shows strong generalization to reference texts and noisy ASR transcripts.

Conclusion: Establishes strong baseline for Bangla punctuation restoration, demonstrates effectiveness in real-world scenarios, and contributes publicly available datasets/code for low-resource NLP research.

Abstract: Punctuation restoration enhances the readability of text and is critical for post-processing tasks in Automatic Speech Recognition (ASR), especially for low-resource languages like Bangla. In this study, we explore the application of transformer-based models, specifically XLM-RoBERTa-large, to automatically restore punctuation in unpunctuated Bangla text. We focus on predicting four punctuation marks: period, comma, question mark, and exclamation mark across diverse text domains. To address the scarcity of annotated resources, we constructed a large, varied training corpus and applied data augmentation techniques. Our best-performing model, trained with an augmentation factor of alpha = 0.20%, achieves an accuracy of 97.1% on the News test set, 91.2% on the Reference set, and 90.2% on the ASR set. Results show strong generalization to reference and ASR transcripts, demonstrating the model’s effectiveness in real-world, noisy scenarios. This work establishes a strong baseline for Bangla punctuation restoration and contributes publicly available datasets and code to support future research in low-resource NLP.

Method: Formulates CKGE as sequential Bayesian inference using posterior update principle as continual learning strategy. Treats each batch of new data as Bayesian update to model’s prior, maintaining posterior distribution to preserve earlier knowledge. Introduces continual clustering method with regularization term to maintain compact cluster structure of entity embeddings, ensuring semantic consistency while allowing controlled adaptation.

Result: Extensive experiments on multiple CKGE benchmarks demonstrate BAKE achieves top performance in vast majority of cases compared to existing approaches.

Conclusion: BAKE effectively addresses catastrophic forgetting in continual knowledge graph embedding through Bayesian inference framework with clustering regularization, enabling preservation of prior knowledge while adapting to new information in dynamic social media environments.

Abstract: As social media and the World Wide Web become hubs for information dissemination, effectively organizing and understanding the vast amounts of dynamically evolving Web content is crucial. Knowledge graphs (KGs) provide a powerful framework for structuring this information. However, the rapid emergence of new hot topics, user relationships, and events in social media renders traditional static knowledge graph embedding (KGE) models rapidly outdated. Continual Knowledge Graph Embedding (CKGE) aims to address this issue, but existing methods commonly suffer from catastrophic forgetting, whereby older, but still valuable, information is lost when learning new knowledge (such as new memes or trending events). This means the model cannot effectively learn the evolution of the data. We propose a novel CKGE framework, BAKE. Unlike existing methods, BAKE formulates CKGE as a sequential Bayesian inference problem and utilizes the Bayesian posterior update principle as a natural continual learning strategy. This principle is insensitive to data order and provides theoretical guarantees to preserve prior knowledge as much as possible. Specifically, we treat each batch of new data as a Bayesian update to the model’s prior. By maintaining the posterior distribution, the model effectively preserves earlier knowledge even as it evolves over multiple snapshots. Furthermore, to constrain the evolution of knowledge across snapshots, we introduce a continual clustering method that maintains the compact cluster structure of entity embeddings through a regularization term, ensuring semantic consistency while allowing controlled adaptation to new knowledge. We conduct extensive experiments on multiple CKGE benchmarks, which demonstrate that BAKE achieves the top performance in the vast majority of cases compared to existing approaches.

Method: SAEMark uses inference-time, feature-based rejection sampling without altering model logits. It extracts deterministic features from generated text and selects outputs whose feature statistics align with key-derived targets. The framework uses Sparse Autoencoders (SAEs) as feature extractors and operates through sampling LLM outputs instead of modifying them.

Result: Achieves 99.7% F1 score on English datasets with strong multi-bit detection accuracy. Shows consistent performance across 4 datasets, superior detection accuracy and text quality preservation compared to existing methods.

Conclusion: SAEMark establishes a new paradigm for scalable watermarking that works out-of-the-box with closed-source LLMs, generalizes across languages and domains, preserves text quality, and enables content attribution while providing theoretical guarantees for watermark success probability.

Abstract: Watermarking LLM-generated text is critical for content attribution and misinformation prevention. However, existing methods compromise text quality, require white-box model access and logit manipulation. These limitations exclude API-based models and multilingual scenarios. We propose SAEMark, a general framework for post-hoc multi-bit watermarking that embeds personalized messages solely via inference-time, feature-based rejection sampling without altering model logits or requiring training. Our approach operates on deterministic features extracted from generated text, selecting outputs whose feature statistics align with key-derived targets. This framework naturally generalizes across languages and domains while preserving text quality through sampling LLM outputs instead of modifying. We provide theoretical guarantees relating watermark success probability and compute budget that hold for any suitable feature extractor. Empirically, we demonstrate the framework’s effectiveness using Sparse Autoencoders (SAEs), achieving superior detection accuracy and text quality. Experiments across 4 datasets show SAEMark’s consistent performance, with 99.7% F1 on English and strong multi-bit detection accuracy. SAEMark establishes a new paradigm for scalable watermarking that works out-of-the-box with closed-source LLMs while enabling content attribution.

Method: Reframe detection as a signal processing problem by analyzing token log-probability sequences in frequency domain using global Discrete Fourier Transform (DFT) and local Short-Time Fourier Transform (STFT). Construct SpecDetect based on DFT total energy feature, and enhanced SpecDetect++ with sampling discrepancy mechanism.

Result: Human-written text consistently exhibits significantly higher spectral energy than LLM-generated text. SpecDetect outperforms state-of-the-art models while running in nearly half the time.

Conclusion: Signal processing techniques offer an efficient, interpretable pathway for LLM-generated text detection, showing classical methods provide powerful solutions to modern challenges.

Abstract: The proliferation of high-quality text from Large Language Models (LLMs) demands reliable and efficient detection methods. While existing training-free approaches show promise, they often rely on surface-level statistics and overlook fundamental signal properties of the text generation process. In this work, we reframe detection as a signal processing problem, introducing a novel paradigm that analyzes the sequence of token log-probabilities in the frequency domain. By systematically analyzing the signal’s spectral properties using the global Discrete Fourier Transform (DFT) and the local Short-Time Fourier Transform (STFT), we find that human-written text consistently exhibits significantly higher spectral energy. This higher energy reflects the larger-amplitude fluctuations inherent in human writing compared to the suppressed dynamics of LLM-generated text. Based on this key insight, we construct SpecDetect, a detector built on a single, robust feature from the global DFT: DFT total energy. We also propose an enhanced version, SpecDetect++, which incorporates a sampling discrepancy mechanism to further boost robustness. Extensive experiments show that our approach outperforms the state-of-the-art model while running in nearly half the time. Our work introduces a new, efficient, and interpretable pathway for LLM-generated text detection, showing that classical signal processing techniques offer a surprisingly powerful solution to this modern challenge.

Method: The paper proposes two contributions: 1) MuDRiC dataset - an extended Arabic commonsense dataset incorporating multiple dialects, and 2) a novel method adapting Graph Convolutional Networks (GCNs) to Arabic commonsense reasoning to enhance semantic relationship modeling.

Result: Experimental results demonstrate that the proposed GCN-based approach consistently outperforms the baseline of direct language model fine-tuning for Arabic commonsense validation.

Conclusion: This work enhances Arabic natural language understanding by providing a foundational multi-dialect dataset and a new method for handling Arabic’s complex linguistic variations, with data and code publicly available.

Abstract: Commonsense validation evaluates whether a sentence aligns with everyday human understanding, a critical capability for developing robust natural language understanding systems. While substantial progress has been made in English, the task remains underexplored in Arabic, particularly given its rich linguistic diversity. Existing Arabic resources have primarily focused on Modern Standard Arabic (MSA), leaving regional dialects underrepresented despite their prevalence in spoken contexts. To bridge this gap, we present two key contributions. We introduce MuDRiC, an extended Arabic commonsense dataset incorporating multiple dialects. To the best of our knowledge, this is the first Arabic multi-dialect commonsense reasoning dataset. We further propose a novel method adapting Graph Convolutional Networks (GCNs) to Arabic commonsense reasoning, which enhances semantic relationship modeling for improved commonsense validation. Our experimental results demonstrate that this approach consistently outperforms the baseline of direct language model fine-tuning. Overall, our work enhances Arabic natural language understanding by providing a foundational dataset and a new method for handling its complex variations. Data and code are available at https://github.com/KareemElozeiri/MuDRiC.

Method: Created VMMU benchmark with 2.5k multimodal questions across 7 diverse tasks: STEM problem solving, data interpretation, rule-governed visual reasoning, and abstract visual reasoning. All questions require genuine multimodal integration (no text-only shortcuts). Evaluated diverse state-of-the-art proprietary and open-source VLMs.

Result: Proprietary models achieve only 66% mean accuracy despite strong Vietnamese OCR performance. Analysis shows primary failure source is not OCR, but multimodal grounding and reasoning over text and visual evidence. Open-source models perform even worse.

Conclusion: Current VLMs struggle with genuine multimodal reasoning in Vietnamese context, revealing limitations beyond OCR capabilities. The benchmark highlights the need for better multimodal integration and reasoning abilities in non-English languages.

Abstract: We introduce VMMU, a Vietnamese Multitask Multimodal Understanding and Reasoning Benchmark designed to evaluate how vision-language models (VLMs) interpret and reason over visual and textual information beyond English. VMMU consists of 2.5k multimodal questions across 7 tasks, covering a diverse range of problem contexts, including STEM problem solving, data interpretation, rule-governed visual reasoning, and abstract visual reasoning. All questions require genuine multimodal integration, rather than reliance on text-only cues or OCR-based shortcuts. We evaluate a diverse set of state-of-the-art proprietary and open-source VLMs on VMMU. Despite strong Vietnamese OCR performance, proprietary models achieve only 66% mean accuracy. Further analysis shows that the primary source of failure is not OCR, but instead multimodal grounding and reasoning over text and visual evidence. Code and data are available at https://vmmu.github.io.

Method: Proposes Cognitive Surgery (CoSur) framework with four modules: representation extraction, subspace construction, authorship discrimination, and cognitive editing to mitigate implicit self-recognition (ISR).

Result: CoSur improves self-recognition performance for three different LLMs in IPP scenarios, achieving average accuracies of 99.00%, 97.69%, and 97.13% respectively.

Conclusion: The paper successfully identifies ISR as the cause of LLMs’ poor self-recognition in IPP scenarios and demonstrates that CoSur effectively bridges the gap between internal representations and output behavior.

Abstract: Large language models (LLMs) have been shown to possess a degree of self-recognition ability, which used to identify whether a given text was generated by themselves. Prior work has demonstrated that this capability is reliably expressed under the pair presentation paradigm (PPP), where the model is presented with two texts and asked to choose which one it authored. However, performance deteriorates sharply under the individual presentation paradigm (IPP), where the model is given a single text to judge authorship. Although this phenomenon has been observed, its underlying causes have not been systematically analyzed. In this paper, we first investigate the cause of this failure and attribute it to implicit self-recognition (ISR). ISR describes the gap between internal representations and output behavior in LLMs: under the IPP scenario, the model encodes self-recognition information in its feature space, yet its ability to recognize self-generated texts remains poor. To mitigate the ISR of LLMs, we propose cognitive surgery (CoSur), a novel framework comprising four main modules: representation extraction, subspace construction, authorship discrimination, and cognitive editing. Experimental results demonstrate that our proposed method improves the self-recognition performance of three different LLMs in the IPP scenario, achieving average accuracies of 99.00%, 97.69%, and 97.13%, respectively.

Method: Created LongEmotion benchmark with Emotion Recognition, Knowledge Application, and Empathetic Generation tasks. Introduced Collaborative Emotional Modeling (CoEM) framework integrating RAG and multi-agent collaboration.

Result: Conducted detailed analysis of various models in long-context settings, investigating reasoning mode activation, RAG-based retrieval strategies, and context-length adaptability on EI performance.

Conclusion: LongEmotion addresses the gap in evaluating LLMs’ emotional intelligence in long-context scenarios, and CoEM framework enhances performance under realistic constraints through collaborative approaches.

Abstract: Large language models (LLMs) have made significant progress in Emotional Intelligence (EI) and long-context modeling. However, existing benchmarks often overlook the fact that emotional information processing unfolds as a continuous long-context process. To address the absence of multidimensional EI evaluation in long-context inference and explore model performance under more challenging conditions, we present LongEmotion, a benchmark that encompasses a diverse suite of tasks targeting the assessment of models’ capabilities in Emotion Recognition, Knowledge Application, and Empathetic Generation, with an average context length of 15,341 tokens. To enhance performance under realistic constraints, we introduce the Collaborative Emotional Modeling (CoEM) framework, which integrates Retrieval-Augmented Generation (RAG) and multi-agent collaboration to improve models’ EI in long-context scenarios. We conduct a detailed analysis of various models in long-context settings, investigating how reasoning mode activation, RAG-based retrieval strategies, and context-length adaptability influence their EI performance. Our project page is: https://longemotion.github.io/

Method: SPECTRA treats LLMs as implicit probabilistic models, probing them to infer probability distributions over interpretable preference clusters rather than directly generating ranked lists. This shifts from sequence generation with decoding heuristics to distributional inference for explicit cluster-level user preference representations.

Result: SPECTRA achieves: 1) 25% reduction in Jensen-Shannon divergence to empirical distributions; 2) 30% increase in global exposure entropy by reducing head concentration; 3) 40% NDCG boost on public datasets and 7x improvement on ranking long-tail preferences against production baselines.

Conclusion: Distributional inference over interpretable preference clusters provides a more effective approach than direct generation for LLM-based user modeling, addressing bias issues and improving both head and long-tail preference representation.

Abstract: Large Language Models (LLMs) are increasingly used to understand user preferences, typically via the direct generation of ranked item lists. However, this end-to-end generative paradigm inherits the bias and opacity of autoregressive decoding, over-emphasizing frequent (head) preferences and obscure long-tail ones, thereby biasing personalization toward head preferences. To address this, we propose SPECTRA (Semantic Preference Extraction and Clustered TRAcking), which treats the LLM as an implicit probabilistic model by probing it to infer a probability distribution over interpretable preference clusters. In doing so, SPECTRA reframes user modeling from sequence generation with decoding heuristics to distributional inference, yielding explicit, cluster-level user preference representations. We evaluate SPECTRA on MovieLens, Yelp, and a large-scale short-video platform, demonstrating significant gains across three dimensions: SPECTRA achieves (i) distributional alignment, reducing Jensen-Shannon divergence to empirical distributions by 25% against strong baselines; (ii) long-tail exposure, reducing decoding-induced head concentration and increasing global exposure entropy by 30%; and (iii) downstream applications such as personalized ranking, translating these gains into a 40% NDCG boost on public datasets and a 7x improvement on ranking long-tail preferences against an industry-leading Transformer-based production baseline.

Method: Uses a permissive-first, risk-mitigated sourcing strategy combining public-domain and permissively licensed text with carefully justified low-risk additions. Implements a three-tier risk categorization scheme with shard-level provenance metadata. Employs a single-stage pretraining recipe integrating synthetic instruction and reasoning data.

Result: Models trained on MixtureVitae consistently outperform other permissive datasets across standard benchmarks. At 1.7B parameters/300B tokens, they approach DCLM performance and match/exceed instruction-tuned baselines on GSM8K, HumanEval, and MBPP despite using 36× fewer tokens.

Conclusion: Permissive-first data with high instruction/reasoning density, tiered by licensing risk, provides a practical and risk-mitigated foundation for training capable LLMs without sacrificing competitiveness or relying heavily on broad web scrapes.

Abstract: We present MixtureVitae, an open-access pretraining corpus built to minimize legal risk while providing strong downstream performance. MixtureVitae follows a permissive-first, risk-mitigated sourcing strategy that combines public-domain and permissively licensed text (e.g., CC-BY/Apache) with carefully justified low-risk additions (e.g., government works and EU TDM-eligible sources). MixtureVitae adopts a simple, single-stage pretraining recipe that integrates a large proportion of permissive synthetic instruction and reasoning data-signals typically introduced during post-training and generally scarce in permissive web corpora. We categorize all sources into a three-tier scheme that reflects varying risk levels and provide shard-level provenance metadata to enable risk-aware usage. In controlled experiments using the open-sci-ref training protocol (fixed architectures and hyperparameters; 50B and 300B token budgets across 130M-1.7B parameters), models trained on MixtureVitae consistently outperform other permissive datasets across a suite of standard benchmarks, and at the 1.7B-parameters/300B-tokens setting, they surpass FineWeb-Edu and approach DCLM late in training. Performance is particularly strong on MMLU and on math and code benchmarks: a 1.7B model pretrained on 300B MixtureVitae tokens matches or exceeds a strong 1.7B instruction-tuned baseline on GSM8K, HumanEval, and MBPP, despite using over 36 times fewer tokens (300B vs. ~11T). Supported by a thorough decontamination analysis, these results show that permissive-first data with high instruction and reasoning density, tiered by licensing and provenance-related risk, can provide a practical and risk-mitigated foundation for training capable LLMs, reducing reliance on broad web scrapes without sacrificing competitiveness. Code: https://github.com/ontocord/mixturevitae

Method: ThinkBrake monitors the log-probability margin between the top continuation token and the token at sentence boundaries, stopping reasoning when this margin narrows. No training required.

Result: ThinkBrake reduces thinking token usage by up to 30% while maintaining or improving accuracy across math, scientific QA, and tool usage benchmarks. Oracle stopping shows potential for 8% accuracy improvement with 72% token reduction.

Conclusion: ThinkBrake provides an effective, training-free solution to overthinking in Chain-of-Thought reasoning, with theoretical grounding showing equivalence to test-time realignment with reward bonuses.

Abstract: Large Reasoning Models (LRMs) allocate substantial inference-time compute to Chain-of-Thought (CoT) reasoning, improving performance on mathematics, scientific QA, and tool usage. However, this introduces overthinking: LRMs often reach a correct intermediate solution, continue reasoning, and overwrite it with an incorrect answer. We first demonstrate that oracle stopping–where we inject at every sentence boundary and select the best stopping point in hindsight–improves average accuracy by 8% while reducing thinking tokens by 72%, exposing substantial overthinking. Motivated by this finding, we propose ThinkBrake, which monitors the log-probability margin between the top continuation token and at sentence boundaries, stopping reasoning when this margin narrows. ThinkBrake requires no training and achieves favorable accuracy-efficiency trade-offs across math, scientific QA, and tool usage benchmarks, reducing thinking token usage by up to 30%. Furthermore, we provide theoretical analysis showing that ThinkBrake is equivalent to test-time realignment with a reward bonus for the token.

Method: Evaluated 12 Reasoning Models across diverse MT benchmarks in three scenarios: direct translation, forced-reasoning extrapolation, and post-editing. Examined how domain-specific fine-tuning affects TTS effectiveness.

Result: TTS provides limited and inconsistent benefits for direct translation with general-purpose models, quickly plateauing. Domain-specific fine-tuning unlocks TTS effectiveness, leading to consistent improvements up to optimal reasoning depth. Forcing models beyond natural stopping points degrades quality. TTS is highly effective in post-editing, reliably turning self-correction into beneficial process.

Conclusion: The value of inference-time computation in MT lies not in enhancing single-pass translation with general models, but in targeted applications like multi-step self-correction workflows and with task-specialized models.

Abstract: Test-time scaling (TTS) has enhanced the performance of Reasoning Models (RMs) on various tasks such as math and coding, yet its efficacy in machine translation (MT) remains underexplored. This paper investigates whether increased inference-time computation improves translation quality. We evaluate 12 RMs across a diverse suite of MT benchmarks spanning multiple domains, examining three scenarios: direct translation, forced-reasoning extrapolation, and post-editing. Our findings show that for general-purpose RMs, TTS provides limited and inconsistent benefits for direct translation, with performance quickly plateauing. However, the effectiveness of TTS is unlocked by domain-specific fine-tuning, which aligns a model’s reasoning process with task requirements, leading to consistent improvements up to an optimal, self-determined reasoning depth. We also find that forcing a model to reason beyond its natural stopping point consistently degrades translation quality. In contrast, TTS proves highly effective in a post-editing context, reliably turning self-correction into a beneficial process. These results indicate that the value of inference-time computation in MT lies not in enhancing single-pass translation with general models, but in targeted applications like multi-step, self-correction workflows and in conjunction with task-specialized models.

Method: Proposes a method that transforms LLM judgments directly into a bag-of-texts representation where texts are initialized as equidistant, without assuming prior distance relationships or requiring embedding optimization. The approach is model-agnostic and works with various embedders and clustering methods.

Result: Achieves comparable or superior results to state-of-the-art methods without embedding optimization or prior knowledge of clusters/labels. Works with diverse datasets, smaller LLMs, different clustering methods, and scales to large datasets with reduced computational cost.

Conclusion: The method offers flexibility and scalability that better aligns with real-world training-free clustering scenarios than existing approaches, making it practical for unsupervised intent clustering in customer-facing chatbots.

Abstract: In this paper, we propose a training-free method for unsupervised short text clustering that relies less on careful selection of embedders than other methods. In customer-facing chatbots, companies are dealing with large amounts of user utterances that need to be clustered according to their intent. In these settings, no labeled data is typically available, and the number of clusters is not known. Recent approaches to short-text clustering in label-free settings incorporate LLM output to refine existing embeddings. While LLMs can identify similar texts effectively, the resulting similarities may not be directly represented by distances in the dense vector space, as they depend on the original embedding. We therefore propose a method for transforming LLM judgments directly into a bag-of-texts representation in which texts are initialized to be equidistant, without assuming any prior distance relationships. Our method achieves comparable or superior results to state-of-the-art methods, but without embeddings optimization or assuming prior knowledge of clusters or labels. Experiments on diverse datasets and smaller LLMs show that our method is model agnostic and can be applied to any embedder, with relatively small LLMs, and different clustering methods. We also show how our method scales to large datasets, reducing the computational cost of the LLM use. The flexibility and scalability of our method make it more aligned with real-world training-free scenarios than existing clustering methods.

Method: CORGI uses synthetic databases inspired by real enterprises (DoorDash, Airbnb, Lululemon) and provides questions across four increasingly complex categories: descriptive, explanatory, predictive, and recommendational. It introduces new evaluation methods for open-ended qualitative responses in data access tasks.

Result: LLM performance degrades on higher-level questions as complexity increases. LLMs exhibit an average 33.12% lower success execution rate (SER) on CORGI compared to existing benchmarks like BIRD, highlighting the substantially higher complexity of real-world business needs.

Conclusion: CORGI expands text-to-SQL to reflect practical database queries and calls for multi-level, multi-step agentic intelligence. The benchmark encourages the community to consider new automatic evaluation methods for open-ended qualitative responses and supports future research with released dataset, framework, and submission website.

Abstract: Text-to-SQL benchmarks have traditionally only tested simple data access as a translation task of natural language to SQL queries. But in reality, users tend to ask diverse questions that require more complex responses including data-driven predictions or recommendations. Using the business domain as a motivating example, we introduce CORGI, a new benchmark that expands text-to-SQL to reflect practical database queries encountered by end users. CORGI is composed of synthetic databases inspired by enterprises such as DoorDash, Airbnb, and Lululemon. It provides questions across four increasingly complicated categories of business queries: descriptive, explanatory, predictive, and recommendational. This challenge calls for causal reasoning, temporal forecasting, and strategic recommendation, reflecting multi-level and multi-step agentic intelligence. We find that LLM performance degrades on higher-level questions as question complexity increases. CORGI also introduces and encourages the text-to-SQL community to consider new automatic methods for evaluating open-ended, qualitative responses in data access tasks. Our experiments show that LLMs exhibit an average 33.12% lower success execution rate (SER) on CORGI compared to existing benchmarks such as BIRD, highlighting the substantially higher complexity of real-world business needs. We release the CORGI dataset, an evaluation framework, and a submission website to support future research.

Method: 1) Create OpenRubrics - diverse, large-scale collection of (prompt, rubric) pairs. 2) Introduce Contrastive Rubric Generation (CRG) that derives both hard rules (explicit constraints) and principles (implicit qualities) by contrasting preferred and rejected responses. 3) Remove noisy rubrics via preference-label consistency preservation.

Result: Rubric-based reward model (Rubric-RM) surpasses strong size-matched baselines by 8.4% across multiple reward-modeling benchmarks. These gains transfer to policy models on instruction-following and biomedical benchmarks.

Conclusion: OpenRubrics provides a scalable solution for generating reliable rubrics, and the contrastive rubric generation approach enables capturing multifaceted human preferences more effectively than traditional reward modeling methods.

Abstract: Reward modeling lies at the core of reinforcement learning from human feedback (RLHF), yet most existing reward models rely on scalar or pairwise judgments that fail to capture the multifaceted nature of human preferences. Recent studies have explored rubrics-as-rewards (RaR) that uses structured criteria to capture multiple dimensions of response quality. However, producing rubrics that are both reliable and scalable remains a key challenge. In this work, we introduce OpenRubrics, a diverse, large-scale collection of (prompt, rubric) pairs for training rubric-generation and rubric-based reward models. To elicit discriminative and comprehensive evaluation signals, we introduce Contrastive Rubric Generation (CRG), which derives both hard rules (explicit constraints) and principles (implicit qualities) by contrasting preferred and rejected responses. We further remove noisy rubrics via preserving preference-label consistency. Across multiple reward-modeling benchmarks, our rubric-based reward model, Rubric-RM, surpasses strong size-matched baselines by 8.4%. These gains transfer to policy models on instruction-following and biomedical benchmarks.

Method: Introduced SLAQ (Short-Long Form Alignment for Factual Question Answering), a controlled evaluation framework comparing LLMs’ answers to the same factual questions asked in isolation (short) vs. integrated into complex queries (long). Evaluated 16 LLMs across 600 queries, conducted mechanistic analysis to examine model internals.

Result: Found systematic misalignment between short and long query answers, position-dependent accuracy loss, and momentum effects where consecutive correct/incorrect answers create self-reinforcing patterns. Mechanistic analysis showed aligned facts activate overlapping model internals, and mechanistic similarity metrics can predict short-long answer alignment with up to 78% accuracy.

Conclusion: Factual consistency over query complexity is crucial for LLM trustworthiness. Current evaluation practices are insufficient as they assume simple query performance implies complex task reliability. The work establishes this consistency gap as an important aspect of model trustworthiness.

Abstract: Large language models (LLMs) can correctly answer “When was Einstein born?” yet fail to provide the same date when writing about Einstein’s life revealing a fundamental inconsistency in how models access factual knowledge across task complexities. While models display impressive accuracy on factual question-answering benchmarks, the reliability gap between simple and complex queries remains poorly understood, eroding their trustworthiness. In this work, we introduce Short-Long Form Alignment for Factual Question Answering (SLAQ), a controlled evaluation framework that compares LLMs’ answers to the same factual questions asked (a) in isolation (short) vs. (b) integrated into complex queries (long). Looking at 16 LLMs across 600 queries, we find a systematic misalignment of answers to the corresponding short and long queries. We further uncover position-dependent accuracy loss and momentum effects where consecutive correct or incorrect answers create self-reinforcing patterns. Through mechanistic analysis, we find that aligned facts activate overlapping model internals, and that metrics based on mechanistic similarity can predict short-long answer alignment with up to 78% accuracy. Our work establishes factual consistency over query complexity as an important aspect of LLMs’ trustworthiness and challenges current evaluation practices, which implicitly assume that good performance for simple factual queries implies reliability in more complex knowledge-seeking tasks too.

Method: Unsupervised framework that: 1) generates aspect-specific summarizing sentences, 2) trains embedding models to map related summaries to nearby positions, 3) distills aspect-specific capabilities into a unified model that predicts multiple aspect embeddings in one forward pass, and 4) includes an embedding decoding pipeline that converts embeddings back to natural language descriptions.

Result: The approach produces multifaceted embeddings that enable fine-grained similarity assessment and adaptive visualizations. The decoding pipeline remains effective even for unoccupied regions in low-dimensional visualizations, improving interpretability in user-centric settings.

Conclusion: SemCSE-Multi provides a novel framework for generating interpretable, multifaceted embeddings of scientific abstracts that support fine-grained analysis and user-driven exploration, with applications demonstrated in invasion biology and medicine.

Abstract: We propose SemCSE-Multi, a novel unsupervised framework for generating multifaceted embeddings of scientific abstracts, evaluated in the domains of invasion biology and medicine. These embeddings capture distinct, individually specifiable aspects in isolation, thus enabling fine-grained and controllable similarity assessments as well as adaptive, user-driven visualizations of scientific domains. Our approach relies on an unsupervised procedure that produces aspect-specific summarizing sentences and trains embedding models to map semantically related summaries to nearby positions in the embedding space. We then distill these aspect-specific embedding capabilities into a unified embedding model that directly predicts multiple aspect embeddings from a scientific abstract in a single, efficient forward pass. In addition, we introduce an embedding decoding pipeline that decodes embeddings back into natural language descriptions of their associated aspects. Notably, we show that this decoding remains effective even for unoccupied regions in low-dimensional visualizations, thus offering vastly improved interpretability in user-centric settings.

Method: Minimal Test-Time Intervention (MTI) includes two key components: (1) Selective CFG intervention that applies classifier-free guidance only at uncertain positions rather than all tokens, and (2) Lightweight negative-prompt guidance that reuses the main model’s KV cache to approximate unconditional decoding efficiently.

Result: MTI achieves consistent gains across general, coding, and STEM tasks: +9.28% average improvement on six benchmarks for DeepSeek-R1-7B and +11.25% on AIME2024 using Ling-mini-2.0, while maintaining high efficiency.

Conclusion: The paper demonstrates that targeted, minimal interventions at uncertain token positions can significantly improve LLM reasoning accuracy and stability without the computational overhead of full test-time scaling approaches, offering an efficient alternative to current methods.

Abstract: Recent progress in large language models (LLMs) has focused on test-time scaling to improve reasoning via increased inference computation, but often at the cost of efficiency. We revisit test-time behavior and uncover a simple yet underexplored phenomenon: reasoning uncertainty is highly localized-only a small subset of high-entropy tokens dominantly affects output correctness. Motivated by this, we propose Minimal Test-Time Intervention (MTI), a training-free framework that enhances reasoning accuracy and stability with minimal overhead. MTI includes: (i) Selective CFG intervention, applying classifier-free guidance only at uncertain positions; and (ii) Lightweight negative-prompt guidance, reusing the main model’s KV cache to approximate unconditional decoding efficiently. MTI yields consistent gains across general, coding, and STEM tasks-e.g., +9.28% average improvement on six benchmarks for DeepSeek-R1-7B and +11.25% on AIME2024 using Ling-mini-2.0-while remaining highly efficient.

Method: Proposes a label-free self-supervised RL framework that eliminates external supervision by deriving reward signals directly from instructions and generating pseudo-labels for reward model training. Uses constraint decomposition strategies and efficient constraint-wise binary classification to address sparse reward challenges while maintaining computational efficiency.

Result: The approach generalizes well, achieving strong improvements across 3 in-domain and 5 out-of-domain datasets, including challenging agentic and multi-turn instruction following.

Conclusion: The proposed self-supervised RL framework effectively addresses multi-constraint instruction following without external supervision, demonstrating strong generalization across diverse datasets and task types.

Abstract: Language models often struggle to follow multi-constraint instructions that are crucial for real-world applications. Existing reinforcement learning (RL) approaches suffer from dependency on external supervision and sparse reward signals from multi-constraint tasks. We propose a label-free self-supervised RL framework that eliminates dependency on external supervision by deriving reward signals directly from instructions and generating pseudo-labels for reward model training. Our approach introduces constraint decomposition strategies and efficient constraint-wise binary classification to address sparse reward challenges while maintaining computational efficiency. Experiments show that our approach generalizes well, achieving strong improvements across 3 in-domain and 5 out-of-domain datasets, including challenging agentic and multi-turn instruction following. The data and code are publicly available at https://github.com/Rainier-rq/verl-if

Method: GlobalRAG decomposes questions into subgoals, coordinates retrieval with reasoning, and refines evidence iteratively. It introduces Planning Quality Reward and SubGoal Completion Reward to encourage coherent planning and reliable execution, plus a progressive weight annealing strategy to balance process-oriented and outcome-based objectives.

Result: Extensive experiments on in-domain and out-of-domain benchmarks show GlobalRAG significantly outperforms strong baselines while using only 8k training data (42% of baselines’ data), achieving average improvements of 14.2% in both EM and F1 scores.

Conclusion: GlobalRAG effectively addresses the core limitations of RL in multi-hop QA by introducing structured global planning and faithful execution mechanisms, demonstrating superior performance with significantly less training data.

Abstract: Reinforcement learning has recently shown promise in improving retrieval-augmented generation (RAG). Despite these advances, its effectiveness in multi-hop question answering (QA) remains limited by two fundamental limitations: (i) global planning absence to structure multi-step reasoning, and (ii) unfaithful execution, which hinders effective query formulation and consistent use of retrieved evidence. We propose GlobalRAG, a reinforcement learning framework designed to enhance global reasoning in multi-hop QA. GlobalRAG decomposes questions into subgoals, coordinates retrieval with reasoning, and refines evidence iteratively. To guide this process, we introduce Planning Quality Reward and SubGoal Completion Reward, which encourage coherent planning and reliable subgoal execution. In addition, a progressive weight annealing strategy balances process-oriented and outcome-based objectives. Extensive experiments on both in-domain and out-of-domain benchmarks demonstrate that GlobalRAG significantly outperforms strong baselines while using only 8k training data (42% of the training data used by strong baselines), achieving average improvements of 14.2% in both EM and F1.

Method: 1) OlaBench: A comprehensive ICS benchmark spanning retrieval-augmented generation, workflow-based systems, and agentic settings that evaluates service capability, safety, and latency sensitivity. 2) OlaMind: Distills reusable reasoning patterns and service strategies from expert dialogues and applies rubric-aware staged exploration-exploitation reinforcement learning to improve model capability.

Result: OlaMind surpasses GPT-5.2 and Gemini 3 Pro on OlaBench (78.72 vs. 70.58/70.84). In online A/B tests, it delivers +23.67% average issue resolution and -6.6% human transfer rate versus baseline, effectively bridging offline gains to deployment.

Conclusion: Together, OlaBench and OlaMind advance ICS systems toward more anthropomorphic, professional, and reliable deployment by closing the gap between offline evaluation and real-world performance.

Abstract: Existing benchmarks and training pipelines for industrial intelligent customer service (ICS) remain misaligned with real-world dialogue requirements, overemphasizing verifiable task success while under-measuring subjective service quality and realistic failure modes, leaving a gap between offline gains and deployable dialogue behavior. We close this gap with a benchmark-to-optimization loop: we first introduce OlaBench, an ICS benchmark spanning retrieval-augmented generation, workflow-based systems, and agentic settings, which evaluates service capability, safety, and latency sensitivity; moreover, motivated by OlaBench results showing state-of-the-art LLMs still fall short, we propose OlaMind, which distills reusable reasoning patterns and service strategies from expert dialogues and applies rubric-aware staged exploration–exploitation reinforcement learning to improve model capability. OlaMind surpasses GPT-5.2 and Gemini 3 Pro on OlaBench (78.72 vs. 70.58/70.84) and, in online A/B tests, delivers an average +23.67% issue resolution and -6.6% human transfer rate versus the baseline, bridging offline gains to deployment. Together, OlaBench and OlaMind advance ICS systems toward more anthropomorphic, professional, and reliable deployment.

Method: Created M4FC dataset with 4,982 images verified by professional fact-checkers from 22 organizations, paired with 6,980 claims in up to 10 languages. Designed to cover six multimodal fact-checking tasks and analyzed task combinations.

Result: The dataset provides diverse cultural/geographic contexts and spans six tasks: visual claim extraction, claimant intent prediction, fake image detection, image contextualization, location verification, and verdict prediction. Baseline results provided for all tasks.

Conclusion: M4FC addresses key limitations of existing datasets and enables comprehensive multimodal fact-checking research across multiple languages and tasks. The dataset and code are publicly available to advance the field.

Abstract: Existing real-world datasets for multimodal fact-checking have multiple limitations: they contain few instances, focus on only one or two languages and tasks, suffer from evidence leakage, or rely on external sets of news articles for sourcing true claims. To address these shortcomings, we introduce M4FC, a new real-world dataset comprising 4,982 images paired with 6,980 claims. The images, verified by professional fact-checkers from 22 organizations, represent a diverse range of cultural and geographic contexts. Each claim is available in one or two out of ten languages. M4FC spans six multimodal fact-checking tasks: visual claim extraction, claimant intent prediction, fake image detection, image contextualization, location verification, and verdict prediction. We provide baseline results for all tasks and analyze how combining intermediate tasks influences verdict prediction performance. We make our dataset and code available.

Method: Created TicToc dataset with 76 scenarios across dynamic environments with varying time sensitivity. Collected human preferences on tool-calling decisions, evaluated LLM alignment with human preferences under different elapsed times, and tested prompt-based and post-training alignment techniques.

Result: Existing models show poor alignment with human temporal perception (no model better than 65% alignment even with time stamps). Prompt-based alignment has limited effectiveness, but specific post-training alignment can improve alignment with human temporal perception.

Conclusion: Temporal blindness is a critical limitation in LLM agents. The TicToc dataset provides a foundation for understanding and mitigating this issue, with post-training alignment showing promise for developing more time-aware and human-aligned agents.

Abstract: Large language model (LLM) agents are increasingly used to interact with and execute tasks in dynamic environments. However, a critical yet overlooked limitation of these agents is that they, by default, assume a stationary context, failing to account for the real-world time elapsed between messages. We refer to this as “temporal blindness”. This limitation hinders decisions about when to invoke tools, leading agents to either over-rely on stale context and skip needed tool calls, or under-rely on it and redundantly repeat tool calls. To study this challenge, we constructed TicToc, a diverse dataset of multi-turn user-agent message trajectories across 76 scenarios, spanning dynamic environments with high, medium, and low time sensitivity. We collected human preferences between “calling a tool” and “directly answering” on each sample, and evaluated how well LLM tool-calling decisions align with human preferences under varying amounts of elapsed time. Our analysis reveals that existing models display poor alignment with human temporal perception, with no model achieving a normalized alignment rate better than 65% when given time stamp information. We also show that naive, prompt-based alignment techniques have limited effectiveness for most models, but specific post-training alignment can be a viable way to align multi-turn LLM tool use with human temporal perception. Our data and findings provide a first step toward understanding and mitigating temporal blindness, offering insights to foster the development of more time-aware and human-aligned agents.

Method: Conducted extensive experiments to analyze entropy dynamics in RLVR-trained LLMs, examining correlations with response diversity, calibration, and performance. Identified three key entropy factors: clipping thresholds, off-policy updates, and training data diversity. Proposed Positive-Advantage Reweighting approach that adjusts loss weights for tokens with positive advantages.

Result: Found that tokens with positive advantages are primary drivers of entropy collapse. The proposed Positive-Advantage Reweighting effectively regulates model entropy while maintaining competitive performance across benchmarks.

Conclusion: The study provides comprehensive understanding of entropy dynamics in RLVR, identifies key influencing factors, and offers a simple yet effective solution (Positive-Advantage Reweighting) to mitigate entropy collapse while preserving model performance.

Abstract: Reinforcement learning with verifiable rewards (RLVR) has emerged as a prominent paradigm for enhancing the reasoning capabilities of large language models (LLMs). However, the entropy of LLMs usually collapses during RLVR training, leading to premature convergence to suboptimal local minima and hindering further performance improvement. Although various approaches have been proposed to mitigate entropy collapse, a comprehensive study of entropy in RLVR remains lacking. To bridge this gap, we conduct extensive experiments to investigate the entropy dynamics of LLMs trained with RLVR and analyze how model entropy correlates with response diversity, calibration, and performance across various benchmarks. Our results identify three key factors that influence entropy: the clipping thresholds in the optimization objective, the number of off-policy updates, and the diversity of the training data. Furthermore, through both theoretical analysis and empirical validation, we demonstrate that tokens with positive advantages are the primary drivers of entropy collapse. Motivated by this insight, we propose Positive-Advantage Reweighting, a simple yet effective approach that regulates model entropy by adjusting the loss weights assigned to tokens with positive advantages during RLVR training, while maintaining competitive performance.

Method: Two approaches: (1) For promptable encoder-decoder ASR (like Whisper), use decoder prompting with first-pass hypotheses and encoder/decoder prefixing with retrieved speech-text exemplars, plus prompt reordering and optional speaker-matched synthetic exemplars. (2) For CTC ASR, introduce proxy-guided n-best selection that chooses from model’s n-best list by minimizing text-level distance to external proxy hypotheses.

Result: Across ten Arabic conditions, context-aware decoding yields average relative WER reductions: 22.29% on MSA, 20.54% on accented MSA, and 9.15% on dialectal Arabic. For CTC models, proxy-guided selection reduces WER by 15.6% relative on MSA and recovers substantial fraction of oracle n-best gains.

Conclusion: Context-aware inference effectively adapts ASR models to diverse Arabic speech without parameter updates, generalizing beyond encoder-decoder architectures to CTC models through proxy-guided selection.

Abstract: Zero-shot ASR for Arabic remains challenging: while multilingual models perform well on Modern Standard Arabic (MSA), error rates rise sharply on dialectal and accented speech due to linguistic mismatch and scarce labeled data. We study context-aware decoding as a lightweight test-time adaptation paradigm that conditions inference on external side information without parameter updates. For promptable encoder-decoder ASR (e.g., Whisper), we incorporate context through (i) decoder prompting with first-pass hypotheses and (ii) encoder/decoder prefixing with retrieved speech-text exemplars, complemented by simple prompt reordering and optional speaker-matched synthetic exemplars to improve robustness in informal and multi-speaker settings. To extend contextual adaptation beyond promptable architectures, we introduce proxy-guided n-best selection for CTC ASR: given one or more external proxy hypotheses, we select from a model’s n-best list by minimizing text-level distance to the proxies, enabling contextual inference without direct prompting. Across ten Arabic conditions spanning MSA, accented MSA, and multiple dialects, context-aware decoding yields average relative WER reductions of 22.29% on MSA, 20.54 on accented MSA, and 9.15% on dialectal Arabic. For CTC models, proxy-guided selection reduces WER by 15.6% relative on MSA and recovers a substantial fraction of oracle n-best gains, demonstrating that context-aware inference generalizes beyond encoder-decoder ASR.

Method: Created a human-curated benchmark with over 13,000 multiple-choice questions covering 11 Indic languages (Nepali, Gujarati, Marathi, Odia as low-resource; Dogri, Maithili, Rajasthani, Sanskrit, Bodo, Santali, Konkani as extremely low-resource) plus Sanskrit-English code-mixed set. Questions are labeled as knowledge-oriented or purely linguistic, and include diverse formats beyond conventional MCQs.

Result: Evaluated 20 LLMs (proprietary and open-weights) showing even top-performing Gemini-2.5 reaches only 58% average accuracy, followed by GPT-5 (45%) and DeepSeek-3.2 (43.1%). The benchmark reveals significant limitations in cross-lingual transfer for Indic languages.

Conclusion: IndicParam establishes a challenging benchmark for Indic languages, provides insights into limitations of cross-lingual transfer, and highlights the need for improved LLM performance on low-resource Indic languages. The dataset and evaluation scripts are publicly available.

Abstract: While large language models excel on high-resource multilingual tasks, low- and extremely low-resource Indic languages remain severely under-evaluated. We present IndicParam, a human-curated benchmark of over 13,000 multiple-choice questions covering 11 such languages (Nepali, Gujarati, Marathi, Odia as low-resource; Dogri, Maithili, Rajasthani, Sanskrit, Bodo, Santali, Konkani as extremely low-resource) plus Sanskrit-English code-mixed set. We evaluated 20 LLMs, both proprietary and open-weights, which reveals that even the top-performing \texttt{Gemini-2.5} reaches 58% average accuracy, followed by \texttt{GPT-5} (45) and \texttt{DeepSeek-3.2} (43.1). We additionally label each question as knowledge-oriented or purely linguistic to discriminate factual recall from grammatical proficiency. Further, we assess the ability of LLMs to handle diverse question formats-such as list-based matching, assertion-reason pairs, and sequence ordering-alongside conventional multiple-choice questions. \benchmark\ provides insights into limitations of cross-lingual transfer and establishes a challenging benchmark for Indic languages. The dataset is available at https://huggingface.co/datasets/bharatgenai/IndicParam. Scripts to run benchmark are present at https://github.com/ayushbits/IndicParam.

Method: Pearmut provides an end-to-end platform with features like document-level context, absolute/contrastive evaluation, attention checks, ESAAI pre-annotations, and both static/active learning-based assignment strategies. It supports standard protocols (DA, ESA, MQM) while being extensible for new protocols.

Result: The platform enables reliable human evaluation to become a practical, routine component of model development rather than an occasional effort, making it as easy to run as automatic evaluation.

Conclusion: Pearmut addresses the critical gap in human evaluation tools by providing a lightweight, feature-rich platform that reduces engineering overhead and makes human assessment accessible for multilingual NLP tasks, particularly machine translation.

Abstract: Human evaluation is the gold standard for multilingual NLP, but is often skipped in practice and substituted with automatic metrics, because it is notoriously complex and slow to set up with existing tools with substantial engineering and operational overhead. We introduce Pearmut, a lightweight yet feature-rich platform that makes end-to-end human evaluation as easy to run as automatic evaluation. Pearmut removes common entry barriers and provides support for evaluating multilingual tasks, with a particular focus on machine translation. The platform implements standard evaluation protocols, including DA, ESA, or MQM, but is also extensible to allow prototyping new protocols. It features document-level context, absolute and contrastive evaluation, attention checks, ESAAI pre-annotations and both static and active learning-based assignment strategies. Pearmut enables reliable human evaluation to become a practical, routine component of model development and diagnosis rather than an occasional effort.

Method: F-DPO extends DPO with: (1) label-flipping transformation to ensure chosen responses are never less factual than rejected ones, (2) factuality-aware margin emphasizing pairs with clear correctness differences, and (3) construction of factuality-aware preference data using binary factuality indicators and synthetic hallucinated variants.

Result: Across 7 LLMs (1B-14B), F-DPO consistently improves factuality and reduces hallucination rates vs base models and standard DPO. Qwen3-8B: hallucination rates reduced 5x (0.424→0.084), factuality scores improved 50% (5.26→7.90). On TruthfulQA, Qwen2.5-14B achieved +17% MC1 accuracy (0.500→0.585) and +49% MC2 accuracy (0.357→0.531).

Conclusion: F-DPO effectively improves LLM factuality without requiring auxiliary reward models, token-level annotations, or multi-stage training, offering a simple yet powerful extension to standard preference alignment methods.

Abstract: Preference alignment methods such as RLHF and Direct Preference Optimization (DPO) improve instruction following, but they can also reinforce hallucinations when preference judgments reward fluency and confidence over factual correctness. We introduce F-DPO (Factuality-aware Direct Preference Optimization), a simple extension of DPO that uses only binary factuality labels. F-DPO (i) applies a label-flipping transformation that corrects misordered preference pairs so the chosen response is never less factual than the rejected one, and (ii) adds a factuality-aware margin that emphasizes pairs with clear correctness differences, while reducing to standard DPO when both responses share the same factuality. We construct factuality-aware preference data by augmenting DPO pairs with binary factuality indicators and synthetic hallucinated variants. Across seven open-weight LLMs (1B-14B), F-DPO consistently improves factuality and reduces hallucination rates relative to both base models and standard DPO. On Qwen3-8B, F-DPO reduces hallucination rates by five times (from 0.424 to 0.084) while improving factuality scores by 50 percent (from 5.26 to 7.90). F-DPO also generalizes to out-of-distribution benchmarks: on TruthfulQA, Qwen2.5-14B achieves plus 17 percent MC1 accuracy (0.500 to 0.585) and plus 49 percent MC2 accuracy (0.357 to 0.531). F-DPO requires no auxiliary reward model, token-level annotations, or multi-stage training.

Method: Augment curated parallel datasets with synthetic sentence pairs generated using high-capacity multilingual translation models; fine-tune mBART models on both curated-only and synthetically augmented data; apply language-specific preprocessing including orthographic normalization and noise-aware filtering.

Result: Experiments show consistent chrF++ improvements for Guarani-Spanish and Quechua-Spanish translation with synthetic data augmentation; diagnostic experiments on Aymara reveal limitations of generic preprocessing for highly agglutinative languages.

Conclusion: Synthetic data generation is effective for improving NMT in low-resource indigenous languages, but language-specific approaches are needed, especially for highly agglutinative languages where generic preprocessing has limitations.

Abstract: Low-resource indigenous languages often lack the parallel corpora required for effective neural machine translation (NMT). Synthetic data generation offers a practical strategy for mitigating this limitation in data-scarce settings. In this work, we augment curated parallel datasets for indigenous languages of the Americas with synthetic sentence pairs generated using a high-capacity multilingual translation model. We fine-tune a multilingual mBART model on curated-only and synthetically augmented data and evaluate translation quality using chrF++, the primary metric used in recent AmericasNLP shared tasks for agglutinative languages. We further apply language-specific preprocessing, including orthographic normalization and noise-aware filtering, to reduce corpus artifacts. Experiments on Guarani-Spanish and Quechua-Spanish translation show consistent chrF++ improvements from synthetic data augmentation, while diagnostic experiments on Aymara highlight the limitations of generic preprocessing for highly agglutinative languages.

Method: Researchers examine the causal relationship between metaphors in training data and LLM misalignment. They conduct interventions using metaphors during pre-training, fine-tuning, and re-alignment phases. They analyze the connection between metaphors and activation of global/local latent features in reasoning models, then design a detector based on monitoring these features.

Result: Strong causal relationship discovered between metaphors in training data and misalignment degree of LLMs’ reasoning. Interventions with metaphors significantly change models’ cross-domain misalignment degrees. Metaphors are connected to activation of latent features, enabling a detector that predicts misaligned content with high accuracy.

Conclusion: Metaphors play a significant role in LLM misalignment across domains, and monitoring latent feature activations related to metaphors enables effective detection of misaligned content in reasoning models.

Abstract: Earlier research has shown that metaphors influence human’s decision making, which raises the question of whether metaphors also influence large language models (LLMs)’ reasoning pathways, considering their training data contain a large number of metaphors. In this work, we investigate the problem in the scope of the emergent misalignment problem where LLMs can generalize patterns learned from misaligned content in one domain to another domain. We discover a strong causal relationship between metaphors in training data and the misalignment degree of LLMs’ reasoning contents. With interventions using metaphors in pre-training, fine-tuning and re-alignment phases, models’ cross-domain misalignment degrees change significantly. As we delve deeper into the causes behind this phenomenon, we observe that there is a connection between metaphors and the activation of global and local latent features of large reasoning models. By monitoring these latent features, we design a detector that predict misaligned content with high accuracy.

Method: 1) Create new multilingual dataset from English Wiktionary dump (16 languages, 75 translation directions, ~10M records); 2) Build Wiktionary search tool from cleaned records (~3M); 3) Train translation agent with RL using novel reward design and adaptive rollout generation based on “translation difficulty.”

Result: Developed comprehensive dataset and search tool for neologism-aware MT, and proposed RL framework that improves translation quality by leveraging translation difficulty metrics and specialized reward mechanisms.

Conclusion: The NeoAMT framework addresses the gap in neologism-aware MT by providing both data resources (dataset and search tool) and methodological innovations (RL with adaptive training), enabling better handling of novel words in translation tasks.

Abstract: Neologism-aware machine translation aims to translate source sentences containing neologisms into target languages. This field remains underexplored compared with general machine translation (MT). In this paper, we propose an agentic framework, NeoAMT, for neologism-aware machine translation using a Wiktionary search tool. Specifically, we first create a new dataset for neologism-aware machine translation and develop a search tool based on Wiktionary. The new dataset covers 16 languages and 75 translation directions and is derived from approximately 10 million records of an English Wiktionary dump. The retrieval corpus of the search tool is also constructed from around 3 million cleaned records of the Wiktionary dump. We then use it for training the translation agent with reinforcement learning (RL) and evaluating the accuracy of neologism-aware machine translation. Based on this, we also propose an RL training framework that contains a novel reward design and an adaptive rollout generation approach by leveraging “translation difficulty” to further improve the translation quality of translation agents using our search tool.

Method: ContextFocus uses activation steering - a lightweight approach that modifies model activations during inference to prioritize external context over internal knowledge. It requires no model fine-tuning and adds minimal inference-time overhead.

Result: Extensive experiments on the ConFiQA benchmark show ContextFocus significantly improves contextual-faithfulness compared to baselines like ContextDPO, COIECD, and prompting methods. It’s complementary to prompting strategies and remains effective on larger models.

Conclusion: ContextFocus is an effective, robust, and efficient solution for improving LLM faithfulness to retrieved context in knowledge-conflict scenarios, offering a practical approach for real-world deployment where external evidence must be prioritized.

Abstract: Large Language Models (LLMs) encode vast amounts of parametric knowledge during pre-training. As world knowledge evolves, effective deployment increasingly depends on their ability to faithfully follow externally retrieved context. When such evidence conflicts with the model’s internal knowledge, LLMs often default to memorized facts, producing unfaithful outputs. In this work, we introduce ContextFocus, a lightweight activation steering approach that improves context faithfulness in such knowledge-conflict settings while preserving fluency and efficiency. Unlike prior approaches, our solution requires no model finetuning and incurs minimal inference-time overhead, making it highly efficient. We evaluate ContextFocus on the ConFiQA benchmark, comparing it against strong baselines including ContextDPO, COIECD, and prompting-based methods. Furthermore, we show that our method is complementary to prompting strategies and remains effective on larger models. Extensive experiments show that ContextFocus significantly improves contextual-faithfulness. Our results highlight the effectiveness, robustness, and efficiency of ContextFocus in improving contextual-faithfulness of LLM outputs.

Method: Two-stage multimodal framework: 1) fine-tuned reranker for evidence retrieval from paper bodies, 2) fine-tuned model to predict overstatement scores with justification. Uses dataset of 10K+ claim-evidence sets from ICLR/NeurIPS papers, annotated by eight LLMs and calibrated using peer-review comments.

Result: RIGOURATE enables improved evidence retrieval and overstatement detection compared to strong baselines, with the framework validated through human evaluation.

Conclusion: The work operationalizes evidential proportionality and supports clearer, more transparent scientific communication by providing a systematic way to assess claim overstatement.

Abstract: Scientific rigour tends to be sidelined in favour of bold statements, leading authors to overstate claims beyond what their results support. We present RIGOURATE, a two-stage multimodal framework that retrieves supporting evidence from a paper’s body and assigns each claim an overstatement score. The framework consists of a dataset of over 10K claim-evidence sets from ICLR and NeurIPS papers, annotated using eight LLMs, with overstatement scores calibrated using peer-review comments and validated through human evaluation. It employes a fine-tuned reranker for evidence retrieval and a fine-tuned model to predict overstatement scores with justification. Compared to strong baselines, RIGOURATE enables improved evidence retrieval and overstatement detection. Overall, our work operationalises evidential proportionality and supports clearer, more transparent scientific communication.

Method: Constructs intra-chunk discourse trees to capture local hierarchies and builds inter-chunk rhetorical graphs to model cross-passage coherence, jointly integrated into a planning blueprint that conditions generation.

Result: Achieves state-of-the-art results on question answering and long-document summarization benchmarks without fine-tuning.

Conclusion: Discourse structure plays an important role in advancing RAG systems, and explicitly injecting discourse signals into generation significantly enhances performance on knowledge-intensive tasks.

Abstract: Retrieval-Augmented Generation (RAG) has emerged as an important means of enhancing the performance of large language models (LLMs) in knowledge-intensive tasks. However, most existing RAG strategies treat retrieved passages in a flat and unstructured way, which prevents the model from capturing structural cues and constrains its ability to synthesize knowledge from dispersed evidence across documents. To overcome these limitations, we propose Disco-RAG, a discourse-aware framework that explicitly injects discourse signals into the generation process. Our method constructs intra-chunk discourse trees to capture local hierarchies and builds inter-chunk rhetorical graphs to model cross-passage coherence. These structures are jointly integrated into a planning blueprint that conditions the generation. Experiments on question answering and long-document summarization benchmarks show the efficacy of our approach. Disco-RAG achieves state-of-the-art results on the benchmarks without fine-tuning. These findings underscore the important role of discourse structure in advancing RAG systems.

Method: Attention head intervention - a paradigm shift from visualization to intervention, moving from observing correlations to causally validating mechanistic hypotheses through direct intervention on transformer attention heads.

Result: Head intervention studies revealed robust empirical findings while highlighting interpretation limitations. Recent work shows mechanistic understanding enables targeted control: suppressing toxic outputs and manipulating semantic content through selective attention head intervention.

Conclusion: Attention head intervention represents a key advancement in mechanistic interpretability, validating the practical utility of interpretability research for AI safety by enabling causal understanding and targeted control of transformer behavior.

Abstract: Neural networks are growing more capable on their own, but we do not understand their neural mechanisms. Understanding these mechanisms’ decision-making processes, or mechanistic interpretability, enables (1) accountability and control in high-stakes domains, (2) the study of digital brains and the emergence of cognition, and (3) discovery of new knowledge when AI systems outperform humans. This paper traces how attention head intervention emerged as a key method for causal interpretability of transformers. The evolution from visualization to intervention represents a paradigm shift from observing correlations to causally validating mechanistic hypotheses through direct intervention. Head intervention studies revealed robust empirical findings while also highlighting limitations that complicate interpretation. Recent work demonstrates that mechanistic understanding now enables targeted control of model behaviour, successfully suppressing toxic outputs and manipulating semantic content through selective attention head intervention, validating the practical utility of interpretability research for AI safety.

Method: SmartSplat uses image-aware features (gradients and color variances) with Gradient-Color Guided Variational Sampling and Exclusion-based Uniform Sampling to improve non-overlapping coverage of Gaussian primitives. It also employs Scale-Adaptive Gaussian Color Sampling for better color initialization across scales, with joint optimization of spatial layout, scale, and color initialization.

Result: Extensive experiments on DIV8K and a new 16K dataset show that SmartSplat outperforms state-of-the-art methods at comparable compression ratios and exceeds their compression limits, demonstrating strong scalability and practical applicability.

Conclusion: SmartSplat provides an effective solution for ultra-high-resolution image compression using adaptive Gaussian Splatting techniques, achieving high reconstruction quality with limited Gaussians under strong compression, making it suitable for real-time decoding on end-user devices.

Abstract: Recent advances in generative AI have accelerated the production of ultra-high-resolution visual content, posing significant challenges for efficient compression and real-time decoding on end-user devices. Inspired by 3D Gaussian Splatting, recent 2D Gaussian image models improve representation efficiency, yet existing methods struggle to balance compression ratio and reconstruction fidelity in ultra-high-resolution scenarios. To address this issue, we propose SmartSplat, a highly adaptive and feature-aware GS-based image compression framework that supports arbitrary image resolutions and compression ratios. SmartSplat leverages image-aware features such as gradients and color variances, introducing a Gradient-Color Guided Variational Sampling strategy together with an Exclusion-based Uniform Sampling scheme to improve the non-overlapping coverage of Gaussian primitives in pixel space. In addition, we propose a Scale-Adaptive Gaussian Color Sampling method to enhance color initialization across scales. Through joint optimization of spatial layout, scale, and color initialization, SmartSplat efficiently captures both local structures and global textures using a limited number of Gaussians, achieving high reconstruction quality under strong compression. Extensive experiments on DIV8K and a newly constructed 16K dataset demonstrate that SmartSplat consistently outperforms state-of-the-art methods at comparable compression ratios and exceeds their compression limits, showing strong scalability and practical applicability. The code is publicly available at https://github.com/lif314/SmartSplat.

Method: Two-step pipeline: 1) Face recognition using pruned VGG-16 classifier trained on augmented dataset (924 images, 5 subjects) with MTCNN face localization; 2) Voice recognition using CNN speaker-verification model trained on LibriSpeech dataset, only performed against the matched identity from step 1.

Result: Face recognition achieves 95.1% accuracy; voice recognition achieves 98.9% accuracy and 3.456% EER on test-clean. The sequential approach reduces computation and improves robustness.

Conclusion: The proposed two-step authentication system effectively combines face and voice biometrics using common hardware, demonstrating high accuracy and computational efficiency for practical deployment.

Abstract: We present a cost-effective two-step authentication system that integrates face identification and speaker verification using only a camera and microphone available on common devices. The pipeline first performs face recognition to identify a candidate user from a small enrolled group, then performs voice recognition only against the matched identity to reduce computation and improve robustness. For face recognition, a pruned VGG-16 based classifier is trained on an augmented dataset of 924 images from five subjects, with faces localized by MTCNN; it achieves 95.1% accuracy. For voice recognition, a CNN speaker-verification model trained on LibriSpeech (train-other-360) attains 98.9% accuracy and 3.456% EER on test-clean. Source code and trained models are available at https://github.com/NCUE-EE-AIAL/Two-step-Authentication-Multi-biometric-System.

Method: HyperTopo-Adapters: Parameter-efficient head trained on frozen encoder, embedding features on hyperbolic+Euclidean+spherical product manifold. Uses topology prior with persistent homology distance for evaluation and differentiable surrogate (soft Euler-characteristic + total variation) for training. Includes warm-ups, per-sample structure-aware metrics, and min-PD within top-K Dice checkpoint selection.

Result: On Kaggle leaf-lesion dataset (N=2,940): consistent gains in boundary and topology metrics (reducing Delta beta_1 hole error by 9%) while maintaining competitive Dice/IoU. Diagnostic ablations show effects of curvature learning, latent dimensions, contrastive temperature, and surrogate settings.

Conclusion: Provides open, reproducible train/eval suite isolating geometric/topological priors, surfaces failure modes to guide stronger topology-preserving architectures. Contribution is diagnostic framework for topology-sensitive segmentation.

Abstract: Leaf-lesion segmentation is topology-sensitive: small merges, splits, or false holes can be biologically meaningful descriptors of biochemical pathways, yet they are weakly penalized by standard pixel-wise losses in Euclidean latents. I explore HyperTopo-Adapters, a lightweight, parameter-efficient head trained on top of a frozen vision encoder, which embeds features on a product manifold – hyperbolic + Euclidean + spherical (H + E + S) – to encourage hierarchical separation (H), local linear detail (E), and global closure (S). A topology prior complements Dice/BCE in two forms: (i) persistent-homology (PH) distance for evaluation and selection, and (ii) a differentiable surrogate that combines a soft Euler-characteristic match with total variation regularization for stable training. I introduce warm-ups for both the hyperbolic contrastive term and the topology prior, per-sample evaluation of structure-aware metrics (Boundary-F1, Betti errors, PD distance), and a min-PD within top-K Dice rule for checkpoint selection. On a Kaggle leaf-lesion dataset (N=2,940), early results show consistent gains in boundary and topology metrics (reducing Delta beta_1 hole error by 9%) while Dice/IoU remain competitive. The study is diagnostic by design: I report controlled ablations (curvature learning, latent dimensions, contrastive temperature, surrogate settings), and ongoing tests varying encoder strength (ResNet-50, DeepLabV3, DINOv2/v3), input resolution, PH weight, and partial unfreezing of late blocks. The contribution is an open, reproducible train/eval suite (available at https://github.com/ChimdiWalter/HyperTopo-Adapters) that isolates geometric/topological priors and surfaces failure modes to guide stronger, topology-preserving architectures.

Method: OptFormer uses an encoder-decoder architecture that combines phase-space reconstruction with a motion-aware attention mechanism guided by optical flow. Unlike conventional attention, it leverages inter-frame motion cues to highlight relative changes in spatial fields, focusing on dynamic regions and capturing long-range temporal dependencies.

Result: Experiments on NOAA SST datasets across multiple spatial scales show OptFormer achieves superior performance under 1:1 training-to-prediction setting, significantly outperforming existing baselines in both accuracy and robustness.

Conclusion: OptFormer effectively addresses SST prediction challenges by incorporating motion-aware attention with optical flow guidance, demonstrating improved performance for capturing complex spatiotemporal dynamics in climate data.

Abstract: Sea Surface Temperature (SST) prediction plays a vital role in climate modeling and disaster forecasting. However, it remains challenging due to its nonlinear spatiotemporal dynamics and extended prediction horizons. To address this, we propose OptFormer, a novel encoder-decoder model that integrates phase-space reconstruction with a motion-aware attention mechanism guided by optical flow. Unlike conventional attention, our approach leverages inter-frame motion cues to highlight relative changes in the spatial field, allowing the model to focus on dynamic regions and capture long-range temporal dependencies more effectively. Experiments on NOAA SST datasets across multiple spatial scales demonstrate that OptFormer achieves superior performance under a 1:1 training-to-prediction setting, significantly outperforming existing baselines in accuracy and robustness.

Method: Detects/tracks objects with YOLOv11, converts proximity patterns into human-object events, organizes into Temporal Scene Graph, then uses query-aware pruning to extract relevant subgraph for verbalization and answer generation with Gemini 2.5 Flash.

Result: Achieves 65.0% accuracy using only 3.47k tokens per query (vs 62.5% at 40.39k tokens for full-log baseline), reducing token usage by 91.4%. Short-context baseline collapses to 2.5% accuracy.

Conclusion: Symbolic temporal graphs serve as effective plug-and-play memory layer for off-the-shelf VLMs, preserving long-range reasoning while making long-form video QA substantially more token- and cost-efficient.

Abstract: Long-form video question answering remains challenging for modern vision-language models, which struggle to reason over hour-scale footage without exceeding practical token and compute budgets. Existing systems typically downsample frames or feed dense visual embeddings to large-context language models, trading off temporal coverage against cost. We propose Semantic Event Graphs (SEG), a lightweight symbolic interface between video and language that replaces raw frames with compact temporal interaction logs. Our pipeline detects and tracks objects with YOLOv11, converts proximity patterns into START/END human-object events, and organizes them into a Temporal Scene Graph (TSG). At inference time, a query-aware pruning module identifies anchor entities and lexically relevant events, returning only a small subgraph which is verbalized and passed to Gemini 2.5 Flash for answer generation. On five YouTube videos (300-500 interactions each) and 120 automatically generated long-horizon questions, SEG achieves 65.0% accuracy using only 3.47k tokens per query, closely matching a full-log baseline (62.5% at 40.39k tokens) while reducing token usage by 91.4%. A short-context baseline restricted to the last 30 seconds collapses to 2.5% accuracy, underscoring the need for explicit temporal memory. These results show that symbolic temporal graphs can serve as an effective, plug-and-play memory layer for off-the-shelf vision-language models, preserving long-range reasoning ability while making long-form video question answering substantially more token- and cost-efficient. Code, logs, and event-extraction tools will be released for reproducibility.

Method: COVR introduces a collaborative framework with bidirectional optimization: 1) Fine-tunes VLM with RL-generated data to improve semantic reasoning for target tasks, 2) Uses enhanced VLM to guide RL policy via action priors. Includes Exploration-Driven Dynamic Filter (preserves valuable exploration samples) and Return-Aware Adaptive Loss Weight (improves training stability). Also uses progressive fine-tuning to reduce resource consumption.

Result: Extensive experiments show COVR achieves strong performance across various challenging visual control tasks, demonstrating effectiveness of the collaborative optimization approach.

Conclusion: The proposed COVR framework successfully enables mutual enhancement between VLMs and RL policies through bidirectional knowledge transfer, addressing the limitations of previous one-way approaches and improving overall performance in visual RL tasks.

Abstract: Visual reinforcement learning (RL) suffers from poor sample efficiency due to high-dimensional observations in complex tasks. While existing works have shown that vision-language models (VLMs) can assist RL, they often focus on knowledge distillation from the VLM to RL, overlooking the potential of RL-generated interaction data to enhance the VLM. To address this, we propose COVR, a collaborative optimization framework that enables the mutual enhancement of the VLM and RL policies. Specifically, COVR fine-tunes the VLM with RL-generated data to enhance the semantic reasoning ability consistent with the target task, and uses the enhanced VLM to further guide policy learning via action priors. To improve fine-tuning efficiency, we introduce two key modules: (1) an Exploration-Driven Dynamic Filter module that preserves valuable exploration samples using adaptive thresholds based on the degree of exploration, and (2) a Return-Aware Adaptive Loss Weight module that improves the stability of training by quantifying the inconsistency of sampling actions via return signals of RL. We further design a progressive fine-tuning strategy to reduce resource consumption. Extensive experiments show that COVR achieves strong performance across various challenging visual control tasks.

Method: Creation of the Keraal dataset - a clinically collected dataset of patients performing low back-pain rehabilitation exercises, benchmarked on state-of-the-art human movement analysis algorithms.

Result: The dataset includes rehabilitation motions in clinical settings with patients in their rehabilitation programs, addressing four key challenges in exercise monitoring: motion assessment, error recognition, spatial localization, and temporal localization.

Conclusion: The Keraal dataset provides valuable clinical data for advancing intelligent tutoring systems in rehabilitation by addressing critical monitoring challenges with real patient data.

Abstract: To allow the development and assessment of physical rehabilitation by an intelligent tutoring system, we propose a medical dataset of clinical patients carrying out low back-pain rehabilitation exercises and benchmark on state of the art human movement analysis algorithms. This dataset is valuable because it includes rehabilitation motions in a clinical setting with patients in their rehabilitation program. This paper introduces the Keraal dataset, a clinically collected dataset to enable intelligent tutoring systems (ITS) for rehabilitation. It addresses four challenges in exercise monitoring: motion assessment, error recognition, spatial localization, temporal localization

Method: FIA uses Contrastive Concept Saliency to quantify weight contributions, identifies Concept-Sensitive Neurons using temporal and spatial information, creates concept masks, and fuses them into a unified multi-concept mask that preserves concept-agnostic neurons while pruning concept-specific ones.

Result: Extensive experiments across three unlearning tasks show FIA achieves more reliable multi-concept unlearning with improved forgetting effectiveness while maintaining semantic fidelity and image quality.

Conclusion: FIA provides a training-free, plug-and-play solution for multi-concept unlearning that requires minimal hyperparameter tuning, addressing real-world needs for selective content removal in diffusion models.

Abstract: The widespread adoption of text-to-image (T2I) diffusion models has raised concerns about their potential to generate copyrighted, inappropriate, or sensitive imagery learned from massive training corpora. As a practical solution, machine unlearning aims to selectively erase unwanted concepts from a pre-trained model without retraining from scratch. While most existing methods are effective for single-concept unlearning, they often struggle in real-world scenarios that require removing multiple concepts, since extending them to this setting is both non-trivial and problematic, causing significant challenges in unlearning effectiveness, generation quality, and sensitivity to hyperparameters and datasets. In this paper, we take a unique perspective on multi-concept unlearning by leveraging model sparsity and propose the Forget It All (FIA) framework. FIA first introduces Contrastive Concept Saliency to quantify each weight connection’s contribution to a target concept. It then identifies Concept-Sensitive Neurons by combining temporal and spatial information, ensuring that only neurons consistently responsive to the target concept are selected. Finally, FIA constructs masks from the identified neurons and fuses them into a unified multi-concept mask, where Concept-Agnostic Neurons that broadly support general content generation are preserved while concept-specific neurons are pruned to remove the targets. FIA is training-free and requires only minimal hyperparameter tuning for new tasks, thereby promoting a plug-and-play paradigm. Extensive experiments across three distinct unlearning tasks demonstrate that FIA achieves more reliable multi-concept unlearning, improving forgetting effectiveness while maintaining semantic fidelity and image quality.

Method: Created HAERAE-Vision benchmark with 653 real-world visual questions from Korean online communities (0.76% survival from 86K candidates), each paired with an explicit rewrite, totaling 1,306 query variants. Evaluated 39 VLMs including state-of-the-art models.

Result: Even top models (GPT-5, Gemini 2.5 Pro) achieve under 50% accuracy on original underspecified queries. Query explicitation alone yields 8-22 point improvements, with smaller models benefiting most. Web search cannot compensate for underspecification.

Conclusion: Substantial VLM difficulty stems from natural query underspecification rather than model capability, highlighting critical gap between benchmark evaluation and real-world deployment.

Abstract: Current vision-language benchmarks predominantly feature well-structured questions with clear, explicit prompts. However, real user queries are often informal and underspecified. Users naturally leave much unsaid, relying on images to convey context. We introduce HAERAE-Vision, a benchmark of 653 real-world visual questions from Korean online communities (0.76% survival from 86K candidates), each paired with an explicit rewrite, yielding 1,306 query variants in total. Evaluating 39 VLMs, we find that even state-of-the-art models (GPT-5, Gemini 2.5 Pro) achieve under 50% on the original queries. Crucially, query explicitation alone yields 8 to 22 point improvements, with smaller models benefiting most. We further show that even with web search, under-specified queries underperform explicit queries without search, revealing that current retrieval cannot compensate for what users leave unsaid. Our findings demonstrate that a substantial portion of VLM difficulty stem from natural query under-specification instead of model capability, highlighting a critical gap between benchmark evaluation and real-world deployment.

Method: Proposes Hierarchical Watermark Learning (HiWL) framework with two-stage optimization: 1) Distribution alignment learning establishes common latent space with visual consistency and information invariance constraints; 2) Generalized watermark representation learning separates unique watermark representations from marked images in RGB space.

Result: Achieves 7.6% higher accuracy in watermark extraction compared to existing methods while maintaining extremely low latency (processing 1000 images in 1 second). Demonstrates effectiveness across extensive experiments.

Conclusion: HiWL successfully addresses the three essential criteria for generalizable watermarking, providing an effective solution for copyright protection of image assets with superior performance and practical efficiency.

Abstract: Deep image watermarking, which refers to enabling imperceptible watermark embedding and reliable extraction in cover images, has been shown to be effective for copyright protection of image assets. However, existing methods face limitations in simultaneously satisfying three essential criteria for generalizable watermarking: (1) invisibility (imperceptible hiding of watermarks), (2) robustness (reliable watermark recovery under diverse conditions), and (3) broad applicability (low latency in the watermarking process). To address these limitations, we propose a Hierarchical Watermark Learning (HiWL) framework, a two-stage optimization that enables a watermarking model to simultaneously achieve all three criteria. In the first stage, distribution alignment learning is designed to establish a common latent space with two constraints: (1) visual consistency between watermarked and non-watermarked images, and (2) information invariance across watermark latent representations. In this way, multimodal inputs – including watermark messages (binary codes) and cover images (RGB pixels) – can be effectively represented, ensuring both the invisibility of watermarks and robustness in the watermarking process. In the second stage, we employ generalized watermark representation learning to separate a unique representation of the watermark from the marked image in RGB space. Once trained, the HiWL model effectively learns generalizable watermark representations while maintaining broad applicability. Extensive experiments demonstrate the effectiveness of the proposed method. Specifically, it achieves 7.6% higher accuracy in watermark extraction compared to existing methods, while maintaining extremely low latency (processing 1000 images in 1 second).

Method: B-FIRE uses a CNN-INR (Implicit Neural Representation) encoder-decoder backbone optimized with diffusion models. It employs a comprehensive loss function enforcing both image-domain fidelity and frequency-aware constraints. The framework is trained on motion-binned image pairs but performs inference on binning-free undersampled data.

Result: Experiments on T1-weighted StarVIBE liver MRI with accelerations from RV8 to RV1 show B-FIRE outperforms direct NuFFT, GRASP-CS, and unrolled CNN methods in reconstruction fidelity and motion trajectory consistency while maintaining reasonable inference latency.

Conclusion: B-FIRE represents a novel paradigm for hyper-accelerated 4DMRI reconstruction that successfully captures instantaneous 3D abdominal anatomy without the blurring artifacts of traditional motion-binned approaches, enabling more accurate dynamic motion representation.

Abstract: Accelerated dynamic volumetric magnetic resonance imaging (4DMRI) is essential for applications relying on motion resolution. Existing 4DMRI produces acceptable artifacts of averaged breathing phases, which can blur and misrepresent instantaneous dynamic information. Recovery of such information requires a new paradigm to reconstruct extremely undersampled non-Cartesian k-space data. We propose B-FIRE, a binning-free diffusion implicit neural representation framework for hyper-accelerated MR reconstruction capable of reflecting instantaneous 3D abdominal anatomy. B-FIRE employs a CNN-INR encoder-decoder backbone optimized using diffusion with a comprehensive loss that enforces image-domain fidelity and frequency-aware constraints. Motion binned image pairs were used as training references, while inference was performed on binning-free undersampled data. Experiments were conducted on a T1-weighted StarVIBE liver MRI cohort, with accelerations ranging from 8 spokes per frame (RV8) to RV1. B-FIRE was compared against direct NuFFT, GRASP-CS, and an unrolled CNN method. Reconstruction fidelity, motion trajectory consistency, and inference latency were evaluated.

Method: Three complementary dimensionality reduction techniques: Principal Component Analysis (PCA), Factor Analysis (FA), and Uniform Manifold Approximation and Projection (UMAP). These methods are used to study intrinsic dimensionality, shared variation, and nonlinear geometry of handwritten digits.

Result: PCA reveals dominant global variance directions and enables high-fidelity reconstructions with few components. FA decomposes digits into interpretable latent handwriting primitives (strokes, loops, symmetry). UMAP uncovers nonlinear manifolds showing smooth stylistic transitions between digit classes.

Conclusion: Handwritten digits occupy a structured low-dimensional manifold, and different statistical frameworks expose complementary aspects of this structure, with each method providing unique insights into digit organization.

Abstract: Handwritten digit images lie in a high-dimensional pixel space but exhibit strong geometric and statistical structure. This paper investigates the latent organization of handwritten digits in the MNIST dataset using three complementary dimensionality reduction techniques: Principal Component Analysis (PCA), Factor Analysis (FA), and Uniform Manifold Approximation and Projection (UMAP). Rather than focusing on classification accuracy, we study how each method characterizes intrinsic dimensionality, shared variation, and nonlinear geometry. PCA reveals dominant global variance directions and enables high-fidelity reconstructions using a small number of components. FA decomposes digits into interpretable latent handwriting primitives corresponding to strokes, loops, and symmetry. UMAP uncovers nonlinear manifolds that reflect smooth stylistic transitions between digit classes. Together, these results demonstrate that handwritten digits occupy a structured low-dimensional manifold and that different statistical frameworks expose complementary aspects of this structure.

Method: TBDN integrates two complementary training-free mechanisms: 1) Hint Instruction (HI) - lightweight prompt engineering to inject task-aware inductive bias and anchor models on contextual mapping rules; 2) Query Contrastive Decoding (QCD) - adjusts language model decoding distributions by contrasting full-input and query-omitted distributions to suppress prior-dominated hallucination.

Result: Achieves State-of-the-Art performance on CoBSAT and Text-to-Image Fast Mini-ImageNet benchmarks. Shows robust generalization across model backbones, prompt designs, and hyperparameters. Maintains promising performance in concept preservation and prompt following on Dreambench++.

Conclusion: TBDN establishes a simple yet effective training-free framework for efficient and reliable T2I-ICL by breaking the two bottlenecks of compliance failure and prior-dominated hallucination, offering flexibility and reduced deployment costs compared to training-based approaches.

Abstract: Text-to-Image In-Context Learning (T2I-ICL) enables customized image synthesis via interleaved text-image examples but faces two mutually reinforcing bottlenecks, compliance failure and prior-dominated hallucination, that form a vicious cycle degrading generation quality. Existing methods rely on tailored training, which limits flexibility and raises deployment costs. To address these challenges effectively, we propose TBDN, a training-free framework integrating two complementary closed-loop mechanisms: Hint Instruction (HI) and Query Contrastive Decoding (QCD). HI injects task-aware inductive bias via lightweight prompt engineering to anchor models on contextual mapping rules, thereby mitigating compliance failure. QCD adjusts the decoding distributions of language models by contrasting full-input and query-omitted distributions, suppressing prior-dominated hallucination. TBDN achieves State-of-the-Art performance on CoBSAT and Text-to-Image Fast Mini-ImageNet, with robust generalization across model backbones, prompt designs, and hyperparameters. It also maintains promising performance in concept preservation and prompt following on Dreambench++. By breaking the two bottlenecks, TBDN establishes a simple yet effective framework for efficient and reliable T2I-ICL.

Method: Proposes a hierarchical adaptive categorization framework using multi-criteria density-based clustering and adaptive quality-driven refinement. Categorizes Gaussians into two types: (1) Sketch Gaussians for high-frequency, boundary-defining features, and (2) Patch Gaussians for low-frequency, smooth regions. This enables layered progressive streaming where compact Sketch Gaussians establish structure first, followed by incremental refinement with Patch Gaussians.

Result: Achieves up to 1.74 dB improvement in PSNR, 6.7% in SSIM, and 41.4% in LPIPS at equivalent model sizes compared to uniform pruning baselines. For indoor scenes, maintains visual quality with only 0.5% of original model size. Works across diverse scenes including both man-made and natural environments.

Conclusion: The semantic separation of Gaussians into Sketch and Patch categories enables efficient storage, adaptive streaming, and high-fidelity rendering across bandwidth-constrained networks and resource-limited devices, outperforming traditional uniform pruning approaches while eliminating dependency on external 3D line primitives.

Abstract: We observe that Gaussians exhibit distinct roles and characteristics analogous to traditional artistic techniques – like how artists first sketch outlines before filling in broader areas with color, some Gaussians capture high-frequency features such as edges and contours, while others represent broader, smoother regions analogous to brush strokes that add volume and depth. Based on this observation, we propose a hybrid representation that categorizes Gaussians into (i) Sketch Gaussians, which represent high-frequency, boundary-defining features, and (ii) Patch Gaussians, which cover low-frequency, smooth regions. This semantic separation naturally enables layered progressive streaming, where the compact Sketch Gaussians establish the structural skeleton before Patch Gaussians incrementally refine volumetric detail. In this work, we extend our previous method to arbitrary 3D scenes by proposing a novel hierarchical adaptive categorization framework that operates directly on the 3DGS representation. Our approach employs multi-criteria density-based clustering, combined with adaptive quality-driven refinement. This method eliminates dependency on external 3D line primitives while ensuring optimal parametric encoding effectiveness. Our comprehensive evaluation across diverse scenes, including both man-made and natural environments, demonstrates that our method achieves up to 1.74 dB improvement in PSNR, 6.7% in SSIM, and 41.4% in LPIPS at equivalent model sizes compared to uniform pruning baselines. For indoor scenes, our method can maintain visual quality with only 0.5% of the original model size. This structure-aware representation enables efficient storage, adaptive streaming, and rendering of high-fidelity 3D content across bandwidth-constrained networks and resource-limited devices.

Method: Propose weighted activation distribution KANs (WKANs) to reduce activation frequency and distribute node information, plus multilevel tensor splitting framework (MTSF) to decompose high-dimensional data and enable tensor-parallel computation.

Result: SpectralKAN (using MTSF) achieves OA 0.9801 and Kappa 0.9514 on Farmland dataset with only 8k parameters, 0.07M FLOPs, 911MB memory, 13.26s training time, and 2.52s testing time across five datasets.

Conclusion: WKANs with MTSF framework effectively address KAN limitations on high-dimensional data, demonstrating superior accuracy-efficiency trade-off for hyperspectral image change detection tasks.

Abstract: Kolmogorov-Arnold networks (KANs) represent data features by learning the activation functions and demonstrate superior accuracy with fewer parameters, FLOPs, GPU memory usage (Memory), shorter training time (TraT), and testing time (TesT) when handling low-dimensional data. However, when applied to high-dimensional data, which contains significant redundant information, the current activation mechanism of KANs leads to unnecessary computations, thereby reducing computational efficiency. KANs require reshaping high-dimensional data into a one-dimensional tensor as input, which inevitably results in the loss of dimensional information. To address these limitations, we propose weighted activation distribution KANs (WKANs), which reduce the frequency of activations per node and distribute node information into different output nodes through weights to avoid extracting redundant information. Furthermore, we introduce a multilevel tensor splitting framework (MTSF), which decomposes high-dimensional data to extract features from each dimension independently and leverages tensor-parallel computation to significantly improve the computational efficiency of WKANs on high-dimensional data. In this paper, we design SpectralKAN for hyperspectral image change detection using the proposed MTSF. SpectralKAN demonstrates outstanding performance across five datasets, achieving an overall accuracy (OA) of 0.9801 and a Kappa coefficient (K) of 0.9514 on the Farmland dataset, with only 8 k parameters, 0.07 M FLOPs, 911 MB Memory, 13.26 S TraT, and 2.52 S TesT, underscoring its superior accuracy-efficiency trade-off. The source code is publicly available at https://github.com/yanhengwang-heu/SpectralKAN.

Method: TIR-Flow operates through three synergistic modules: 1) HDD (Hypothesis Decomposition and Deduction) decomposes complex queries into verifiable sub-tasks; 2) HAP (Hypothesis-Aware Perception) actively directs visual attention to gather high-resolution evidence for hypothesis validation; 3) EBA (Evidence-Based Accumulation) maintains a persistent workspace to accumulate and update discovered clues for logical reasoning. The framework works with frozen VLMs without additional data or parameter updates.

Result: Extensive experiments on seven benchmarks show TIR-Flow significantly outperforms recent strong baselines, delivering an average performance boost of 5.9%, with gains reaching 10.5% on Egoschema. The framework demonstrates that empowering frozen VLMs with System-2-like active perception is a scalable path toward solving long-horizon video reasoning.

Conclusion: TIR-Flow represents a paradigm shift from passive processing to active video searching and reasoning, enabling frozen Video-LLMs to achieve superior reasoning capabilities without additional training. The approach demonstrates that activating intrinsic intelligence through active perception is more effective than traditional data engineering methods for complex video reasoning tasks.

Abstract: While Large Video-Language Models (Video-LLMs) have achieved remarkable progress in perception, their reasoning capabilities remain a bottleneck. Existing solutions typically resort to a heavy “data engineering” paradigm-synthesizing large-scale Chain-of-Thought (CoT) datasets followed by Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL). This pipeline primarily optimizes probability sampling efficiency and aligns output distributions, but fails to activate the intrinsic intelligence required for dynamic visual exploration. In this work, we propose TIR-Flow, a novel framework that shifts the paradigm from passive processing to active video searching and reasoning without additional data or parameter updating. Concretely, our framework operates through three synergistic modules: HDD decomposes complex queries into a set of verifiable sub-tasks; HAP actively directs visual attention to gather high-resolution evidence for hypothesis validation; EBA maintains a persistent workspace to accumulate and update the discovered clues for logical reasoning. Extensive experiments on seven benchmarks demonstrate that TIR-Flow significantly outperforms recent strong baselines, delivering an average performance boost of 5.9%, with gains reaching 10.5% on Egoschema. Our analysis confirms that empowering frozen VLMs with System-2-like active perception is a scalable path toward solving long-horizon video reasoning.

Method: TLRN uses a temporal residual network with carefully designed residual blocks in latent deformation spaces parameterized by time-sequential initial velocity fields. The system treats residual blocks over time as a dynamic training system, where each block learns residual functions between desired deformation features and current input accumulated from previous time frames.

Result: TLRN achieves substantially improved registration accuracy compared to state-of-the-art methods on both synthetic data and real-world cine cardiac magnetic resonance (CMR) image videos.

Conclusion: The proposed TLRN effectively handles large motions in time-series image registration by leveraging temporal smoothness through a novel residual network architecture in latent deformation spaces, demonstrating superior performance on cardiac imaging applications.

Abstract: This paper presents a novel approach, termed {\em Temporal Latent Residual Network (TLRN)}, to predict a sequence of deformation fields in time-series image registration. The challenge of registering time-series images often lies in the occurrence of large motions, especially when images differ significantly from a reference (e.g., the start of a cardiac cycle compared to the peak stretching phase). To achieve accurate and robust registration results, we leverage the nature of motion continuity and exploit the temporal smoothness in consecutive image frames. Our proposed TLRN highlights a temporal residual network with residual blocks carefully designed in latent deformation spaces, which are parameterized by time-sequential initial velocity fields. We treat a sequence of residual blocks over time as a dynamic training system, where each block is designed to learn the residual function between desired deformation features and current input accumulated from previous time frames. We validate the effectivenss of TLRN on both synthetic data and real-world cine cardiac magnetic resonance (CMR) image videos. Our experimental results shows that TLRN is able to achieve substantially improved registration accuracy compared to the state-of-the-art. Our code is publicly available at https://github.com/nellie689/TLRN.

Method: Unified Attention U-Net architecture trained jointly on MRI (BraTS 2021) and CT (LIDC-IDRI) datasets with modality-harmonized preprocessing, attention-gated skip connections, and modality-aware Focal Tversky loss function.

Result: The unified model demonstrates competitive performance in terms of Dice coefficient, IoU, and AUC on both MRI and CT domains, showing generalizability across modalities.

Conclusion: A single Attention U-Net can effectively handle multiple imaging modalities without modality-specific encoders or domain adaptation, establishing a robust baseline for future cross-modality tumor segmentation research.

Abstract: This study presents a unified Attention U-Net architecture trained jointly on MRI (BraTS 2021) and CT (LIDC-IDRI) datasets to investigate the generalizability of a single model across diverse imaging modalities and anatomical sites. Our proposed pipeline incorporates modality-harmonized preprocessing, attention-gated skip connections, and a modality-aware Focal Tversky loss function. To the best of our knowledge, this study is among the first to evaluate a single Attention U-Net trained simultaneously on separate MRI (BraTS) and CT (LIDC-IDRI) tumor datasets, without relying on modality-specific encoders or domain adaptation. The unified model demonstrates competitive performance in terms of Dice coefficient, IoU, and AUC on both domains, thereby establishing a robust and reproducible baseline for future research in cross-modality tumor segmentation.

Method: A cascading multi-agent framework with early modules for reconstruction-gated filtering and object-level assessment, plus higher-level reasoning agents selectively invoked for ambiguous events. Uses adaptive escalation thresholds and publish-subscribe communication for asynchronous coordination across heterogeneous hardware.

Result: Achieves 3x latency reduction compared to direct vision-language inference while maintaining high perceptual fidelity (PSNR = 38.3 dB, SSIM = 0.965) and consistent semantic labeling. Provides efficient, interpretable anomaly detection.

Conclusion: The framework advances anomaly detection by combining early-exit efficiency, adaptive multi-agent reasoning, and explainable anomaly attribution, establishing a reproducible and energy-efficient foundation for scalable intelligent visual monitoring.

Abstract: Intelligent anomaly detection in dynamic visual environments requires reconciling real-time performance with semantic interpretability. Conventional approaches address only fragments of this challenge. Reconstruction-based models capture low-level deviations without contextual reasoning, object detectors provide speed but limited semantics, and large vision-language systems deliver interpretability at prohibitive computational cost. This work introduces a cascading multi-agent framework that unifies these complementary paradigms into a coherent and interpretable architecture. Early modules perform reconstruction-gated filtering and object-level assessment, while higher-level reasoning agents are selectively invoked to interpret semantically ambiguous events. The system employs adaptive escalation thresholds and a publish-subscribe communication backbone, enabling asynchronous coordination and scalable deployment across heterogeneous hardware. Extensive evaluation on large-scale monitoring data demonstrates that the proposed cascade achieves a threefold reduction in latency compared to direct vision-language inference, while maintaining high perceptual fidelity (PSNR = 38.3 dB, SSIM = 0.965) and consistent semantic labeling. The framework advances beyond conventional detection pipelines by combining early-exit efficiency, adaptive multi-agent reasoning, and explainable anomaly attribution, establishing a reproducible and energy-efficient foundation for scalable intelligent visual monitoring.

Method: The paper proposes: 1) Realistic Anomaly Augmentation to inject synthetic anomalies while preserving realistic spatial/occlusion patterns, creating VAA-KITTI and VAA-KITTI-360 datasets; 2) OccOoD framework integrating OoD detection into 3D semantic occupancy prediction using Cross-Space Semantic Refinement (CSSR) to refine predictions from complementary voxel and BEV representations.

Result: OccOoD achieves state-of-the-art OoD detection with AuROC of 65.50% and AuPRCr of 31.83% within 1.2m region, while maintaining competitive semantic occupancy prediction performance and generalization in real-world urban driving scenes.

Conclusion: The proposed OccOoD framework effectively addresses OoD detection in 3D semantic occupancy prediction, enhancing safety for autonomous driving systems. The datasets and source code will be publicly released to support further research.

Abstract: 3D semantic occupancy prediction is crucial for autonomous driving, providing a dense, semantically rich environmental representation. However, existing methods focus on in-distribution scenes, making them susceptible to Out-of-Distribution (OoD) objects and long-tail distributions, which increases the risk of undetected anomalies and misinterpretations, posing safety hazards. To address these challenges, we introduce Out-of-Distribution Semantic Occupancy Prediction, targeting OoD detection in 3D voxel space. To fill dataset gaps, we propose a Realistic Anomaly Augmentation that injects synthetic anomalies while preserving realistic spatial and occlusion patterns, enabling the creation of two datasets: VAA-KITTI and VAA-KITTI-360. Then, a novel framework that integrates OoD detection into 3D semantic occupancy prediction, OccOoD, is proposed, which uses Cross-Space Semantic Refinement (CSSR) to refine semantic predictions from complementary voxel and BEV representations, improving OoD detection. Experimental results demonstrate that OccOoD achieves state-of-the-art OoD detection with an AuROC of 65.50% and an AuPRCr of 31.83 within a 1.2m region, while maintaining competitive semantic occupancy prediction performance and generalization in real-world urban driving scenes. The established datasets and source code will be made publicly available at https://github.com/7uHeng/OccOoD.

Method: A multi-stage framework using vision-language models (VLMs) to segment videos into procedural units, align them with audio transcripts and recipe text, and a video comparison pipeline to identify procedural differences between regional variants. The approach employs multiple VLMs in complementary roles with human-in-the-loop verification.

Result: Created the first large-scale curated dataset of 120 biryani preparation videos across 12 regional styles, developed a comprehensive QA benchmark for evaluating procedural understanding in VLMs, and benchmarked several state-of-the-art models under zero-shot and fine-tuned settings.

Conclusion: The dataset, comparison methodology, and QA benchmark provide a new testbed for evaluating VLMs on structured, multimodal reasoning tasks and open new directions for computational analysis of cultural heritage through cooking videos.

Abstract: Biryani, one of India’s most celebrated dishes, exhibits remarkable regional diversity in its preparation, ingredients, and presentation. With the growing availability of online cooking videos, there is unprecedented potential to study such culinary variations using computational tools systematically. However, existing video understanding methods fail to capture the fine-grained, multimodal, and culturally grounded differences in procedural cooking videos. This work presents the first large-scale, curated dataset of biryani preparation videos, comprising 120 high-quality YouTube recordings across 12 distinct regional styles. We propose a multi-stage framework leveraging recent advances in vision-language models (VLMs) to segment videos into fine-grained procedural units and align them with audio transcripts and canonical recipe text. Building on these aligned representations, we introduce a video comparison pipeline that automatically identifies and explains procedural differences between regional variants. We construct a comprehensive question-answer (QA) benchmark spanning multiple reasoning levels to evaluate procedural understanding in VLMs. Our approach employs multiple VLMs in complementary roles, incorporates human-in-the-loop verification for high-precision tasks, and benchmarks several state-of-the-art models under zero-shot and fine-tuned settings. The resulting dataset, comparison methodology, and QA benchmark provide a new testbed for evaluating VLMs on structured, multimodal reasoning tasks and open new directions for computational analysis of cultural heritage through cooking videos. We release all data, code, and the project website at https://farzanashaju.github.io/how-does-india-cook-biryani/.

Method: Neighborhood Feature Pooling (NFP) layer that captures relationships between neighboring spatial features by aggregating local similarity patterns across feature dimensions. Implemented using standard convolutional operations, it can be seamlessly integrated into existing neural network architectures with minimal additional parameters.

Result: Extensive experiments across multiple benchmark datasets and backbone models demonstrate that NFP consistently improves classification performance compared to conventional pooling strategies while maintaining computational efficiency.

Conclusion: NFP effectively captures discriminative texture information in remote sensing imagery through neighborhood-based feature aggregation, offering a computationally efficient enhancement to existing neural network architectures.

Abstract: In this work, we introduce Neighborhood Feature Pooling (NFP), a novel pooling layer designed to enhance texture-aware representation learning for remote sensing image classification. The proposed NFP layer captures relationships between neighboring spatial features by aggregating local similarity patterns across feature dimensions. Implemented using standard convolutional operations, NFP can be seamlessly integrated into existing neural network architectures with minimal additional parameters. Extensive experiments across multiple benchmark datasets and backbone models demonstrate that NFP consistently improves classification performance compared to conventional pooling strategies, while maintaining computational efficiency. These results highlight the effectiveness of neighborhood-based feature aggregation for capturing discriminative texture information in remote sensing imagery.

Method: Proposed QwenStyle model trained on Qwen-Image-Edit with Curriculum Continual Learning framework. Used collected/filtered high-quality style data and synthesized triplets with thousands of style categories. Trained with mixture of clean and noisy triplets to generalize to unseen styles.

Result: QwenStyle V1 achieves state-of-the-art performance in three core metrics: style similarity, content consistency, and aesthetic quality. Successfully activates Qwen-Image-Edit’s content preservation and style customization capabilities.

Conclusion: The proposed method effectively addresses the content-style entanglement problem in DiTs through curriculum continual learning, enabling robust style transfer with strong content preservation and generalization to unseen styles.

Abstract: Content-Preserving Style transfer, given content and style references, remains challenging for Diffusion Transformers (DiTs) due to its internal entangled content and style features. In this technical report, we propose the first content-preserving style transfer model trained on Qwen-Image-Edit, which activates Qwen-Image-Edit’s strong content preservation and style customization capability. We collected and filtered high quality data of limited specific styles and synthesized triplets with thousands categories of style images in-the-wild. We introduce the Curriculum Continual Learning framework to train QwenStyle with such mixture of clean and noisy triplets, which enables QwenStyle to generalize to unseen styles without degradation of the precise content preservation capability. Our QwenStyle V1 achieves state-of-the-art performance in three core metrics: style similarity, content consistency, and aesthetic quality.

Method: Simulation study using two open-source datasets with controlled levels of class imbalance and label shift. Compared three Active Learning selection strategies: random sampling, entropy-based selection, and core-set selection.

Result: Active learning strategies (especially entropy-based and core-set selections) remain effective even for highly imbalanced datasets. However, strong label shift causes measurable efficiency loss.

Conclusion: Active Learning is valuable for imbalanced vision datasets, but practitioners should be aware of label shift’s negative impact on efficiency.

Abstract: The aim of Active Learning is to select the most informative samples from an unlabelled set of data. This is useful in cases where the amount of data is large and labelling is expensive, such as in machine vision or medical imaging. Two particularities of machine vision are first, that most of the images produced are free of defects, and second, that the amount of images produced is so big that we cannot store all acquired images. This results, on the one hand, in a strong class imbalance in defect distribution and, on the other hand, in a potential label shift caused by limited storage. To understand how these two forms of imbalance affect active learning algorithms, we propose a simulation study based on two open-source datasets. We artificially create datasets for which we control the levels of class imbalance and label shift. Three standard active learning selection strategies are compared: random sampling, entropy-based selection, and core-set selection. We demonstrate that active learning strategies, and in particular the entropy-based and core-set selections, remain interesting and efficient even for highly imbalanced datasets. We also illustrate and measure the loss of efficiency that occurs in the situation a strong label shift.

Method: Integrates H-SSD with VL-JEPA using Mamba-3 Selective State Space Model augmented by Sparse Mixture of Hamiltonian Experts (SMoE-HE) that enforces latent physical conservation laws. For visual synthesis, introduces Hamiltonian Flow Matching (HFM) and persistent 3D Gaussian Splatting (3DGS).

Result: Achieves state-of-the-art video prediction (FVD: 287), 4x faster visual synthesis than diffusion models, 3-18x inference speedup over transformer baselines, ultra-low latency (<50ms) on mobile hardware, and maintains energy conservation over extended horizons.

Conclusion: Incorporating physics-inspired inductive biases into neural architectures yields significant improvements in spatiotemporal coherence, computational efficiency, and energy conservation, establishing a new paradigm for latent world models.

Abstract: We present Akasha 2, a state-of-the-art multimodal architecture that integrates Hamiltonian State Space Duality (H-SSD) with Visual-Language Joint Embedding Predictive Architecture (VL-JEPA). The system leverages the Mamba-3 Selective State Space Model (SSM) augmented by a Sparse Mixture of Hamiltonian Experts (SMoE-HE) that enforces latent physical conservation laws through symplectic integration. For visual synthesis, we introduce Hamiltonian Flow Matching (HFM) and persistent 3D Gaussian Splatting (3DGS), enabling ultra-low latency (<50ms) on mobile hardware. This work establishes a new paradigm in latent world models, achieving unprecedented spatiotemporal coherence through a holographic memory architecture. Our approach demonstrates that incorporating physics-inspired inductive biases into neural architectures yields significant improvements: state-of-the-art video prediction (FVD: 287), 4x faster visual synthesis than diffusion models, and 3-18x inference speedup over transformer baselines while maintaining energy conservation over extended horizons.

Method: Proposes Semantic-Agnostic Prompt Learning (SAPL) in CLIP with two modules: Edge-aware Contextual Prompt Learning (ECPL) uses edge-enhanced image features to generate learnable textual prompts via attention, embedding semantic-irrelevant information to focus on manipulation edges. Hierarchical Edge Contrastive Learning (HECL) extracts genuine/manipulated edge patches and uses contrastive learning to boost discrimination between them.

Result: Extensive experiments on multiple public benchmarks demonstrate SAPL significantly outperforms existing approaches, achieving state-of-the-art localization performance.

Conclusion: SAPL effectively addresses the gap in weakly supervised manipulation localization by focusing on boundary cues rather than object semantics, providing a cost-effective solution without requiring pixel-level annotations.

Abstract: Malicious image manipulation threatens public safety and requires efficient localization methods. Existing approaches depend on costly pixel-level annotations which make training expensive. Existing weakly supervised methods rely only on image-level binary labels and focus on global classification, often overlooking local edge cues that are critical for precise localization. We observe that feature variations at manipulated boundaries are substantially larger than in interior regions. To address this gap, we propose Semantic-Agnostic Prompt Learning (SAPL) in CLIP, which learns text prompts that intentionally encode non-semantic, boundary-centric cues so that CLIPs multimodal similarity highlights manipulation edges rather than high-level object semantics. SAPL combines two complementary modules Edge-aware Contextual Prompt Learning (ECPL) and Hierarchical Edge Contrastive Learning (HECL) to exploit edge information in both textual and visual spaces. The proposed ECPL leverages edge-enhanced image features to generate learnable textual prompts via an attention mechanism, embedding semantic-irrelevant information into text features, to guide CLIP focusing on manipulation edges. The proposed HECL extract genuine and manipulated edge patches, and utilize contrastive learning to boost the discrimination between genuine edge patches and manipulated edge patches. Finally, we predict the manipulated regions from the similarity map after processing. Extensive experiments on multiple public benchmarks demonstrate that SAPL significantly outperforms existing approaches, achieving state-of-the-art localization performance.

Method: Proposes a three-module framework: 1) Enhanced visual localization with planning and captioning stages before reasoning, using quality-based caption rewards; 2) Improved exploration by categorizing samples based on reward distribution mean/variance, prioritizing high-variance samples; 3) Mitigation of sample interference via NTK similarity regulation using InfoNCE loss to balance gradient interactions.

Result: Experimental results show the proposed method significantly reduces hallucination rates and effectively enhances the inference accuracy of MLLMs.

Conclusion: The paper successfully identifies key factors causing hallucinations in MLLMs during RL training and provides an effective framework to address them, improving model reliability and practical deployment potential.

Abstract: While Multimodal Large Language Models (MLLMs) have achieved remarkable success across diverse tasks, their practical deployment is severely hindered by hallucination issues, which become particularly acute during Reinforcement Learning (RL) optimization. This paper systematically analyzes the root causes of hallucinations in MLLMs under RL training, identifying three critical factors: (1) an over-reliance on chained visual reasoning, where inaccurate initial descriptions or redundant information anchor subsequent inferences to incorrect premises; (2) insufficient exploration diversity during policy optimization, leading the model to generate overly confident but erroneous outputs; and (3) destructive conflicts between training samples, where Neural Tangent Kernel (NTK) similarity causes false associations and unstable parameter updates. To address these challenges, we propose a comprehensive framework comprising three core modules. First, we enhance visual localization by introducing dedicated planning and captioning stages before the reasoning phase, employing a quality-based caption reward to ensure accurate initial anchoring. Second, to improve exploration, we categorize samples based on the mean and variance of their reward distributions, prioritizing samples with high variance to focus the model on diverse and informative data. Finally, to mitigate sample interference, we regulate NTK similarity by grouping sample pairs and applying an InfoNCE loss to push overly similar pairs apart and pull dissimilar ones closer, thereby guiding gradient interactions toward a balanced range. Experimental results demonstrate that our proposed method significantly reduces hallucination rates and effectively enhances the inference accuracy of MLLMs.

Method: Proposes a conditional generative diffusion model for synthesizing FMCW radar Range-Azimuth Maps using Confidence Maps for semantic conditioning, with Geometry Aware Conditioning and Temporal Consistency Regularization for radar-specific characteristics.

Result: Improves signal reconstruction quality by 3.6 dB in PSNR over baselines, and training with real+synthetic data improves mean Average Precision by 4.15% compared to conventional augmentation methods.

Conclusion: The framework produces physically plausible and diverse radar spectrum while substantially improving model generalization in downstream perception tasks.

Abstract: The scarcity and low diversity of well-annotated automotive radar datasets often limit the performance of deep-learning-based environmental perception. To overcome these challenges, we propose a conditional generative framework for synthesizing realistic Frequency-Modulated Continuous-Wave radar Range-Azimuth Maps. Our approach leverages a generative diffusion model to generate radar data for multiple object categories, including pedestrians, cars, and cyclists. Specifically, conditioning is achieved via Confidence Maps, where each channel represents a semantic class and encodes Gaussian-distributed annotations at target locations. To address radar-specific characteristics, we incorporate Geometry Aware Conditioning and Temporal Consistency Regularization into the generative process. Experiments on the ROD2021 dataset demonstrate that signal reconstruction quality improves by \SI{3.6}{dB} in Peak Signal-to-Noise Ratio over baseline methods, while training with a combination of real and synthetic datasets improves overall mean Average Precision by 4.15% compared with conventional image-processing-based augmentation. These results indicate that our generative framework not only produces physically plausible and diverse radar spectrum but also substantially improves model generalization in downstream tasks.

Method: The paper conducts a systematic literature review of facial recognition methods, analyzing nine major algorithms: Hidden Markov Models, PCA, Elastic Cluster Plot Matching, SVM, Gabor Waves, ANN, Eigenfaces, ICA, and 3D Morphable Model. It also evaluates these methods using six standard facial databases (JAFEE, FEI, Yale, LFW, AT&T/ORL, and AR).

Result: The survey provides a comprehensive analysis of how different methods address key facial recognition challenges. It presents experimental results comparing algorithm performance across various databases, highlighting strengths and weaknesses of each approach in handling specific challenges like lighting variations, aging, pose differences, occlusion, and expressions.

Conclusion: The paper concludes that effective facial recognition requires addressing multiple challenging factors simultaneously. It provides a thorough review of existing methodologies and their applications, offering insights into current state-of-the-art approaches and identifying areas for future research in developing more robust facial recognition systems.

Abstract: As multimedia content is quickly growing, the field of facial recognition has become one of the major research fields, particularly in the recent years. The most problematic area to researchers in image processing and computer vision is the human face which is a complex object with myriads of distinctive features that can be used to identify the face. The survey of this survey is particularly focused on most challenging facial characteristics, including differences in the light, ageing, variation in poses, partial occlusion, and facial expression and presents methodological solutions. The factors, therefore, are inevitable in the creation of effective facial recognition mechanisms used on facial images. This paper reviews the most sophisticated methods of facial detection which are Hidden Markov Models, Principal Component Analysis (PCA), Elastic Cluster Plot Matching, Support Vector Machine (SVM), Gabor Waves, Artificial Neural Networks (ANN), Eigenfaces, Independent Component Analysis (ICA), and 3D Morphable Model. Alongside the works mentioned above, we have also analyzed the images of a number of facial databases, namely JAFEE, FEI, Yale, LFW, AT&T (then called ORL), and AR (created by Martinez and Benavente), to analyze the results. However, this survey is aimed at giving a thorough literature review of face recognition, and its applications, and some experimental results are provided at the end after a detailed discussion.

Method: Combines MediaPipe-based facial landmark extraction with a convolutional neural network inspired by iTracker architecture. Investigates two strategies: 1) adapting a model pretrained on mobile data, and 2) training from scratch on desktop-oriented data. Includes optional lightweight user-specific fine-tuning for calibration.

Result: Validation on MPIIFaceGaze shows comparable performance between both approaches before calibration, with user-specific fine-tuning consistently reducing gaze prediction error. Evaluation in Dot-Probe task shows strong agreement with commercial SeeSo SDK in left-right gaze allocation, though with higher temporal variability.

Conclusion: EyeTheia provides an effective, transparent, and extensible solution for low-cost gaze tracking suitable for scalable experimental and clinical studies, with code, models, and materials publicly available.

Abstract: We introduce EyeTheia, a lightweight and open deep learning pipeline for webcam-based gaze estimation, designed for browser-based experimental platforms and real-world cognitive and clinical research. EyeTheia enables real-time gaze tracking using only a standard laptop webcam, combining MediaPipe-based landmark extraction with a convolutional neural network inspired by iTracker and optional user-specific fine-tuning. We investigate two complementary strategies: adapting a model pretrained on mobile data and training the same architecture from scratch on a desktop-oriented dataset. Validation results on MPIIFaceGaze show comparable performance between both approaches prior to calibration, while lightweight user-specific fine-tuning consistently reduces gaze prediction error. We further evaluate EyeTheia in a realistic Dot-Probe task and compare it to the commercial webcam-based tracker SeeSo SDK. Results indicate strong agreement in left-right gaze allocation during stimulus presentation, despite higher temporal variability. Overall, EyeTheia provides a transparent and extensible solution for low-cost gaze tracking, suitable for scalable and reproducible experimental and clinical studies. The code, trained models, and experimental materials are publicly available.

Method: Proposes NAS-GS framework with: 1) Two-Ways Splatting technique for accurate modeling of dual directions in sonar imaging (intensity accumulation and transmittance calculation), improving rendering speed; 2) Gaussian Mixture Model (GMM) based noise model to capture complex sonar noise patterns (side-lobes, speckle, multi-path noise), preventing 3D Gaussian overfitting to noise.

Result: Achieves state-of-the-art performance on both simulated and real-world large-scale offshore sonar scenarios, with superior results in novel view synthesis and 3D reconstruction.

Conclusion: NAS-GS effectively addresses the unique challenges of sonar imaging through innovative splatting techniques and noise modeling, enabling high-quality 3D reconstruction and novel view synthesis for underwater applications.

Abstract: Underwater sonar imaging plays a crucial role in various applications, including autonomous navigation in murky water, marine archaeology, and environmental monitoring. However, the unique characteristics of sonar images, such as complex noise patterns and the lack of elevation information, pose significant challenges for 3D reconstruction and novel view synthesis. In this paper, we present NAS-GS, a novel Noise-Aware Sonar Gaussian Splatting framework specifically designed to address these challenges. Our approach introduces a Two-Ways Splatting technique that accurately models the dual directions for intensity accumulation and transmittance calculation inherent in sonar imaging, significantly improving rendering speed without sacrificing quality. Moreover, we propose a Gaussian Mixture Model (GMM) based noise model that captures complex sonar noise patterns, including side-lobes, speckle, and multi-path noise. This model enhances the realism of synthesized images while preventing 3D Gaussian overfitting to noise, thereby improving reconstruction accuracy. We demonstrate state-of-the-art performance on both simulated and real-world large-scale offshore sonar scenarios, achieving superior results in novel view synthesis and 3D reconstruction.

Method: Organized five consolidated tracks: unified video QA, unified object/point tracking, unified action/sound localization, grounded video QA, and hour-long video QA. Required unified approaches rather than task-specific engineered pipelines.

Result: The challenge highlighted significant difficulties existing models face when tackling diverse perception tasks through unified interfaces, with unified tracks merging previously separate tasks.

Conclusion: Perception Test 2025 demonstrates that current multimodal models struggle with unified task interfaces, indicating need for more versatile and integrated approaches to diverse perception challenges.

Abstract: The Third Perception Test challenge was organised as a full-day workshop alongside the IEEE/CVF International Conference on Computer Vision (ICCV) 2025. Its primary goal is to benchmark state-of-the-art video models and measure the progress in multimodal perception. This year, the workshop featured 2 guest tracks as well: KiVA (an image understanding challenge) and Physic-IQ (a video generation challenge). In this report, we summarise the results from the main Perception Test challenge, detailing both the existing tasks as well as novel additions to the benchmark. In this iteration, we placed an emphasis on task unification, as this poses a more challenging test for current SOTA multimodal models. The challenge included five consolidated tracks: unified video QA, unified object and point tracking, unified action and sound localisation, grounded video QA, and hour-long video QA, alongside an analysis and interpretability track that is still open for submissions. Notably, the unified video QA track introduced a novel subset that reformulates traditional perception tasks (such as point tracking and temporal action localisation) as multiple-choice video QA questions that video-language models can natively tackle. The unified object and point tracking merged the original object tracking and point tracking tasks, whereas the unified action and sound localisation merged the original temporal action localisation and temporal sound localisation tracks. Accordingly, we required competitors to use unified approaches rather than engineered pipelines with task-specific models. By proposing such a unified challenge, Perception Test 2025 highlights the significant difficulties existing models face when tackling diverse perception tasks through unified interfaces.

Method: Construct synthetic long-context training samples by splicing together short, captioned videos from existing datasets. Study different data composition strategies like random vs visually clustered splicing and caption enrichment.

Result: Under identical compute constraints, models trained with VideoWeave achieve higher accuracy than conventional video finetuning on video question answering tasks.

Conclusion: Reorganizing training data rather than altering architectures offers a simple and scalable path for training video-language models, improving data efficiency.

Abstract: Training video-language models is often prohibitively expensive due to the high cost of processing long frame sequences and the limited availability of annotated long videos. We present VideoWeave, a simple yet effective approach to improve data efficiency by constructing synthetic long-context training samples that splice together short, captioned videos from existing datasets. Rather than modifying model architectures or optimization objectives, VideoWeave reorganizes available video-text pairs to expand temporal diversity within fixed compute. We systematically study how different data composition strategies like random versus visually clustered splicing and caption enrichment affect downstream performance on downstream video question answering. Under identical compute constraints, models trained with VideoWeave achieve higher accuracy than conventional video finetuning. Our results highlight that reorganizing training data, rather than altering architectures, may offer a simple and scalable path for training video-language models. We link our code for all experiments here.

Method: Leverages pre-trained text-to-video DiT; localizes relevant tokens via cross-attention; fuses user mask with effect mask; inverts video to structured noise; reinitializes masked tokens with Gaussian noise while preserving background; copies background tokens during denoising.

Result: Outperforms both training-based and training-free baselines on DAVIS and new WIPER-Bench; achieves clean removal and temporally stable reconstruction without retraining; introduces new evaluation metric.

Conclusion: Object-WIPER provides an effective training-free solution for removing dynamic objects and effects from videos with temporal coherence, supported by a new benchmark and evaluation metric.

Abstract: In this paper, we introduce Object-WIPER, a training-free framework for removing dynamic objects and their associated visual effects from videos, and inpainting them with semantically consistent and temporally coherent content. Our approach leverages a pre-trained text-to-video diffusion transformer (DiT). Given an input video, a user-provided object mask, and query tokens describing the target object and its effects, we localize relevant visual tokens via visual-text cross-attention and visual self-attention. This produces an intermediate effect mask that we fuse with the user mask to obtain a final foreground token mask to replace. We first invert the video through the DiT to obtain structured noise, then reinitialize the masked tokens with Gaussian noise while preserving background tokens. During denoising, we copy values for the background tokens saved during inversion to maintain scene fidelity. To address the lack of suitable evaluation, we introduce a new object removal metric that rewards temporal consistency among foreground tokens across consecutive frames, coherence between foreground and background tokens within each frame, and dissimilarity between the input and output foreground tokens. Experiments on DAVIS and a newly curated real-world associated effect benchmark (WIPER-Bench) show that Object-WIPER surpasses both training-based and training-free baselines in terms of the metric, achieving clean removal and temporally stable reconstruction without any retraining. Our new benchmark, source code, and pre-trained models will be publicly available.

Method: Three-stage approach: 1) Few-shot fine-tuning of vision-language model for student action recognition, 2) Sliding window segmentation of 2-minute videos into action sequences, 3) LLM classification of action sequences with classroom context for engagement prediction.

Result: Experimental results demonstrate effectiveness in identifying student engagement using the proposed framework.

Conclusion: The framework successfully addresses limitations of existing methods by enabling few-shot learning and incorporating classroom context for improved student engagement measurement.

Abstract: Understanding student behavior in the classroom is essential to improve both pedagogical quality and student engagement. Existing methods for predicting student engagement typically require substantial annotated data to model the diversity of student behaviors, yet privacy concerns often restrict researchers to their own proprietary datasets. Moreover, the classroom context, represented in peers’ actions, is ignored. To address the aforementioned limitation, we propose a novel three-stage framework for video-based student engagement measurement. First, we explore the few-shot adaptation of the vision-language model for student action recognition, which is fine-tuned to distinguish among action categories with a few training samples. Second, to handle continuous and unpredictable student actions, we utilize the sliding temporal window technique to divide each student’s 2-minute-long video into non-overlapping segments. Each segment is assigned an action category via the fine-tuned VLM model, generating a sequence of action predictions. Finally, we leverage the large language model to classify this entire sequence of actions, together with the classroom context, as belonging to an engaged or disengaged student. The experimental results demonstrate the effectiveness of the proposed approach in identifying student engagement.

Method: Hierarchical pipeline: first outpaints key frames, then completes intermediate frames via interpolation conditioned on completed boundaries. Uses enhanced pretrained image inpainting backbone with 3D windowed attention for spatiotemporal interaction and global feature guidance using OpenCLIP features distilled into compact global tokens.

Result: Comprehensive evaluations on benchmark datasets demonstrate improved reconstruction quality and more natural motion compared to prior methods.

Conclusion: GlobalPaint effectively addresses video outpainting challenges by combining hierarchical processing with enhanced spatiotemporal modeling, achieving better temporal coherence and visual quality than previous approaches.

Abstract: Video outpainting extends a video beyond its original boundaries by synthesizing missing border content. Compared with image outpainting, it requires not only per-frame spatial plausibility but also long-range temporal coherence, especially when outpainted content becomes visible across time under camera or object motion. We propose GlobalPaint, a diffusion-based framework for spatiotemporal coherent video outpainting. Our approach adopts a hierarchical pipeline that first outpaints key frames and then completes intermediate frames via an interpolation model conditioned on the completed boundaries, reducing error accumulation in sequential processing. At the model level, we augment a pretrained image inpainting backbone with (i) an Enhanced Spatial-Temporal module featuring 3D windowed attention for stronger spatiotemporal interaction, and (ii) global feature guidance that distills OpenCLIP features from observed regions across all frames into compact global tokens using a dedicated extractor. Comprehensive evaluations on benchmark datasets demonstrate improved reconstruction quality and more natural motion compared to prior methods. Our demo page is https://yuemingpan.github.io/GlobalPaint/

Method: Created WHU-PCPR dataset with: 1) cross-platform heterogeneous point clouds from survey-grade vehicle-mounted MLS systems and low-cost portable helmet-mounted PLS systems with different mechanical/solid-state LiDARs; 2) complex urban/campus scenes with real-time and long-term changes; 3) large-scale coverage (82.3km trajectory over 60 months).

Result: Established comprehensive dataset with 82.3km trajectory over 60-month period, ~30km unrepeated route. Conducted extensive evaluation of representative PCPR methods and provided analysis of key challenges and future directions.

Conclusion: WHU-PCPR addresses critical gaps in PCPR research by providing diverse, cross-platform, heterogeneous point cloud data with complex scenes and large-scale coverage, enabling more robust place recognition method development and evaluation.

Abstract: Point Cloud-based Place Recognition (PCPR) demonstrates considerable potential in applications such as autonomous driving, robot localization and navigation, and map update. In practical applications, point clouds used for place recognition are often acquired from different platforms and LiDARs across varying scene. However, existing PCPR datasets lack diversity in scenes, platforms, and sensors, which limits the effective development of related research. To address this gap, we establish WHU-PCPR, a cross-platform heterogeneous point cloud dataset designed for place recognition. The dataset differentiates itself from existing datasets through its distinctive characteristics: 1) cross-platform heterogeneous point clouds: collected from survey-grade vehicle-mounted Mobile Laser Scanning (MLS) systems and low-cost Portable helmet-mounted Laser Scanning (PLS) systems, each equipped with distinct mechanical and solid-state LiDAR sensors. 2) Complex localization scenes: encompassing real-time and long-term changes in both urban and campus road scenes. 3) Large-scale spatial coverage: featuring 82.3 km of trajectory over a 60-month period and an unrepeated route of approximately 30 km. Based on WHU-PCPR, we conduct extensive evaluation and in-depth analysis of several representative PCPR methods, and provide a concise discussion of key challenges and future research directions. The dataset and benchmark code are available at https://github.com/zouxianghong/WHU-PCPR.

Method: The authors conduct empirical analysis comparing model performance before and after unsupervised adaptation, providing practical insights for selecting and adapting foundation models for datasets with limited size and labels while leveraging larger unlabeled datasets through unsupervised training.

Result: The paper includes comprehensive quantitative and visual analyses comparing model performance, though specific results are not detailed in the abstract. The study aims to answer critical questions about model appropriateness, unsupervised training strategies, feasible training scale under computational constraints, and downstream performance benefits.

Conclusion: The research provides practical guidance for social health researchers on how to effectively adapt computer vision foundation models for neighborhood built environment analysis, addressing domain transfer challenges and offering evidence-based recommendations for model selection and training strategies in resource-constrained settings.

Abstract: A substantial body of health research demonstrates a strong link between neighborhood environments and health outcomes. Recently, there has been increasing interest in leveraging advances in computer vision to enable large-scale, systematic characterization of neighborhood built environments. However, the generalizability of vision models across fundamentally different domains remains uncertain, for example, transferring knowledge from ImageNet to the distinct visual characteristics of Google Street View (GSV) imagery. In applied fields such as social health research, several critical questions arise: which models are most appropriate, whether to adopt unsupervised training strategies, what training scale is feasible under computational constraints, and how much such strategies benefit downstream performance. These decisions are often costly and require specialized expertise. In this paper, we answer these questions through empirical analysis and provide practical insights into how to select and adapt foundation models for datasets with limited size and labels, while leveraging larger, unlabeled datasets through unsupervised training. Our study includes comprehensive quantitative and visual analyses comparing model performance before and after unsupervised adaptation.

Method: Introduce Ghost-100, a procedurally generated dataset of synthetic scenes with deliberately removed visual details. Use structured 5-Level Prompt Intensity Framework varying prompts from neutral queries to toxic demands and rigid formatting constraints. Evaluate three open-weight VLMs: MiniCPM-V 2.6-8B, Qwen2-VL-7B, and Qwen3-VL-8B.

Result: Across all three models, hallucination rates do not increase monotonically with prompt intensity. All models exhibit reductions at higher intensity levels at different thresholds, though not all show sustained reduction under maximum coercion.

Conclusion: Current safety alignment is more effective at detecting semantic hostility than structural coercion, revealing model-specific limitations in handling compliance pressure. The Ghost-100 dataset enables controlled analysis of absence-based hallucinations.

Abstract: Vision-Language Models (VLMs) are increasingly used in safety-critical applications that require reliable visual grounding. However, these models often hallucinate details that are not present in the image to satisfy user prompts. While recent datasets and benchmarks have been introduced to evaluate systematic hallucinations in VLMs, many hallucination behaviors remain insufficiently characterized. In particular, prior work primarily focuses on object presence or absence, leaving it unclear how prompt phrasing and structural constraints can systematically induce hallucinations. In this paper, we investigate how different forms of prompt pressure influence hallucination behavior. We introduce Ghost-100, a procedurally generated dataset of synthetic scenes in which key visual details are deliberately removed, enabling controlled analysis of absence-based hallucinations. Using a structured 5-Level Prompt Intensity Framework, we vary prompts from neutral queries to toxic demands and rigid formatting constraints. We evaluate three representative open-weight VLMs: MiniCPM-V 2.6-8B, Qwen2-VL-7B, and Qwen3-VL-8B. Across all three models, hallucination rates do not increase monotonically with prompt intensity. All models exhibit reductions at higher intensity levels at different thresholds, though not all show sustained reduction under maximum coercion. These results suggest that current safety alignment is more effective at detecting semantic hostility than structural coercion, revealing model-specific limitations in handling compliance pressure. Our dataset is available at: https://github.com/bli1/tone-matters

Method: Proposes two complementary attack strategies: 1) Vision Attack - perturbs visual token features from 3D encoder and projector to test vision-language alignment robustness; 2) Caption Attack - directly manipulates output token sequences to evaluate end-to-end system robustness. Both include untargeted and targeted variants.

Result: 3D VLMs exhibit significant adversarial vulnerabilities under untargeted attacks, but demonstrate greater resilience against targeted attacks aimed at forcing specific harmful outputs compared to their 2D counterparts.

Conclusion: The study highlights the importance of improving adversarial robustness of 3D VLMs, especially as they are deployed in safety-critical applications, revealing both vulnerabilities and relative strengths compared to 2D models.

Abstract: 3D Vision-Language Models (VLMs), such as PointLLM and GPT4Point, have shown strong reasoning and generalization abilities in 3D understanding tasks. However, their adversarial robustness remains largely unexplored. Prior work in 2D VLMs has shown that the integration of visual inputs significantly increases vulnerability to adversarial attacks, making these models easier to manipulate into generating toxic or misleading outputs. In this paper, we investigate whether incorporating 3D vision similarly compromises the robustness of 3D VLMs. To this end, we present the first systematic study of adversarial robustness in point-based 3D VLMs. We propose two complementary attack strategies: \textit{Vision Attack}, which perturbs the visual token features produced by the 3D encoder and projector to assess the robustness of vision-language alignment; and \textit{Caption Attack}, which directly manipulates output token sequences to evaluate end-to-end system robustness. Each attack includes both untargeted and targeted variants to measure general vulnerability and susceptibility to controlled manipulation. Our experiments reveal that 3D VLMs exhibit significant adversarial vulnerabilities under untargeted attacks, while demonstrating greater resilience against targeted attacks aimed at forcing specific harmful outputs, compared to their 2D counterparts. These findings highlight the importance of improving the adversarial robustness of 3D VLMs, especially as they are deployed in safety-critical applications.

Method: Proposes SparseOccVLA with: 1) Lightweight Sparse Occupancy Encoder generating compact sparse occupancy queries as bridge between vision and language, 2) LLM reasoning on aligned queries for unified scene understanding and future occupancy forecasting, 3) LLM-guided Anchor-Diffusion Planner with decoupled anchor scoring/denoising and cross-model trajectory-condition fusion.

Result: Achieves 7% relative improvement in CIDEr over SOTA on OmniDrive-nuScenes, 0.5 increase in mIoU score on Occ3D-nuScenes, and sets SOTA open-loop planning metric on nuScenes benchmark, demonstrating strong holistic capability.

Conclusion: SparseOccVLA effectively bridges VLMs and semantic occupancy through sparse occupancy queries, enabling unified scene understanding, occupancy forecasting, and trajectory planning with superior performance across multiple autonomous driving benchmarks.

Abstract: In autonomous driving, Vision Language Models (VLMs) excel at high-level reasoning , whereas semantic occupancy provides fine-grained details. Despite significant progress in individual fields, there is still no method that can effectively integrate both paradigms. Conventional VLMs struggle with token explosion and limited spatiotemporal reasoning, while semantic occupancy provides a unified, explicit spatial representation but is too dense to integrate efficiently with VLMs. To address these challenges and bridge the gap between VLMs and occupancy, we propose SparseOccVLA, a novel vision-language-action model that unifies scene understanding, occupancy forecasting, and trajectory planning powered by sparse occupancy queries. Starting with a lightweight Sparse Occupancy Encoder, SparseOccVLA generates compact yet highly informative sparse occupancy queries that serve as the single bridge between vision and language. These queries are aligned into the language space and reasoned by the LLM for unified scene understanding and future occupancy forecasting. Furthermore, we introduce an LLM-guided Anchor-Diffusion Planner featuring decoupled anchor scoring and denoising, as well as cross-model trajectory-condition fusion. SparseOccVLA achieves a 7% relative improvement in CIDEr over the state-of-the-art on OmniDrive-nuScenes, a 0.5 increase in mIoU score on Occ3D-nuScenes, and sets state-of-the-art open-loop planning metric on nuScenes benchmark, demonstrating its strong holistic capability.

Method: VVTRec transforms sparse visibility data into both image-form and text-form features to enhance spatial and semantic information. It leverages pre-trained Vision-Language Models (VLMs) to extract knowledge without additional training, using visibility as a foreign modality to supplement the VLMs’ capabilities.

Result: VVTRec effectively enhances imaging results by exploiting multimodal information without introducing excessive computational overhead, improving structural integrity and accuracy of reconstructed radio astronomy images.

Conclusion: The proposed multimodal approach with visibility-guided visual and textual modality enrichment successfully addresses limitations of single-modality methods, providing cleaner radio astronomy images with better artifact removal and correlation modeling.

Abstract: Radio astronomy is an indispensable discipline for observing distant celestial objects. Measurements of wave signals from radio telescopes, called visibility, need to be transformed into images for astronomical observations. These dirty images blend information from real sources and artifacts. Therefore, astronomers usually perform reconstruction before imaging to obtain cleaner images. Existing methods consider only a single modality of sparse visibility data, resulting in images with remaining artifacts and insufficient modeling of correlation. To enhance the extraction of visibility information and emphasize output quality in the image domain, we propose VVTRec, a multimodal radio interferometric data reconstruction method with visibility-guided visual and textual modality enrichment. In our VVTRec, sparse visibility is transformed into image-form and text-form features to obtain enhancements in terms of spatial and semantic information, improving the structural integrity and accuracy of images. Also, we leverage Vision-Language Models (VLMs) to achieve additional training-free performance improvements. VVTRec enables sparse visibility, as a foreign modality unseen by VLMs, to accurately extract pre-trained knowledge as a supplement. Our experiments demonstrate that VVTRec effectively enhances imaging results by exploiting multimodal information without introducing excessive computational overhead.

Method: 1) Created SRFlow dataset using high-resolution facial optical flow data; 2) Developed SRFlowNet model with tailored regularization losses using masks and gradients (difference or Sobel operator) to suppress noise in texture-less/repetitive-pattern regions; 3) Used Gaussian splatting rasterization to guide high-resolution skin motion capture.

Result: Training with SRFlow dataset reduces EPE by up to 42% (0.5081 to 0.2953) across various optical flow models. SRFlowNet with SRFlow dataset achieves up to 48% F1-score improvement (0.4733 to 0.6947) on composite micro-expression datasets.

Conclusion: The SRFlow dataset and SRFlowNet model significantly advance facial optical flow estimation and micro-expression recognition, demonstrating the value of high-resolution facial motion data and tailored regularization techniques.

Abstract: Facial optical flow supports a wide range of tasks in facial motion analysis. However, the lack of high-resolution facial optical flow datasets has hindered progress in this area. In this paper, we introduce Splatting Rasterization Flow (SRFlow), a high-resolution facial optical flow dataset, and Splatting Rasterization Guided FlowNet (SRFlowNet), a facial optical flow model with tailored regularization losses. These losses constrain flow predictions using masks and gradients computed via difference or Sobel operator. This effectively suppresses high-frequency noise and large-scale errors in texture-less or repetitive-pattern regions, enabling SRFlowNet to be the first model explicitly capable of capturing high-resolution skin motion guided by Gaussian splatting rasterization. Experiments show that training with the SRFlow dataset improves facial optical flow estimation across various optical flow models, reducing end-point error (EPE) by up to 42% (from 0.5081 to 0.2953). Furthermore, when coupled with the SRFlow dataset, SRFlowNet achieves up to a 48% improvement in F1-score (from 0.4733 to 0.6947) on a composite of three micro-expression datasets. These results demonstrate the value of advancing both facial optical flow estimation and micro-expression recognition.

Method: Combines intrinsic geometric descriptors (HKS/WKS) with mesh-agnostic latent embeddings that disentangle facial identity and expression. Uses Jacobian loss with vertex-position and Laplacian losses for geometric consistency. Trained only on human expression pairs.

Result: Achieves plausible cross-species expression transfer, effectively narrowing the geometric gap between human and animal facial shapes without animal training data.

Conclusion: The framework successfully enables zero-shot expression transfer from humans to animals by learning disentangled representations that generalize across species, demonstrating the effectiveness of geometric consistency losses and intrinsic descriptors.

Abstract: We present a zero-shot framework for transferring human facial expressions to 3D animal face meshes. Our method combines intrinsic geometric descriptors (HKS/WKS) with a mesh-agnostic latent embedding that disentangles facial identity and expression. The ID latent space captures species-independent facial structure, while the expression latent space encodes deformation patterns that generalize across humans and animals. Trained only with human expression pairs, the model learns the embeddings, decoupling, and recoupling of cross-identity expressions, enabling expression transfer without requiring animal expression data. To enforce geometric consistency, we employ Jacobian loss together with vertex-position and Laplacian losses. Experiments show that our approach achieves plausible cross-species expression transfer, effectively narrowing the geometric gap between human and animal facial shapes.

Method: Combines frozen CLIP semantic prior with spatially-aware 3D encoder and multimodal decoder optimized with contrastive + captioning objectives, plus test-time search for reward-guided caption selection.

Result: Improves over 3D CoCa by +1.50 CIDEr@0.5IoU on ScanRefer, +1.61 CIDEr@0.5IoU on Nr3D, and +3.8 CIDEr@0.25 in zero-shot OOD evaluation on TOD3Cap.

Conclusion: 3D CoCa v2 provides a generalizable 3D captioning framework that achieves better performance and robustness across diverse 3D environments without parameter updates.

Abstract: Spatial intelligence refers to the ability to perceive, reason about, and describe objects and their relationships within three-dimensional environments, forming a foundation for embodied perception and scene understanding. 3D captioning aims to describe 3D scenes in natural language; however, it remains challenging due to the sparsity and irregularity of point clouds and, more critically, the weak grounding and limited out-of-distribution (OOD) generalization of existing captioners across drastically different environments, including indoor and outdoor 3D scenes. To address this challenge, we propose 3D CoCa v2, a generalizable 3D captioning framework that unifies contrastive vision-language learning with 3D caption generation and further improves robustness via test-time search (TTS) without updating the captioner parameters. 3D CoCa v2 builds on a frozen CLIP-based semantic prior, a spatially-aware 3D scene encoder for geometry, and a multimodal decoder jointly optimized with contrastive and captioning objectives, avoiding external detectors or handcrafted proposals. At inference, TTS produces diverse caption candidates and performs reward-guided selection using a compact scene summary. Experiments show improvements over 3D CoCa of +1.50 CIDEr@0.5IoU on ScanRefer and +1.61 CIDEr@0.5IoU on Nr3D, and +3.8 CIDEr@0.25 in zero-shot OOD evaluation on TOD3Cap. Code will be released at https://github.com/AIGeeksGroup/3DCoCav2.

Method: Hybrid Attention U-Net GAN: integrates Attention Gates into a lightweight U-Net backbone and trains within a conditional adversarial framework to approximate high-frequency fidelity of generative models in a single forward pass, avoiding heavy iterative sampling.

Result: Achieves best-in-class LPIPS score of 0.112 among efficient models on SID dataset, significantly outperforming efficient baselines (SID, EnlightenGAN) while maintaining 0.06s inference latency (40x speedup over latent diffusion models).

Conclusion: Demonstrates that heavy iterative sampling of diffusion models is not strictly necessary for texture recovery; the proposed hybrid Attention U-Net GAN provides generative-level texture recovery at near real-time speeds suitable for edge deployment.

Abstract: Recent advancements in Low-Light Image Enhancement (LLIE) have focused heavily on Diffusion Probabilistic Models, which achieve high perceptual quality but suffer from significant computational latency (often exceeding 2-4 seconds per image). Conversely, traditional CNN-based baselines offer real-time inference but struggle with “over-smoothing,” failing to recover fine structural details in extreme low-light conditions. This creates a practical gap in the literature: the lack of a model that provides generative-level texture recovery at edge-deployable speeds. In this paper, we address this trade-off by proposing a hybrid Attention U-Net GAN. We demonstrate that the heavy iterative sampling of diffusion models is not strictly necessary for texture recovery. Instead, by integrating Attention Gates into a lightweight U-Net backbone and training within a conditional adversarial framework, we can approximate the high-frequency fidelity of generative models in a single forward pass. Extensive experiments on the SID dataset show that our method achieves a best-in-class LPIPS score of 0.112 among efficient models, significantly outperforming efficient baselines (SID, EnlightenGAN) while maintaining an inference latency of 0.06s. This represents a 40x speedup over latent diffusion models, making our approach suitable for near real-time applications.

Method: Introduced BabyVision benchmark with 388 items across 22 subclasses in four key categories to assess core visual abilities independent of linguistic knowledge. Also proposed BabyVision-Gen and automatic evaluation toolkit for solving visual reasoning with generation models.

Result: Leading MLLMs perform significantly below human baselines. Gemini3-Pro-Preview scores only 49.7, lagging behind 6-year-old humans and far below average adult score of 94.1. Despite excelling in knowledge-heavy evaluations, current MLLMs lack fundamental visual primitives.

Conclusion: Progress in BabyVision represents a step toward human-level visual perception and reasoning capabilities. The benchmark reveals that despite strong language performance, MLLMs still lack core visual understanding that humans develop early in life.

Abstract: While humans develop core visual skills long before acquiring language, contemporary Multimodal LLMs (MLLMs) still rely heavily on linguistic priors to compensate for their fragile visual understanding. We uncovered a crucial fact: state-of-the-art MLLMs consistently fail on basic visual tasks that humans, even 3-year-olds, can solve effortlessly. To systematically investigate this gap, we introduce BabyVision, a benchmark designed to assess core visual abilities independent of linguistic knowledge for MLLMs. BabyVision spans a wide range of tasks, with 388 items divided into 22 subclasses across four key categories. Empirical results and human evaluation reveal that leading MLLMs perform significantly below human baselines. Gemini3-Pro-Preview scores 49.7, lagging behind 6-year-old humans and falling well behind the average adult score of 94.1. These results show despite excelling in knowledge-heavy evaluations, current MLLMs still lack fundamental visual primitives. Progress in BabyVision represents a step toward human-level visual perception and reasoning capabilities. We also explore solving visual reasoning with generation models by proposing BabyVision-Gen and automatic evaluation toolkit. Our code and benchmark data are released at https://github.com/UniPat-AI/BabyVision for reproduction.

Method: Proposes Blur Pattern Pretraining (BPP) to acquire blur priors from simulation datasets and transfer them via joint fine-tuning on real data. Introduces Motion and Semantic Guidance (MoSeG) to strengthen priors under severe degradation. Implements GLOWDeblur with convolution-based pre-reconstruction & domain alignment module combined with lightweight diffusion backbone.

Result: Extensive experiments on six widely-used benchmarks and two real-world datasets validate the approach, demonstrating robust generalization and confirming the importance of blur priors. The lightweight design ensures practicality for real-world applications.

Conclusion: Blur pattern diversity is crucial for robust generalization in deblurring. The proposed GLOWDeblur framework effectively addresses generalization limitations through blur pattern pretraining and motion/semantic guidance while maintaining practical lightweight design for real-world use.

Abstract: Image deblurring has advanced rapidly with deep learning, yet most methods exhibit poor generalization beyond their training datasets, with performance dropping significantly in real-world scenarios. Our analysis shows this limitation stems from two factors: datasets face an inherent trade-off between realism and coverage of diverse blur patterns, and algorithmic designs remain restrictive, as pixel-wise losses drive models toward local detail recovery while overlooking structural and semantic consistency, whereas diffusion-based approaches, though perceptually strong, still fail to generalize when trained on narrow datasets with simplistic strategies. Through systematic investigation, we identify blur pattern diversity as the decisive factor for robust generalization and propose Blur Pattern Pretraining (BPP), which acquires blur priors from simulation datasets and transfers them through joint fine-tuning on real data. We further introduce Motion and Semantic Guidance (MoSeG) to strengthen blur priors under severe degradation, and integrate it into GLOWDeblur, a Generalizable reaL-wOrld lightWeight Deblur model that combines convolution-based pre-reconstruction & domain alignment module with a lightweight diffusion backbone. Extensive experiments on six widely-used benchmarks and two real-world datasets validate our approach, confirming the importance of blur priors for robust generalization and demonstrating that the lightweight design of GLOWDeblur ensures practicality in real-world applications. The project page is available at https://vegdog007.github.io/GLOWDeblur_Website/.

Method: The approach estimates keypoint z-coordinates in a virtual camera space normalized by focal length and image size for camera-agnostic depth prediction. It then uses a self-supervised test-time optimization strategy with a 3D consistency loss between predicted and scale-transformed hand poses to adapt to target domains without ground truth.

Result: V-HPOT achieves 71% reduction in mean pose error on H2O dataset and 41% reduction on AssemblyHands dataset. It outperforms all single-stage approaches and competes closely with two-stage methods while requiring 3.5x to 14x less data.

Conclusion: V-HPOT effectively addresses cross-domain generalization challenges in 3D hand pose estimation through camera-agnostic depth prediction and self-supervised test-time adaptation, demonstrating strong performance with significantly less data requirements.

Abstract: We present V-HPOT, a novel approach for improving the cross-domain performance of 3D hand pose estimation from egocentric images across diverse, unseen domains. State-of-the-art methods demonstrate strong performance when trained and tested within the same domain. However, they struggle to generalise to new environments due to limited training data and depth perception – overfitting to specific camera intrinsics. Our method addresses this by estimating keypoint z-coordinates in a virtual camera space, normalised by focal length and image size, enabling camera-agnostic depth prediction. We further leverage this invariance to camera intrinsics to propose a self-supervised test-time optimisation strategy that refines the model’s depth perception during inference. This is achieved by applying a 3D consistency loss between predicted and in-space scale-transformed hand poses, allowing the model to adapt to target domain characteristics without requiring ground truth annotations. V-HPOT significantly improves 3D hand pose estimation performance in cross-domain scenarios, achieving a 71% reduction in mean pose error on the H2O dataset and a 41% reduction on the AssemblyHands dataset. Compared to state-of-the-art methods, V-HPOT outperforms all single-stage approaches across all datasets and competes closely with two-stage methods, despite needing approximately x3.5 to x14 less data.

Method: Bionic design decoupling localization (Grounding DINO as eyes) from understanding (LLaVA-OneVision as brain); Spatio-Temporal Fusion Module for trajectory comprehension; progressive three-stage training (Visual Alignment, Temporal Fine-tuning, Semantic Injection via LoRA).

Result: State-of-the-art performance on BenSMOT benchmark, significantly outperforming existing methods in instance description, interaction recognition, and video summarization while maintaining robust tracking stability.

Conclusion: LLMTrack successfully bridges geometric perception with cognitive reasoning, enabling semantic understanding of object behaviors beyond traditional tracking capabilities.

Abstract: Traditional Multi-Object Tracking (MOT) systems have achieved remarkable precision in localization and association, effectively answering \textit{where} and \textit{who}. However, they often function as autistic observers, capable of tracing geometric paths but blind to the semantic \textit{what} and \textit{why} behind object behaviors. To bridge the gap between geometric perception and cognitive reasoning, we propose \textbf{LLMTrack}, a novel end-to-end framework for Semantic Multi-Object Tracking (SMOT). We adopt a bionic design philosophy that decouples strong localization from deep understanding, utilizing Grounding DINO as the eyes and the LLaVA-OneVision multimodal large model as the brain. We introduce a Spatio-Temporal Fusion Module that aggregates instance-level interaction features and video-level contexts, enabling the Large Language Model (LLM) to comprehend complex trajectories. Furthermore, we design a progressive three-stage training strategy, Visual Alignment, Temporal Fine-tuning, and Semantic Injection via LoRA to efficiently adapt the massive model to the tracking domain. Extensive experiments on the BenSMOT benchmark demonstrate that LLMTrack achieves state-of-the-art performance, significantly outperforming existing methods in instance description, interaction recognition, and video summarization while maintaining robust tracking stability.

Method: ArrowGEV uses reinforcement learning to explicitly model temporal directionality. It categorizes events into time-sensitive (meaning changes with reversal) and time-insensitive (meaning unchanged with reversal). For time-sensitive events, it rewards VLMs for discriminating between forward/backward videos; for time-insensitive events, it enforces consistent grounding across both directions.

Result: Extensive experiments show ArrowGEV improves grounding precision, temporal directionality recognition, and enhances general video understanding and reasoning ability.

Conclusion: Explicitly modeling temporal directionality through reinforcement learning significantly improves VLM performance on video event grounding tasks, demonstrating the importance of capturing the inherent directionality of temporal processes.

Abstract: Grounding events in videos serves as a fundamental capability in video analysis. While Vision-Language Models (VLMs) are increasingly employed for this task, existing approaches predominantly train models to associate events with timestamps in the forward video only. This paradigm hinders VLMs from capturing the inherent temporal structure and directionality of events, thereby limiting robustness and generalization. To address this limitation, inspired by the arrow of time in physics, which characterizes the intrinsic directionality of temporal processes, we propose ArrowGEV, a reinforcement learning framework that explicitly models temporal directionality in events to improve both event grounding and temporal directionality understanding in VLMs. Specifically, we categorize events into time-sensitive (e.g., putting down a bag) and time-insensitive (e.g., holding a towel in the left hand). The former denote events whose reversal substantially alters their meaning, while the latter remain semantically unchanged under reversal. For time-sensitive events, ArrowGEV introduces a reward that encourages VLMs to discriminate between forward and backward videos, whereas for time-insensitive events, it enforces consistent grounding across both directions. Extensive experiments demonstrate that ArrowGEV not only improves grounding precision and temporal directionality recognition, but also enhances general video understanding and reasoning ability.

Method: QCaption fuses three models: 1) key frame extraction to identify important video frames, 2) a Large Multimodal Model (LMM) for image-text analysis of extracted frames, and 3) a Large Language Model (LLM) for text analysis and integration. This multi-model fusion enables comprehensive video understanding.

Result: Experimental results show QCaption achieves up to 44.2% improvement in video captioning and 48.9% improvement in Q&A tasks compared to existing methods. Ablation studies confirm the importance of LLM fusion, and additional benchmarking shows QCaption outperforms other proposed approaches.

Conclusion: QCaption demonstrates the effectiveness of model fusion for advancing video analytics, providing significant performance gains while maintaining self-contained, on-premises deployment capability, showing promise for practical video understanding applications.

Abstract: This paper introduces QCaption, a novel video captioning and Q&A pipeline that enhances video analytics by fusing three models: key frame extraction, a Large Multimodal Model (LMM) for image-text analysis, and a Large Language Model (LLM) for text analysis. This approach enables integrated analysis of text, images, and video, achieving performance improvements over existing video captioning and Q&A models; all while remaining fully self-contained, adept for on-premises deployment. Experimental results using QCaption demonstrated up to 44.2% and 48.9% improvements in video captioning and Q&A tasks, respectively. Ablation studies were also performed to assess the role of LLM on the fusion on the results. Moreover, the paper proposes and evaluates additional video captioning approaches, benchmarking them against QCaption and existing methodologies. QCaption demonstrate the potential of adopting a model fusion approach in advancing video analytics.

Method: APEX (Adaptive Priority-based Efficient X-objective Alignment) uses Dual-Stage Adaptive Normalization to stabilize heterogeneous rewards and P^3 Adaptive Priorities (learning potential, conflict penalty, progress need) to dynamically schedule objectives.

Result: On Stable Diffusion 3.5, APEX achieves balanced gains: +1.31 PickScore, +0.35 DeQA, +0.53 Aesthetics while maintaining competitive OCR accuracy, improving Pareto trade-offs across four heterogeneous objectives.

Conclusion: APEX effectively mitigates multi-objective alignment instability by addressing mechanistic causes of optimization imbalance, providing a more robust approach for heterogeneous reward alignment in text-to-image generation.

Abstract: Multi-objective alignment for text-to-image generation is commonly implemented via static linear scalarization, but fixed weights often fail under heterogeneous rewards, leading to optimization imbalance where models overfit high-variance, high-responsiveness objectives (e.g., OCR) while under-optimizing perceptual goals. We identify two mechanistic causes: variance hijacking, where reward dispersion induces implicit reweighting that dominates the normalized training signal, and gradient conflicts, where competing objectives produce opposing update directions and trigger seesaw-like oscillations. We propose APEX (Adaptive Priority-based Efficient X-objective Alignment), which stabilizes heterogeneous rewards with Dual-Stage Adaptive Normalization and dynamically schedules objectives via P^3 Adaptive Priorities that combine learning potential, conflict penalty, and progress need. On Stable Diffusion 3.5, APEX achieves improved Pareto trade-offs across four heterogeneous objectives, with balanced gains of +1.31 PickScore, +0.35 DeQA, and +0.53 Aesthetics while maintaining competitive OCR accuracy, mitigating the instability of multi-objective alignment.

Method: Reformulates style-guided synthesis as in-context learning task. Uses pretrained ReFlow-based inpainting model to concatenate reference style image with masked target image. Proposes Dynamic Semantic-Style Integration (DSSI) mechanism to reweight attention between textual semantic and style visual tokens, resolving guidance conflicts.

Result: Achieves high-fidelity stylization with superior semantic-style balance and visual quality, outperforming complex prior methods that suffer from artifacts.

Conclusion: Offers a simple yet powerful training-free alternative to existing artifact-prone methods for precise text-guided image stylization with visual exemplars.

Abstract: Text-guided image generation has advanced rapidly with large-scale diffusion models, yet achieving precise stylization with visual exemplars remains difficult. Existing approaches often depend on task-specific retraining or expensive inversion procedures, which can compromise content integrity, reduce style fidelity, and lead to an unsatisfactory trade-off between semantic prompt adherence and style alignment. In this work, we introduce a training-free framework that reformulates style-guided synthesis as an in-context learning task. Guided by textual semantic prompts, our method concatenates a reference style image with a masked target image, leveraging a pretrained ReFlow-based inpainting model to seamlessly integrate semantic content with the desired style through multimodal attention fusion. We further analyze the imbalance and noise sensitivity inherent in multimodal attention fusion and propose a Dynamic Semantic-Style Integration (DSSI) mechanism that reweights attention between textual semantic and style visual tokens, effectively resolving guidance conflicts and enhancing output coherence. Experiments show that our approach achieves high-fidelity stylization with superior semantic-style balance and visual quality, offering a simple yet powerful alternative to complex, artifact-prone prior methods.

Method: Proposes Pseudo-Label Unmixing (PLU) that identifies erroneous pseudo-labels for overlapping instances and regenerates organoid labels through instance decomposition. Uses contour-based synthesis for efficient organoid instance generation, with instance-level augmentations on pseudo-labels before synthesis to enhance synthetic data effectiveness.

Result: Achieves performance comparable to fully supervised models using only 10% labeled data, with state-of-the-art results on two organoid datasets. Ablation studies validate contributions of PLU, contour-based synthesis, and augmentation-aware training.

Conclusion: By addressing overlap at both pseudo-label and synthesis levels, this work advances scalable, label-efficient organoid analysis, unlocking potential for high-throughput applications in precision medicine.

Abstract: Organoids, sophisticated in vitro models of human tissues, are crucial for medical research due to their ability to simulate organ functions and assess drug responses accurately. Accurate organoid instance segmentation is critical for quantifying their dynamic behaviors, yet remains profoundly limited by high-quality annotated datasets and pervasive overlap in microscopy imaging. While semi-supervised learning (SSL) offers a solution to alleviate reliance on scarce labeled data, conventional SSL frameworks suffer from biases induced by noisy pseudo-labels, particularly in overlapping regions. Synthesis-assisted SSL (SA-SSL) has been proposed for mitigating training biases in semi-supervised semantic segmentation. We present the first adaptation of SA-SSL to organoid instance segmentation and reveal that SA-SSL struggles to disentangle intertwined organoids, often misrepresenting overlapping instances as a single entity. To overcome this, we propose Pseudo-Label Unmixing (PLU), which identifies erroneous pseudo-labels for overlapping instances and then regenerates organoid labels through instance decomposition. For image synthesis, we apply a contour-based approach to synthesize organoid instances efficiently, particularly for overlapping cases. Instance-level augmentations (IA) on pseudo-labels before image synthesis further enhances the effect of synthetic data (SD). Rigorous experiments on two organoid datasets demonstrate our method’s effectiveness, achieving performance comparable to fully supervised models using only 10% labeled data, and state-of-the-art results. Ablation studies validate the contributions of PLU, contour-based synthesis, and augmentation-aware training. By addressing overlap at both pseudo-label and synthesis levels, our work advances scalable, label-efficient organoid analysis, unlocking new potential for high-throughput applications in precision medicine.

Method: Created eSkiTB dataset by converting SkiTB RGB videos to event data without neural interpolation for fair comparison. Benchmarked SDTrack (spiking transformer) against STARK (RGB transformer) for tracking performance.

Result: Event-based tracking (SDTrack) achieved 0.685 IoU in scenes with static overlays, outperforming RGB by +20.0 points. Overall mean IoU of 0.711 across dataset, showing temporal contrast is reliable for tracking ballistic motion in congested environments.

Conclusion: eSkiTB establishes first controlled benchmark for event-based tracking in winter sports. Event cameras show promise for ski tracking, especially in cluttered broadcast environments with static overlays.

Abstract: Tracking skiers in RGB broadcast footage is challenging due to motion blur, static overlays, and clutter that obscure the fast-moving athlete. Event cameras, with their asynchronous contrast sensing, offer natural robustness to such artifacts, yet a controlled benchmark for winter-sport tracking has been missing. We introduce event SkiTB (eSkiTB), a synthetic event-based ski tracking dataset generated from SkiTB using direct video-to-event conversion without neural interpolation, enabling an iso-informational comparison between RGB and event modalities. Benchmarking SDTrack (spiking transformer) against STARK (RGB transformer), we find that event-based tracking is substantially resilient to broadcast clutter in scenes dominated by static overlays, achieving 0.685 IoU, outperforming RGB by +20.0 points. Across the dataset, SDTrack attains a mean IoU of 0.711, demonstrating that temporal contrast is a reliable cue for tracking ballistic motion in visually congested environments. eSkiTB establishes the first controlled setting for event-based tracking in winter sports and highlights the promise of event cameras for ski tracking. The dataset and code will be released at https://github.com/eventbasedvision/eSkiTB.

Method: 1) Interactive quantification tool using Segment Anything Model (SAM) for near-perfect particle segmentation with minimal user input. 2) Classification pipeline using segmentation masks to spatially constrain DINOv2 vision transformer, extracting features exclusively from particle regions while suppressing background noise.

Result: Achieves 95.5% accuracy in distinguishing between four different CNT morphologies on 1,800 TEM images, significantly outperforming current baseline despite using less training data. Can resolve mixed samples by correctly classifying distinct particle types within single field of view.

Conclusion: Integrating zero-shot segmentation with self-supervised feature learning enables high-throughput, reproducible nanomaterial analysis, transforming labor-intensive processes into scalable, data-driven workflows.

Abstract: Accurate characterization of carbon nanotube morphologies in electron microscopy images is vital for exposure assessment and toxicological studies, yet current workflows rely on slow, subjective manual segmentation. This work presents a unified framework leveraging vision foundation models to automate the quantification and classification of CNTs in electron microscopy images. First, we introduce an interactive quantification tool built on the Segment Anything Model (SAM) that segments particles with near-perfect accuracy using minimal user input. Second, we propose a novel classification pipeline that utilizes these segmentation masks to spatially constrain a DINOv2 vision transformer, extracting features exclusively from particle regions while suppressing background noise. Evaluated on a dataset of 1,800 TEM images, this architecture achieves 95.5% accuracy in distinguishing between four different CNT morphologies, significantly outperforming the current baseline despite using a fraction of the training data. Crucially, this instance-level processing allows the framework to resolve mixed samples, correctly classifying distinct particle types co-existing within a single field of view. These results demonstrate that integrating zero-shot segmentation with self-supervised feature learning enables high-throughput, reproducible nanomaterial analysis, transforming a labor-intensive bottleneck into a scalable, data-driven process.

Method: Two controlled experiments: (1) varying pupil sizes with/without pupil size normalization using autoencoder-based identity-preserving image-to-image translation, and (2) synthetic iris generation with both genuine and impostor pairs. Human participants performed verification tasks comparing iris image pairs.

Result: Pupil size normalization significantly improves human verification accuracy. Humans can determine same/different eyes for authentic or synthetic pairs, but accuracy declines when comparing authentic vs. high-quality synthetic same-eye counterparts. Synthetic same-eye images are more often judged as different-eye images compared to authentic same-eye pairs.

Conclusion: Pupil size alignment is crucial for human-involved iris matching, and despite high-fidelity generative models, humans struggle to recognize synthetic same-eye iris pairs as matching, highlighting limitations in synthetic iris detection and verification accuracy.

Abstract: Iris recognition is a mature biometric technology offering remarkable precision and speed, and allowing for large-scale deployments to populations exceeding a billion enrolled users (e.g., AADHAAR in India). However, in forensic applications, a human expert may be needed to review and confirm a positive identification before an iris matching result can be presented as evidence in court, especially in cases where processed samples are degraded (e.g., in post-mortem cases) or where there is a need to judge whether the sample is authentic, rather than a result of a presentation attack. This paper presents a study that examines human performance in iris verification in two controlled scenarios: (a) under varying pupil sizes, with and without a linear/nonlinear alignment of the pupil size between compared images, and (b) when both genuine and impostor iris image pairs are synthetically generated. The results demonstrate that pupil size normalization carried out by a modern autoencoder-based identity-preserving image-to-image translation model significantly improves verification accuracy. Participants were also able to determine whether iris pairs corresponded to the same or different eyes when both images were either authentic or synthetic. However, accuracy declined when subjects were comparing authentic irises against high-quality, same-eye synthetic counterparts. These findings (a) demonstrate the importance of pupil-size alignment for iris matching tasks in which humans are involved, and (b) indicate that despite the high fidelity of modern generative models, same-eye synthetic iris images are more often judged by humans as different-eye images, compared to same-eye authentic image pairs. We offer data and human judgments along with this paper to allow full replicability of this study and future works.

Method: Introduces MedGaze-Bench using clinician gaze as Cognitive Cursor with Three-Dimensional Clinical Intent Framework: (1) Spatial Intent for target discrimination amid visual noise, (2) Temporal Intent for causal rationale through retrospective/prospective reasoning, and (3) Standard Intent for protocol compliance verification. Includes Trap QA mechanisms to stress-test reliability by penalizing hallucinations and cognitive sycophancy.

Result: Experiments reveal current MLLMs struggle with egocentric intent understanding due to over-reliance on global features, leading to fabricated observations and uncritical acceptance of invalid instructions.

Conclusion: MedGaze-Bench addresses critical gaps in evaluating Med-MLLMs’ clinical intent understanding and reveals fundamental limitations in current models that must be addressed for safe real-world deployment.

Abstract: Medical Multimodal Large Language Models (Med-MLLMs) require egocentric clinical intent understanding for real-world deployment, yet existing benchmarks fail to evaluate this critical capability. To address these challenges, we introduce MedGaze-Bench, the first benchmark leveraging clinician gaze as a Cognitive Cursor to assess intent understanding across surgery, emergency simulation, and diagnostic interpretation. Our benchmark addresses three fundamental challenges: visual homogeneity of anatomical structures, strict temporal-causal dependencies in clinical workflows, and implicit adherence to safety protocols. We propose a Three-Dimensional Clinical Intent Framework evaluating: (1) Spatial Intent: discriminating precise targets amid visual noise, (2) Temporal Intent: inferring causal rationale through retrospective and prospective reasoning, and (3) Standard Intent: verifying protocol compliance through safety checks. Beyond accuracy metrics, we introduce Trap QA mechanisms to stress-test clinical reliability by penalizing hallucinations and cognitive sycophancy. Experiments reveal current MLLMs struggle with egocentric intent due to over-reliance on global features, leading to fabricated observations and uncritical acceptance of invalid instructions.

Method: Introduces a differentiable Normalized Difference Layer as a neural network module that learns band coefficients from data. Uses softplus reparameterization to ensure positive coefficients and bounded denominators. Provides complete mathematical framework with forward/backward pass algorithms for end-to-end training via backpropagation. Extends to work with signed inputs for stacking in larger architectures.

Result: Models using the Normalized Difference Layer achieve similar classification accuracy to standard MLPs while using ~75% fewer parameters. They demonstrate strong robustness to multiplicative noise - at 10% noise, accuracy drops only 0.17% vs 3.03% for baseline MLPs. Learned coefficient patterns remain consistent across different network depths.

Conclusion: The Normalized Difference Layer successfully bridges classical remote sensing techniques with modern deep learning by preserving key benefits of normalized differences while enabling data-driven coefficient optimization, resulting in parameter-efficient and noise-robust models.

Abstract: Normalized difference indices have been a staple in remote sensing for decades. They stay reliable under lighting changes produce bounded values and connect well to biophysical signals. Even so, they are usually treated as a fixed pre processing step with coefficients set to one, which limits how well they can adapt to a specific learning task. In this study, we introduce the Normalized Difference Layer that is a differentiable neural network module. The proposed method keeps the classical idea but learns the band coefficients from data. We present a complete mathematical framework for integrating this layer into deep learning architectures that uses softplus reparameterization to ensure positive coefficients and bounded denominators. We describe forward and backward pass algorithms enabling end to end training through backpropagation. This approach preserves the key benefits of normalized differences, namely illumination invariance and outputs bounded to $[-1,1]$ while allowing gradient descent to discover task specific band weightings. We extend the method to work with signed inputs, so the layer can be stacked inside larger architectures. Experiments show that models using this layer reach similar classification accuracy to standard multilayer perceptrons while using about 75% fewer parameters. They also handle multiplicative noise well, at 10% noise accuracy drops only 0.17% versus 3.03% for baseline MLPs. The learned coefficient patterns stay consistent across different depths.

Method: Introduces Clifford Algebra Network (CAN) using Clifford Geometric Product (uv = u·v + u∧v) as a unified interaction mechanism that simultaneously captures feature coherence (inner product) and structural variation (wedge product), implemented via efficient sparse rolling with O(N) complexity.

Result: Achieves new Pareto frontier: Nano variant gets 76.41% accuracy on CIFAR-100 with only 1.4M parameters (8× fewer than ResNet-18’s 11.2M), Base variant sets new SOTA for tiny models at 78.05%, and geometric interactions make FFNs redundant.

Conclusion: Global understanding can emerge from algebraically complete local geometric interactions, suggesting a potential paradigm shift where “geometry is all you need” in vision architectures.

Abstract: Modern computer vision architectures, from CNNs to Transformers, predominantly rely on the stacking of heuristic modules: spatial mixers (Attention/Conv) followed by channel mixers (FFNs). In this work, we challenge this paradigm by returning to mathematical first principles. We propose the \textbf{Clifford Algebra Network (CAN)}, also referred to as CliffordNet, a vision backbone grounded purely in Geometric Algebra. Instead of engineering separate modules for mixing and memory, we derive a unified interaction mechanism based on the \textbf{Clifford Geometric Product} ($uv = u \cdot v + u \wedge v$). This operation ensures algebraic completeness regarding the Geometric Product by simultaneously capturing feature coherence (via the generalized inner product) and structural variation (via the exterior wedge product). Implemented via an efficient sparse rolling mechanism with \textbf{strict linear complexity $\mathcal{O}(N)$}, our model reveals a surprising emergent property: the geometric interaction is so representationally dense that standard Feed-Forward Networks (FFNs) become redundant. Empirically, CliffordNet establishes a new Pareto frontier: our \textbf{Nano} variant achieves \textbf{76.41%} accuracy on CIFAR-100 with only \textbf{1.4M} parameters, effectively matching the heavy-weight ResNet-18 (11.2M) with \textbf{$8\times$ fewer parameters}, while our \textbf{Base} variant sets a new SOTA for tiny models at \textbf{78.05%}. Our results suggest that global understanding can emerge solely from rigorous, algebraically complete local interactions, potentially signaling a shift where \textit{geometry is all you need}. Code is available at https://github.com/ParaMind2025/CAN.

Method: 1) Allow full environment exploration before task execution, 2) Construct Spatial Scene Graph (SSG) to capture global spatial structure and semantics, 3) Integrate agent-centric spatial map, compass-aligned visual representation, and remote object localization strategy.

Result: SpatialNav significantly outperforms existing zero-shot agents in both discrete and continuous environments, and clearly narrows the gap with state-of-the-art learning-based methods.

Conclusion: Global spatial representations are crucial for generalizable navigation in zero-shot VLN settings, as demonstrated by the success of SpatialNav’s SSG-based approach.

Abstract: Although learning-based vision-and-language navigation (VLN) agents can learn spatial knowledge implicitly from large-scale training data, zero-shot VLN agents lack this process, relying primarily on local observations for navigation, which leads to inefficient exploration and a significant performance gap. To deal with the problem, we consider a zero-shot VLN setting that agents are allowed to fully explore the environment before task execution. Then, we construct the Spatial Scene Graph (SSG) to explicitly capture global spatial structure and semantics in the explored environment. Based on the SSG, we introduce SpatialNav, a zero-shot VLN agent that integrates an agent-centric spatial map, a compass-aligned visual representation, and a remote object localization strategy for efficient navigation. Comprehensive experiments in both discrete and continuous environments demonstrate that SpatialNav significantly outperforms existing zero-shot agents and clearly narrows the gap with state-of-the-art learning-based methods. Such results highlight the importance of global spatial representations for generalizable navigation.

Method: SARA uses geometry-first pair selection with a lightweight pre-matching stage using mutual nearest neighbors and RANSAC to estimate overlap and parallax cues. It constructs an Information-Weighted Spanning Tree (IWST) augmented with targeted edges for loop closure, long-baseline anchors, and weak-view reinforcement.

Result: Compared to exhaustive matching, SARA reduces rotation errors by 46.5±5.5% and translation errors by 12.5±6.5% across modern learned detectors. Achieves up to 50x speedup through 98% pair reduction (from 30,848 to 580 pairs), reducing matching complexity from quadratic to quasi-linear.

Conclusion: SARA demonstrates that geometry-driven pair selection significantly improves reconstruction accuracy while dramatically reducing computational cost, maintaining within ±3% of baseline reconstruction metrics for downstream applications like 3D Gaussian Splatting and SVRaster.

Abstract: We present SARA (Scene-Aware Reconstruction Accelerator), a geometry-driven pair selection module for Structure-from-Motion (SfM). Unlike conventional pipelines that select pairs based on visual similarity alone, SARA introduces geometry-first pair selection by scoring reconstruction informativeness - the product of overlap and parallax - before expensive matching. A lightweight pre-matching stage uses mutual nearest neighbors and RANSAC to estimate these cues, then constructs an Information-Weighted Spanning Tree (IWST) augmented with targeted edges for loop closure, long-baseline anchors, and weak-view reinforcement. Compared to exhaustive matching, SARA reduces rotation errors by 46.5+-5.5% and translation errors by 12.5+-6.5% across modern learned detectors, while achieving at most 50x speedup through 98% pair reduction (from 30,848 to 580 pairs). This reduces matching complexity from quadratic to quasi-linear, maintaining within +-3% of baseline reconstruction metrics for 3D Gaussian Splatting and SVRaster.

Method: Proposes LR2Flow framework that integrates wavelet tight frame blocks with normalizing flows to learn low-resolution representations. The design uses invertible neural networks in the wavelet tight frame domain based on reconstruction error analysis.

Result: Experimental results on image rescaling, compression, and denoising tasks demonstrate the effectiveness of the learned representations and the robustness of the proposed framework.

Conclusion: LR2Flow provides an effective nonlinear framework for learning low-resolution image representations that balances efficiency with reconstruction accuracy, validated across multiple image processing applications.

Abstract: Low-resolution image representation is a special form of sparse representation that retains only low-frequency information while discarding high-frequency components. This property reduces storage and transmission costs and benefits various image processing tasks. However, a key challenge is to preserve essential visual content while maintaining the ability to accurately reconstruct the original images. This work proposes LR2Flow, a nonlinear framework that learns low-resolution image representations by integrating wavelet tight frame blocks with normalizing flows. We conduct a reconstruction error analysis of the proposed network, which demonstrates the necessity of designing invertible neural networks in the wavelet tight frame domain. Experimental results on various tasks, including image rescaling, compression, and denoising, demonstrate the effectiveness of the learned representations and the robustness of the proposed framework.

Method: Three core contributions: 1) Cross-Modal Semantic Alignment with Optical-Aware SAR Encoder using teacher-student distillation, 2) Semantically-Grounded Generative Guidance with ControlNet using class-aware text prompts and hierarchical visual prompts, 3) Uncertainty-Aware Objective that models aleatoric uncertainty to dynamically modulate reconstruction focus.

Result: Extensive experiments demonstrate superior perceptual quality and semantic consistency compared to state-of-the-art approaches.

Conclusion: The proposed framework effectively addresses limitations of SAR-to-optical translation by mitigating artifacts caused by speckle-induced ambiguity through integrated semantic alignment, generative guidance, and uncertainty modeling.

Abstract: Synthetic Aperture Radar (SAR) provides robust all-weather imaging capabilities; however, translating SAR observations into photo-realistic optical images remains a fundamentally ill-posed problem. Current approaches are often hindered by the inherent speckle noise and geometric distortions of SAR data, which frequently result in semantic misinterpretation, ambiguous texture synthesis, and structural hallucinations. To address these limitations, a novel SAR-to-Optical (S2O) translation framework is proposed, integrating three core technical contributions: (i) Cross-Modal Semantic Alignment, which establishes an Optical-Aware SAR Encoder by distilling robust semantic priors from an Optical Teacher into a SAR Student (ii) Semantically-Grounded Generative Guidance, realized by a Semantically-Grounded ControlNet that integrates class-aware text prompts for global context with hierarchical visual prompts for local spatial guidance; and (iii) an Uncertainty-Aware Objective, which explicitly models aleatoric uncertainty to dynamically modulate the reconstruction focus, effectively mitigating artifacts caused by speckle-induced ambiguity. Extensive experiments demonstrate that the proposed method achieves superior perceptual quality and semantic consistency compared to state-of-the-art approaches.

Method: PRISM treats RGB color space as the stratification domain and imposes a maximum capacity k per color bin. It allocates sampling density proportional to chromatic diversity, preserving texture-rich regions with high color variation while reducing homogeneous surfaces.

Result: The method shifts the sampling space from spatial coverage to visual complexity, producing sparser point clouds that retain essential features for 3D reconstruction tasks.

Conclusion: PRISM offers a novel approach to point cloud downsampling that leverages color information to preserve visually important features while achieving significant data reduction, outperforming conventional spatial-only methods like Random Sampling, Voxel Grid, and Normal Space Sampling.

Abstract: We present PRISM, a novel color-guided stratified sampling method for RGB-LiDAR point clouds. Our approach is motivated by the observation that unique scene features often exhibit chromatic diversity while repetitive, redundant features are homogeneous in color. Conventional downsampling methods (Random Sampling, Voxel Grid, Normal Space Sampling) enforce spatial uniformity while ignoring this photometric content. In contrast, PRISM allocates sampling density proportional to chormatic diversity. By treating RGB color space as the stratification domain and imposing a maximum capacity k per color bin, the method preserves texture-rich regions with high color variation while substantially reducing visually homogeneous surfaces. This shifts the sampling space from spatial coverage to visual complexity to produce sparser point clouds that retain essential features for 3D reconstruction tasks.

Method: Proposes three designs to relax positional continuity: Overlapped, Group-Decoupled, and Gap-Isolated. These enable parallel streaming by allowing simultaneous perception and generation, breaking the sequential perception-generation cycle constraint.

Result: Group-Decoupled achieves the best efficiency-performance balance, maintaining high fluency and accuracy while significantly reducing latency. The framework yields up to 2x acceleration under balanced perception-generation workloads.

Conclusion: The proposed parallel streaming framework establishes a principled pathway toward speak-while-watching real-time systems by addressing the fundamental positional continuity bottleneck in MLLMs for video understanding.

Abstract: Multimodal Large Language Models (MLLMs) have achieved strong performance across many tasks, yet most systems remain limited to offline inference, requiring complete inputs before generating outputs. Recent streaming methods reduce latency by interleaving perception and generation, but still enforce a sequential perception-generation cycle, limiting real-time interaction. In this work, we target a fundamental bottleneck that arises when extending MLLMs to real-time video understanding: the global positional continuity constraint imposed by standard positional encoding schemes. While natural in offline inference, this constraint tightly couples perception and generation, preventing effective input-output parallelism. To address this limitation, we propose a parallel streaming framework that relaxes positional continuity through three designs: Overlapped, Group-Decoupled, and Gap-Isolated. These designs enable simultaneous perception and generation, allowing the model to process incoming inputs while producing responses in real time. Extensive experiments reveal that Group-Decoupled achieves the best efficiency-performance balance, maintaining high fluency and accuracy while significantly reducing latency. We further show that the proposed framework yields up to 2x acceleration under balanced perception-generation workloads, establishing a principled pathway toward speak-while-watching real-time systems. We make all our code publicly available: https://github.com/EIT-NLP/Speak-While-Watching.

Method: Automated pipeline uses expert masks as spatial anchors to derive localization targets, extracts shape/spatial cues, guides VLMs to synthesize natural clinical queries, and implements multi-stage verification with formatting checks, geometry/medical-prior rules, and visual judging.

Result: Created MedGround-35K dataset; VLMs trained with it show improved referring grounding performance, better multi-object semantic disambiguation, and strong generalization to unseen grounding settings.

Conclusion: MedGround provides a scalable, data-driven approach to anchor medical language to verifiable visual evidence, addressing the visual grounding limitation in clinical VLMs through automated high-quality dataset generation.

Abstract: Vision-Language Models (VLMs) can generate convincing clinical narratives, yet frequently struggle to visually ground their statements. We posit this limitation arises from the scarcity of high-quality, large-scale clinical referring-localization pairs. To address this, we introduce MedGround, an automated pipeline that transforms segmentation resources into high-quality medical referring grounding data. Leveraging expert masks as spatial anchors, MedGround precisely derives localization targets, extracts shape and spatial cues, and guides VLMs to synthesize natural, clinically grounded queries that reflect morphology and location. To ensure data rigor, a multi-stage verification system integrates strict formatting checks, geometry- and medical-prior rules, and image-based visual judging to filter out ambiguous or visually unsupported samples. Finally, we present MedGround-35K, a novel multimodal medical dataset. Extensive experiments demonstrate that VLMs trained with MedGround-35K consistently achieve improved referring grounding performance, enhance multi-object semantic disambiguation, and exhibit strong generalization to unseen grounding settings. This work highlights MedGround as a scalable, data-driven approach to anchor medical language to verifiable visual evidence. Dataset and code will be released publicly upon acceptance.

Method: Proposes MVGGT (Multimodal Visual Geometry Grounded Transformer), an end-to-end dual-branch framework integrating language into sparse-view geometric reasoning. Introduces PVSO (Per-view No-target Suppression Optimization) to address Foreground Gradient Dilution (FGD) in sparse 3D supervision.

Result: MVGGT establishes the first strong baseline for MV-3DRES, achieving both high accuracy and fast inference while outperforming existing alternatives. The MVRefer benchmark provides standardized evaluation settings and metrics.

Conclusion: The proposed MVGGT framework with PVSO optimization effectively solves the MV-3DRES problem from sparse multi-view images, offering efficient and accurate segmentation while addressing optimization challenges in sparse 3D supervision.

Abstract: Most existing 3D referring expression segmentation (3DRES) methods rely on dense, high-quality point clouds, while real-world agents such as robots and mobile phones operate with only a few sparse RGB views and strict latency constraints. We introduce Multi-view 3D Referring Expression Segmentation (MV-3DRES), where the model must recover scene structure and segment the referred object directly from sparse multi-view images. Traditional two-stage pipelines, which first reconstruct a point cloud and then perform segmentation, often yield low-quality geometry, produce coarse or degraded target regions, and run slowly. We propose the Multimodal Visual Geometry Grounded Transformer (MVGGT), an efficient end-to-end framework that integrates language information into sparse-view geometric reasoning through a dual-branch design. Training in this setting exposes a critical optimization barrier, termed Foreground Gradient Dilution (FGD), where sparse 3D signals lead to weak supervision. To resolve this, we introduce Per-view No-target Suppression Optimization (PVSO), which provides stronger and more balanced gradients across views, enabling stable and efficient learning. To support consistent evaluation, we build MVRefer, a benchmark that defines standardized settings and metrics for MV-3DRES. Experiments show that MVGGT establishes the first strong baseline and achieves both high accuracy and fast inference, outperforming existing alternatives. Code and models are publicly available at https://mvggt.github.io.

Method: Proposes SAM-RefiSeR framework combining SAM (Segment Anything Model) with refinement and self-training strategies for unsupervised domain adaptation in brain tumor segmentation.

Result: Improved segmentation performance on target domains without requiring labeled target data, demonstrating effective domain adaptation for brain tumor segmentation tasks.

Conclusion: SAM-RefiSeR provides an effective solution for domain adaptation in medical image segmentation, enabling robust brain tumor segmentation across different imaging domains without additional labeling.

Abstract: Unsupervised Domain Adaptation with SAM-RefiSeR for Enhanced Brain Tumor Segmentation

Method: Direct point matching between query and reference images using multi-view information with a lightweight network, plus a reference image fusion strategy that reduces the number of needed reference images.

Result: Achieves comparable results to methods requiring more reference images and larger parameters on seven core BOP challenge datasets, while reducing memory requirements and inference time.

Conclusion: MixRI demonstrates that lightweight networks with efficient reference image strategies can achieve competitive pose estimation performance while meeting real-world application demands for speed and memory efficiency.

Abstract: We present MixRI, a lightweight network that solves the CAD-based novel object pose estimation problem in RGB images. It can be instantly applied to a novel object at test time without finetuning. We design our network to meet the demands of real-world applications, emphasizing reduced memory requirements and fast inference time. Unlike existing works that utilize many reference images and have large network parameters, we directly match points based on the multi-view information between the query and reference images with a lightweight network. Thanks to our reference image fusion strategy, we significantly decrease the number of reference images, thus decreasing the time needed to process these images and the memory required to store them. Furthermore, with our lightweight network, our method requires less inference time. Though with fewer reference images, experiments on seven core datasets in the BOP challenge show that our method achieves comparable results with other methods that require more reference images and larger network parameters.

Method: Replace both vision and text encoders in CLIP with Mamba architecture. Use VMamba for vision encoder to capture visual spatial inductive biases, and autoregressive Mamba for text encoder. The model naturally supports variable input resolutions without positional encoding interpolation or specialized training.

Result: CLIMP surpasses OpenAI’s CLIP-ViT-B by 7.5% on ImageNet-O for out-of-distribution robustness. Achieves up to 6.6% higher retrieval accuracy at 16x training resolution while using 5x less memory and 1.8x fewer FLOPs. The autoregressive text encoder enables dense captioning retrieval, overcoming CLIP’s fixed context limitation.

Conclusion: Mamba exhibits advantageous properties for vision-language learning, making it a compelling alternative to Transformer-based CLIP. The architecture reduces reliance on spurious correlations, improves efficiency, and provides flexibility for variable resolutions and dense captioning tasks.

Abstract: Contrastive Language-Image Pre-training (CLIP) relies on Vision Transformers whose attention mechanism is susceptible to spurious correlations, and scales quadratically with resolution. To address these limitations, We present CLIMP, the first fully Mamba-based contrastive vision-language model that replaces both the vision and text encoders with Mamba. The new architecture encodes sequential structure in both vision and language, with VMamba capturing visual spatial inductive biases, reducing reliance on spurious correlations and producing an embedding space favorable for cross-modal retrieval and out-of-distribution robustness-surpassing OpenAI’s CLIP-ViT-B by 7.5% on ImageNet-O. CLIMP naturally supports variable input resolutions without positional encoding interpolation or specialized training, achieving up to 6.6% higher retrieval accuracy at 16x training resolution while using 5x less memory and 1.8x fewer FLOPs. The autoregressive text encoder further overcomes CLIP’s fixed context limitation, enabling dense captioning retrieval. Our findings suggest that Mamba exhibits advantageous properties for vision-language learning, making it a compelling alternative to Transformer-based CLIP.

Method: UDPNet uses depth priors from DepthAnything V2 pretrained model with two key modules: Depth-Guided Attention Module (DGAM) for adaptive feature modulation via lightweight depth-guided channel attention, and Depth Prior Fusion Module (DPFM) for hierarchical fusion of multi-scale depth map features using dual sliding-window multi-head cross-attention.

Result: Significant performance improvements over state-of-the-art methods: 0.85 dB PSNR improvement on SOTS dataset, 1.19 dB on Haze4K dataset, and 1.79 dB PSNR on NHR dataset. The framework demonstrates robustness across varying haze densities, illumination conditions, and domain gaps.

Conclusion: UDPNet establishes a new benchmark for depth-aware dehazing by effectively integrating depth priors from large-scale pretrained models, offering both computational efficiency and superior performance across various scenarios.

Abstract: Image dehazing has witnessed significant advancements with the development of deep learning models. However, a few methods predominantly focus on single-modal RGB features, neglecting the inherent correlation between scene depth and haze distribution. Even those that jointly optimize depth estimation and image dehazing often suffer from suboptimal performance due to inadequate utilization of accurate depth information. In this paper, we present UDPNet, a general framework that leverages depth-based priors from large-scale pretrained depth estimation model DepthAnything V2 to boost existing image dehazing models. Specifically, our architecture comprises two typical components: the Depth-Guided Attention Module (DGAM) adaptively modulates features via lightweight depth-guided channel attention, and the Depth Prior Fusion Module (DPFM) enables hierarchical fusion of multi-scale depth map features by dual sliding-window multi-head cross-attention mechanism. These modules ensure both computational efficiency and effective integration of depth priors. Moreover, the intrinsic robustness of depth priors empowers the network to dynamically adapt to varying haze densities, illumination conditions, and domain gaps across synthetic and real-world data. Extensive experimental results demonstrate the effectiveness of our UDPNet, outperforming the state-of-the-art methods on popular dehazing datasets, such as 0.85 dB PSNR improvement on the SOTS dataset, 1.19 dB on the Haze4K dataset and 1.79 dB PSNR on the NHR dataset. Our proposed solution establishes a new benchmark for depth-aware dehazing across various scenarios. Pretrained models and codes will be released at our project https://github.com/Harbinzzy/UDPNet.

Method: Proposes RenderFlow, an end-to-end deterministic single-step neural rendering framework based on flow matching paradigm. Includes an efficient module for sparse keyframe guidance to enhance rendering quality and generalization. Also introduces a lightweight adapter-based module for repurposing the forward model for inverse rendering tasks like intrinsic decomposition.

Result: Achieves near real-time performance with photorealistic rendering quality, significantly accelerating the rendering process while enhancing physical plausibility and visual quality through optional sparse keyframe guidance.

Conclusion: RenderFlow effectively bridges the efficiency of modern generative models with the precision of traditional physically based rendering, offering a versatile framework that can be adapted for both forward rendering and inverse rendering tasks.

Abstract: Conventional physically based rendering (PBR) pipelines generate photorealistic images through computationally intensive light transport simulations. Although recent deep learning approaches leverage diffusion model priors with geometry buffers (G-buffers) to produce visually compelling results without explicit scene geometry or light simulation, they remain constrained by two major limitations. First, the iterative nature of the diffusion process introduces substantial latency. Second, the inherent stochasticity of these generative models compromises physical accuracy and temporal consistency. In response to these challenges, we propose a novel, end-to-end, deterministic, single-step neural rendering framework, RenderFlow, built upon a flow matching paradigm. To further strengthen both rendering quality and generalization, we propose an efficient and effective module for sparse keyframe guidance. Our method significantly accelerates the rendering process and, by optionally incorporating sparsely rendered keyframes as guidance, enhances both the physical plausibility and overall visual quality of the output. The resulting pipeline achieves near real-time performance with photorealistic rendering quality, effectively bridging the gap between the efficiency of modern generative models and the precision of traditional physically based rendering. Furthermore, we demonstrate the versatility of our framework by introducing a lightweight, adapter-based module that efficiently repurposes the pretrained forward model for the inverse rendering task of intrinsic decomposition.

Method: Proposes a face-only counterfactual evaluation paradigm: starting from real photographs, generate counterfactual variants by editing only facial attributes related to race and gender while keeping all other visual factors fixed. Creates FOCUS dataset (480 scene-matched counterfactual images across 6 occupations and 10 demographic groups) and REFLECT benchmark with three decision-oriented tasks.

Result: Experiments on five state-of-the-art VLMs show that demographic disparities persist even under strict visual control, and these disparities vary substantially across different task formulations.

Conclusion: Controlled, counterfactual audits are necessary for measuring social bias in multimodal models, and task design is a critical factor in evaluating such bias.

Abstract: Vision-Language Models (VLMs) are increasingly deployed in socially consequential settings, raising concerns about social bias driven by demographic cues. A central challenge in measuring such social bias is attribution under visual confounding: real-world images entangle race and gender with correlated factors such as background and clothing, obscuring attribution. We propose a \textbf{face-only counterfactual evaluation paradigm} that isolates demographic effects while preserving real-image realism. Starting from real photographs, we generate counterfactual variants by editing only facial attributes related to race and gender, keeping all other visual factors fixed. Based on this paradigm, we construct \textbf{FOCUS}, a dataset of 480 scene-matched counterfactual images across six occupations and ten demographic groups, and propose \textbf{REFLECT}, a benchmark comprising three decision-oriented tasks: two-alternative forced choice, multiple-choice socioeconomic inference, and numeric salary recommendation. Experiments on five state-of-the-art VLMs reveal that demographic disparities persist under strict visual control and vary substantially across task formulations. These findings underscore the necessity of controlled, counterfactual audits and highlight task design as a critical factor in evaluating social bias in multimodal models.

Method: Constructed VideoDR benchmark with video-conditioned open-domain QA, requiring cross-frame visual anchor extraction, interactive web retrieval, and multi-hop reasoning over joint video-web evidence. Used rigorous human annotation across six semantic domains.

Result: Evaluated multimodal LLMs under Workflow and Agentic paradigms; found Agentic not consistently superior - gains depend on model’s ability to maintain initial video anchors over long retrieval chains. Goal drift and long-horizon consistency identified as core bottlenecks.

Conclusion: VideoDR provides systematic benchmark for studying video agents in open-web settings and reveals key challenges for next-generation video deep research agents, particularly around maintaining consistency in long reasoning chains.

Abstract: In real-world video question answering scenarios, videos often provide only localized visual cues, while verifiable answers are distributed across the open web; models therefore need to jointly perform cross-frame clue extraction, iterative retrieval, and multi-hop reasoning-based verification. To bridge this gap, we construct the first video deep research benchmark, VideoDR. VideoDR centers on video-conditioned open-domain video question answering, requiring cross-frame visual anchor extraction, interactive web retrieval, and multi-hop reasoning over joint video-web evidence; through rigorous human annotation and quality control, we obtain high-quality video deep research samples spanning six semantic domains. We evaluate multiple closed-source and open-source multimodal large language models under both the Workflow and Agentic paradigms, and the results show that Agentic is not consistently superior to Workflow: its gains depend on a model’s ability to maintain the initial video anchors over long retrieval chains. Further analysis indicates that goal drift and long-horizon consistency are the core bottlenecks. In sum, VideoDR provides a systematic benchmark for studying video agents in open-web settings and reveals the key challenges for next-generation video deep research agents.

Method: The authors introduce SketchJudge, a benchmark with 1,015 hand-drawn student responses across four STEM domains (geometry, physics, charts, flowcharts) featuring diverse stylistic variations and distinct error types. They evaluate advanced MLLMs on this benchmark.

Result: Evaluations show that even advanced MLLMs significantly lag behind humans in grading hand-drawn diagrams, validating SketchJudge’s effectiveness in exposing the fragility of current vision-language alignment in symbolic and noisy contexts.

Conclusion: SketchJudge successfully reveals critical limitations in MLLMs’ ability to handle unstructured sketches and complex reasoning required for visual grading, highlighting the need for improved vision-language alignment in symbolic and noisy contexts.

Abstract: While Multimodal Large Language Models (MLLMs) have achieved remarkable progress in visual understanding, they often struggle when faced with the unstructured and ambiguous nature of human-generated sketches. This limitation is particularly pronounced in the underexplored task of visual grading, where models should not only solve a problem but also diagnose errors in hand-drawn diagrams. Such diagnostic capabilities depend on complex structural, semantic, and metacognitive reasoning. To bridge this gap, we introduce SketchJudge, a novel benchmark tailored for evaluating MLLMs as graders of hand-drawn STEM diagrams. SketchJudge encompasses 1,015 hand-drawn student responses across four domains: geometry, physics, charts, and flowcharts, featuring diverse stylistic variations and distinct error types. Evaluations on SketchJudge demonstrate that even advanced MLLMs lag significantly behind humans, validating the benchmark’s effectiveness in exposing the fragility of current vision-language alignment in symbolic and noisy contexts. All data, code, and evaluation scripts are publicly available at https://github.com/yuhangsu82/SketchJudge.

Method: OmniPersona introduces structurally decoupled concept tokens that allocate dedicated subspaces for different tasks (understanding, generation, editing) to minimize interference. It incorporates an explicit knowledge replay mechanism that propagates personalized attribute knowledge across tasks for consistent behavior. The framework is end-to-end and unified.

Result: The authors propose OmniPBench, an evaluation benchmark extending UnifyBench with personalized editing tasks and cross-task evaluation protocols. Experimental results show OmniPersona delivers competitive and robust performance across diverse personalization tasks.

Conclusion: OmniPersona successfully integrates personalized understanding, generation, and editing within a single architecture, providing a strong baseline for controllable, unified personalization in multimodal models.

Abstract: Unified large multimodal models (LMMs) have achieved remarkable progress in general-purpose multimodal understanding and generation. However, they still operate under a ``one-size-fits-all’’ paradigm and struggle to model user-specific concepts (e.g., generate a photo of \texttt{}) in a consistent and controllable manner. Existing personalization methods typically rely on external retrieval, which is inefficient and poorly integrated into unified multimodal pipelines. Recent personalized unified models introduce learnable soft prompts to encode concept information, yet they either couple understanding and generation or depend on complex multi-stage training, leading to cross-task interference and ultimately to fuzzy or misaligned personalized knowledge. We present \textbf{OmniPersona}, an end-to-end personalization framework for unified LMMs that, for the first time, integrates personalized understanding, generation, and image editing within a single architecture. OmniPersona introduces structurally decoupled concept tokens, allocating dedicated subspaces for different tasks to minimize interference, and incorporates an explicit knowledge replay mechanism that propagates personalized attribute knowledge across tasks, enabling consistent personalized behavior. To systematically evaluate unified personalization, we propose \textbf{\texttt{OmniPBench}}, extending the public UnifyBench concept set with personalized editing tasks and cross-task evaluation protocols integrating understanding, generation, and editing. Experimental results demonstrate that OmniPersona delivers competitive and robust performance across diverse personalization tasks. We hope OmniPersona will serve as a strong baseline and spur further research on controllable, unified personalization.

Method: Systematically examined CoT’s role in FGVC through zero-shot evaluation and multiple training paradigms. Developed \alg (normalization method for multi-reward optimization) and ReFine-RFT framework combining ensemble rewards with \alg to constrain reasoning length while providing dense accuracy feedback.

Result: Identified “Cost of Thinking” phenomenon: longer textual reasoning consistently lowers classification accuracy. ReFine-RFT achieves state-of-the-art performance across FGVC benchmarks.

Conclusion: CoT’s degradation in FGVC is driven by reasoning length. The proposed ReFine-RFT framework effectively addresses this by balancing reward signals and constraining reasoning length, enabling better FGVC performance.

Abstract: Multi-modal large language models (MLLMs) exhibit strong general-purpose capabilities, yet still struggle on Fine-Grained Visual Classification (FGVC), a core perception task that requires subtle visual discrimination and is crucial for many real-world applications. A widely adopted strategy for boosting performance on challenging tasks such as math and coding is Chain-of-Thought (CoT) reasoning. However, several prior works have reported that CoT can actually harm performance on visual perception tasks. These studies, though, examine the issue from relatively narrow angles and leave open why CoT degrades perception-heavy performance. We systematically re-examine the role of CoT in FGVC through the lenses of zero-shot evaluation and multiple training paradigms. Across these settings, we uncover a central paradox: the degradation induced by CoT is largely driven by the reasoning length, in which longer textual reasoning consistently lowers classification accuracy. We term this phenomenon the ``Cost of Thinking’’. Building on this finding, we make two key contributions: (1) \alg, a simple and general plug-and-play normalization method for multi-reward optimization that balances heterogeneous reward signals, and (2) ReFine-RFT, a framework that combines ensemble rewards with \alg to constrain reasoning length while providing dense accuracy-oriented feedback. Extensive experiments demonstrate the effectiveness of our findings and the proposed ReFine-RFT, achieving state-of-the-art performance across FGVC benchmarks. Code and models are available at \href{https://github.com/jiezhu23/ReFine-RFT}{Project Link}.

Method: A spatial multi-task learning framework that simultaneously predicts ER, PR, HER2 status, and Ki-67 index from single-phase DCE-MRI. The architecture integrates deep feature extraction with multi-scale spatial attention to capture intratumoral and peritumoral characteristics, plus a region-of-interest weighting module emphasizing tumor core, rim, and surrounding tissue. Multi-task learning exploits biological correlations among biomarkers through shared representations with task-specific branches.

Result: On 960 cases (886 internal split 7:1:2, 74 external with five-fold cross-validation), the method achieved AUCs of 0.893 (ER), 0.824 (PR), and 0.857 (HER2), and mean absolute error of 8.2% for Ki-67 regression. Significantly outperformed radiomics and single-task deep learning baselines.

Conclusion: Demonstrates feasibility of accurate, non-invasive molecular subtype prediction using standard single-phase DCE-MRI protocols, potentially enabling personalized breast cancer treatment without invasive biopsies.

Abstract: Accurate molecular subtype classification is essential for personalized breast cancer treatment, yet conventional immunohistochemical analysis relies on invasive biopsies and is prone to sampling bias. Although dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) enables non-invasive tumor characterization, clinical workflows typically acquire only single-phase post-contrast images to reduce scan time and contrast agent dose. In this study, we propose a spatial multi-task learning framework for breast cancer molecular subtype prediction from clinically practical single-phase DCE-MRI. The framework simultaneously predicts estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2) status, and the Ki-67 proliferation index – biomarkers that collectively define molecular subtypes. The architecture integrates a deep feature extraction network with multi-scale spatial attention to capture intratumoral and peritumoral characteristics, together with a region-of-interest weighting module that emphasizes the tumor core, rim, and surrounding tissue. Multi-task learning exploits biological correlations among biomarkers through shared representations with task-specific prediction branches. Experiments on a dataset of 960 cases (886 internal cases split 7:1:2 for training/validation/testing, and 74 external cases evaluated via five-fold cross-validation) demonstrate that the proposed method achieves an AUC of 0.893, 0.824, and 0.857 for ER, PR, and HER2 classification, respectively, and a mean absolute error of 8.2% for Ki-67 regression, significantly outperforming radiomics and single-task deep learning baselines. These results indicate the feasibility of accurate, non-invasive molecular subtype prediction using standard imaging protocols.

Method: Proposes a targeted adversarial attack framework with two components: 1) Local Pixel Dependency Attack that exploits spatial correlations among neighboring pixels, and 2) Multiscale Information Attack that perturbs features across hierarchical spectral-spatial scales.

Result: Experiments on Brain and MDC datasets show the attacks significantly degrade classification performance, especially in tumor regions, while remaining visually imperceptible. Outperforms existing methods in revealing unique vulnerabilities of medical HSI models.

Conclusion: The work reveals specific vulnerabilities in medical HSI models and underscores the need for robust, structure-aware defenses in clinical applications to ensure reliable disease diagnosis.

Abstract: Medical hyperspectral imaging (HSI) enables accurate disease diagnosis by capturing rich spectral-spatial tissue information, but recent advances in deep learning have exposed its vulnerability to adversarial attacks. In this work, we identify two fundamental causes of this fragility: the reliance on local pixel dependencies for preserving tissue structure and the dependence on multiscale spectral-spatial representations for hierarchical feature encoding. Building on these insights, we propose a targeted adversarial attack framework for medical HSI, consisting of a Local Pixel Dependency Attack that exploits spatial correlations among neighboring pixels, and a Multiscale Information Attack that perturbs features across hierarchical spectral-spatial scales. Experiments on the Brain and MDC datasets demonstrate that our attacks significantly degrade classification performance, especially in tumor regions, while remaining visually imperceptible. Compared with existing methods, our approach reveals the unique vulnerabilities of medical HSI models and underscores the need for robust, structure-aware defenses in clinical applications.

Method: Two-stage pipeline: (1) YOLO-based object detection trained on Mapillary Vistas and fine-tuned on BillboardLamac for billboard detection (94% mAP@50), (2) classifier using bounding box positions and DINOv2 features to estimate gaze duration from individual frames.

Result: Achieved 68.1% accuracy on BillboardLamac dataset for individual frame gaze estimation, with additional validation using Google Street View images.

Conclusion: The proposed automated pipeline successfully detects billboards and estimates driver gaze duration, providing a scalable alternative to manual methods for evaluating billboard relevance and potential distraction.

Abstract: Roadside billboards represent a central element of outdoor advertising, yet their presence may contribute to driver distraction and accident risk. This study introduces a fully automated pipeline for billboard detection and driver gaze duration estimation, aiming to evaluate billboard relevance without reliance on manual annotations or eye-tracking devices. Our pipeline operates in two stages: (1) a YOLO-based object detection model trained on Mapillary Vistas and fine-tuned on BillboardLamac images achieved 94% mAP@50 in the billboard detection task (2) a classifier based on the detected bounding box positions and DINOv2 features. The proposed pipeline enables estimation of billboard driver gaze duration from individual frames. We show that our method is able to achieve 68.1% accuracy on BillboardLamac when considering individual frames. These results are further validated using images collected from Google Street View.

Method: Proposes SRC-Pipeline, an efficient VLM framework that learns to compress early frame tokens into a small number of high-level tokens while retaining full patch tokens for recent frames. This selective compression reduces computational load while maintaining important recent visual information.

Result: Achieves 66% FLOPs reduction while maintaining comparable performance on autonomous driving video question answering tasks, enabling VLMs to operate effectively in real-time safety-critical settings.

Conclusion: SRC-Pipeline successfully addresses the efficiency bottleneck in autonomous driving VQA by balancing computational efficiency with performance, making large VLMs practical for real-time deployment in safety-critical autonomous driving applications.

Abstract: Autonomous driving increasingly relies on Visual Question Answering (VQA) to enable vehicles to understand complex surroundings by analyzing visual inputs and textual queries. Currently, a paramount concern for VQA in this domain is the stringent requirement for fast latency and real-time processing, as delays directly impact real-world safety in this safety-critical application. However, current state-of-the-art VQA models, particularly large vision-language models (VLMs), often prioritize performance over computational efficiency. These models typically process dense patch tokens for every frame, leading to prohibitive computational costs (FLOPs) and significant inference latency, especially with long video sequences. This focus limits their practical deployment in real-time autonomous driving scenarios. To tackle this issue, we propose an efficient VLM framework for autonomous driving VQA tasks, SRC-Pipeline. It learns to compress early frame tokens into a small number of high-level tokens while retaining full patch tokens for recent frames. Experiments on autonomous driving video question answering tasks show that our approach achieves 66% FLOPs reduction while maintaining comparable performance, enabling VLMs to operate more effectively in real-time, safety-critical autonomous driving settings.

Method: Proposes Wavelet-Conditioned ControlNet (WCC-Net), a fully 3D diffusion-based framework that introduces explicit frequency-domain structural priors via wavelet representations. It injects wavelet-based structural guidance into a frozen pretrained diffusion backbone through a lightweight control branch, decoupling anatomical structure from noise while preserving generative expressiveness and 3D structural continuity.

Result: WCC-Net consistently outperforms CNN-, GAN-, and diffusion-based baselines. On internal 1/20-dose test set, it improves PSNR by +1.21 dB and SSIM by +0.008 over strong diffusion baseline, while reducing structural distortion (GMSD) and intensity error (NMAE). Generalizes robustly to unseen dose levels (1/50 and 1/4), achieving superior quantitative performance and improved volumetric anatomical consistency.

Conclusion: WCC-Net effectively addresses the challenge of anatomical consistency in diffusion-based PET denoising by incorporating wavelet-based structural guidance, making it a promising solution for low-dose PET imaging with improved image quality and diagnostic reliability.

Abstract: Low-dose Positron Emission Tomography (PET) imaging reduces patient radiation exposure but suffers from increased noise that degrades image quality and diagnostic reliability. Although diffusion models have demonstrated strong denoising capability, their stochastic nature makes it challenging to enforce anatomically consistent structures, particularly in low signal-to-noise regimes and volumetric whole-body imaging. We propose Wavelet-Conditioned ControlNet (WCC-Net), a fully 3D diffusion-based framework that introduces explicit frequency-domain structural priors via wavelet representations to guide volumetric PET denoising. By injecting wavelet-based structural guidance into a frozen pretrained diffusion backbone through a lightweight control branch, WCC-Net decouples anatomical structure from noise while preserving generative expressiveness and 3D structural continuity. Extensive experiments demonstrate that WCC-Net consistently outperforms CNN-, GAN-, and diffusion-based baselines. On the internal 1/20-dose test set, WCC-Net improves PSNR by +1.21 dB and SSIM by +0.008 over a strong diffusion baseline, while reducing structural distortion (GMSD) and intensity error (NMAE). Moreover, WCC-Net generalizes robustly to unseen dose levels (1/50 and 1/4), achieving superior quantitative performance and improved volumetric anatomical consistency.

Method: Introduces MedVistaGym, a scalable interactive training environment that incentivizes tool-integrated visual reasoning. It trains VLMs to determine when and which tools to invoke, localize relevant image regions, and integrate evidence through trajectory sampling and end-to-end reinforcement learning.

Result: MedVistaGym-R1-8B exceeds comparably sized tool-augmented baselines by 19.10% to 24.21% across six medical VQA benchmarks, demonstrating that structured agentic training (not just tool access) enables effective tool-integrated reasoning.

Conclusion: Structured agentic training through MedVistaGym unlocks effective tool-integrated reasoning for medical image analysis, addressing the limitations of current medical VLMs in multi-step reasoning and iterative visual interaction.

Abstract: Vision language models (VLMs) achieve strong performance on general image understanding but struggle to think with medical images, especially when performing multi-step reasoning through iterative visual interaction. Medical VLMs often rely on static visual embeddings and single-pass inference, preventing models from re-examining, verifying, or refining visual evidence during reasoning. While tool-integrated reasoning offers a promising path forward, open-source VLMs lack the training infrastructure to learn effective tool selection, invocation, and coordination in multi-modal medical reasoning. We introduce MedVistaGym, a scalable and interactive training environment that incentivizes tool-integrated visual reasoning for medical image analysis. MedVistaGym equips VLMs to determine when and which tools to invoke, localize task-relevant image regions, and integrate single or multiple sub-image evidence into interleaved multimodal reasoning within a unified, executable interface for agentic training. Using MedVistaGym, we train MedVistaGym-R1 to interleave tool use with agentic reasoning through trajectory sampling and end-to-end reinforcement learning. Across six medical VQA benchmarks, MedVistaGym-R1-8B exceeds comparably sized tool-augmented baselines by 19.10% to 24.21%, demonstrating that structured agentic training–not tool access alone–unlocks effective tool-integrated reasoning for medical image analysis.

Method: 1) Base learning with generative domain adaptation finetuning using MAE decoder for reconstruction and classifier for classification. 2) Construct two class-wise memories: representation memory (mean features) and weight memory (classifier weights). 3) Memory-regularized incremental learning with co-memory regularization that updates memories incrementally and collaboratively regularizes learning.

Result: Extensive experiments on popular benchmarks demonstrate that the approach outperforms state-of-the-art methods in recognition accuracy while effectively mitigating catastrophic forgetting and overfitting.

Conclusion: The generative co-memory regularization approach successfully addresses FSCIL challenges by combining generative domain adaptation with collaborative memory regularization, achieving superior performance compared to existing methods.

Abstract: Few-shot class-incremental learning (FSCIL) aims to incrementally learn models from a small amount of novel data, which requires strong representation and adaptation ability of models learned under few-example supervision to avoid catastrophic forgetting on old classes and overfitting to novel classes. This work proposes a generative co-memory regularization approach to facilitate FSCIL. In the approach, the base learning leverages generative domain adaptation finetuning to finetune a pretrained generative encoder on a few examples of base classes by jointly incorporating a masked autoencoder (MAE) decoder for feature reconstruction and a fully-connected classifier for feature classification, which enables the model to efficiently capture general and adaptable representations. Using the finetuned encoder and learned classifier, we construct two class-wise memories: representation memory for storing the mean features for each class, and weight memory for storing the classifier weights. After that, the memory-regularized incremental learning is performed to train the classifier dynamically on the examples of few-shot classes in each incremental session by simultaneously optimizing feature classification and co-memory regularization. The memories are updated in a class-incremental manner and they collaboratively regularize the incremental learning. In this way, the learned models improve recognition accuracy, while mitigating catastrophic forgetting over old classes and overfitting to novel classes. Extensive experiments on popular benchmarks clearly demonstrate that our approach outperforms the state-of-the-arts.

Method: Leverages foundation model for camera pose estimation, introduces system-level optimizations for efficient inference, uses sliding batch inference strategy for real-time performance with manageable memory consumption.

Result: Achieves real-time performance on collected egocentric action dataset with manageable memory consumption, making motion-centric analysis practical for edge deployment.

Conclusion: Provides a complementary perspective to existing egocentric studies on sports and fast-movement activities by making motion-centric analysis practical for real-time edge deployment.

Abstract: From Vision-Language-Action (VLA) systems to robotics, existing egocentric datasets primarily focus on action recognition tasks, while largely overlooking the inherent role of motion analysis in sports and other fast-movement scenarios. To bridge this gap, we propose a real-time motion focus recognition method that estimates the subject’s locomotion intention from any egocentric video. Our approach leverages the foundation model for camera pose estimation and introduces system-level optimizations to enable efficient and scalable inference. Evaluated on a collected egocentric action dataset, our method achieves real-time performance with manageable memory consumption through a sliding batch inference strategy. This work makes motion-centric analysis practical for edge deployment and offers a complementary perspective to existing egocentric studies on sports and fast-movement activities.

Method: TAHCD uses Adaptive Stable Subspace Alignment and Sample-Adaptive Confidence Alignment to remove heterogeneous noise at global and instance levels. It also employs test-time cooperative enhancement to adaptively update the model in response to input noise without labels.

Result: Experiments on multiple benchmarks show superior classification performance, robustness, and generalization compared to state-of-the-art reliable multimodal learning approaches.

Conclusion: TAHCD effectively addresses multimodal noise challenges by combining hierarchical denoising with test-time adaptation, achieving reliable multimodal learning in noisy environments.

Abstract: Reliable learning on low-quality multimodal data is a widely concerning issue, especially in safety-critical applications. However, multimodal noise poses a major challenge in this domain and leads existing methods to suffer from two key limitations. First, they struggle to reliably remove heterogeneous data noise, hindering robust multimodal representation learning. Second, they exhibit limited adaptability and generalization when encountering previously unseen noise. To address these issues, we propose Test-time Adaptive Hierarchical Co-enhanced Denoising Network (TAHCD). On one hand, TAHCD introduces the Adaptive Stable Subspace Alignment and Sample-Adaptive Confidence Alignment to reliably remove heterogeneous noise. They account for noise at both global and instance levels and enable jointly removal of modality-specific and cross-modality noise, achieving robust learning. On the other hand, TAHCD introduces test-time cooperative enhancement, which adaptively updates the model in response to input noise in a label-free manner, improving adaptability and generalization. This is achieved by collaboratively enhancing the joint removal process of modality-specific and cross-modality noise across global and instance levels according to sample noise. Experiments on multiple benchmarks demonstrate that the proposed method achieves superior classification performance, robustness, and generalization compared with state-of-the-art reliable multimodal learning approaches.

Method: DIVER uses a progressive evidence-driven reasoning paradigm: 1) establishes text-based baseline using language analysis and intra-modal consistency filtering, 2) introduces visual information only when textual evidence is insufficient, 3) uses inter-modal alignment verification to decide if deeper visual inspection is needed, 4) selectively invokes fine-grained visual tools (OCR, dense captioning) for cross-modal discrepancies, and 5) iteratively aggregates evidence via uncertainty-aware fusion.

Result: Outperforms state-of-the-art baselines by average 2.72% on Weibo, Weibo21, and GossipCop datasets, while reducing inference latency by 4.12 seconds.

Conclusion: DIVER provides an effective and efficient solution for multimodal fake news detection through dynamic, evidence-driven reasoning that minimizes unnecessary visual processing while maintaining strong detection performance.

Abstract: Multimodal fake news detection is crucial for mitigating adversarial misinformation. Existing methods, relying on static fusion or LLMs, face computational redundancy and hallucination risks due to weak visual foundations. To address this, we propose DIVER (Dynamic Iterative Visual Evidence Reasoning), a framework grounded in a progressive, evidence-driven reasoning paradigm. DIVER first establishes a strong text-based baseline through language analysis, leveraging intra-modal consistency to filter unreliable or hallucinated claims. Only when textual evidence is insufficient does the framework introduce visual information, where inter-modal alignment verification adaptively determines whether deeper visual inspection is necessary. For samples exhibiting significant cross-modal semantic discrepancies, DIVER selectively invokes fine-grained visual tools (e.g., OCR and dense captioning) to extract task-relevant evidence, which is iteratively aggregated via uncertainty-aware fusion to refine multimodal reasoning. Experiments on Weibo, Weibo21, and GossipCop demonstrate that DIVER outperforms state-of-the-art baselines by an average of 2.72%, while optimizing inference efficiency with a reduced latency of 4.12 s.

Method: Four-component pipeline: 1) Recorder captures screen video with precise user interactions; 2) Learner semantically interprets raw interactions and visual context into natural language captions; 3) Planner reads parsed demonstrations, maintains task states, and formulates next action plans; 4) Executor carries out actions at OS level with safety checks and real-time feedback.

Result: The framework provides a scalable solution for collecting and parsing real-world human data, demonstrating a viable path toward building general-purpose GUI agents that can learn from human observations.

Conclusion: ShowUI-Aloha enables effective learning for GUI agents by transforming unstructured human demonstrations into structured, actionable tasks, addressing the data scarcity problem in GUI automation.

Abstract: Graphical User Interfaces (GUIs) are central to human-computer interaction, yet automating complex GUI tasks remains a major challenge for autonomous agents, largely due to a lack of scalable, high-quality training data. While recordings of human demonstrations offer a rich data source, they are typically long, unstructured, and lack annotations, making them difficult for agents to learn from.To address this, we introduce ShowUI-Aloha, a comprehensive pipeline that transforms unstructured, in-the-wild human screen recordings from desktop environments into structured, actionable tasks. Our framework includes four key components: A recorder that captures screen video along with precise user interactions like mouse clicks, keystrokes, and scrolls. A learner that semantically interprets these raw interactions and the surrounding visual context, translating them into descriptive natural language captions. A planner that reads the parsed demonstrations, maintains task states, and dynamically formulates the next high-level action plan based on contextual reasoning. An executor that faithfully carries out these action plans at the OS level, performing precise clicks, drags, text inputs, and window operations with safety checks and real-time feedback. Together, these components provide a scalable solution for collecting and parsing real-world human data, demonstrating a viable path toward building general-purpose GUI agents that can learn effectively from simply observing humans.

Method: 1) Generate synthetic dataset by path-tracing 3D glass models over real backgrounds with varied properties; 2) Fine-tune LMM using concatenated image layers with joint captioning and task-specific LoRA instead of full-parameter training.

Result: Achieves improved reflection removal and separation performance compared to state-of-the-art methods.

Conclusion: Combining physically accurate synthetic data generation with efficient LMM fine-tuning enables better single-image reflection removal.

Abstract: Glass surfaces create complex interactions of reflected and transmitted light, making single-image reflection removal (SIRR) challenging. Existing datasets suffer from limited physical realism in synthetic data or insufficient scale in real captures. We introduce a synthetic dataset generation framework that path-traces 3D glass models over real background imagery to create physically accurate reflection scenarios with varied glass properties, camera settings, and post-processing effects. To leverage the capabilities of Large Multimodal Model (LMM), we concatenate the image layers into a single composite input, apply joint captioning, and fine-tune the model using task-specific LoRA rather than full-parameter training. This enables our approach to achieve improved reflection removal and separation performance compared to state-of-the-art methods.

Method: SceneNAT uses a single-stage masked non-autoregressive Transformer trained via masked modeling over discretized semantic and spatial attributes. It employs dual-level masking (attribute and instance) to capture intra/inter-object structure and includes a dedicated triplet predictor for relational reasoning using learnable relation queries mapped to symbolic triplets (subject, predicate, object).

Result: On the 3D-FRONT dataset, SceneNAT achieves superior performance compared to state-of-the-art autoregressive and diffusion baselines in both semantic compliance and spatial arrangement accuracy, while operating with substantially lower computational cost.

Conclusion: SceneNAT demonstrates that masked non-autoregressive Transformers can effectively generate 3D scenes from text with improved efficiency and performance, offering a promising direction for scalable 3D scene synthesis.

Abstract: We present SceneNAT, a single-stage masked non-autoregressive Transformer that synthesizes complete 3D indoor scenes from natural language instructions through only a few parallel decoding passes, offering improved performance and efficiency compared to prior state-of-the-art approaches. SceneNAT is trained via masked modeling over fully discretized representations of both semantic and spatial attributes. By applying a masking strategy at both the attribute level and the instance level, the model can better capture intra-object and inter-object structure. To boost relational reasoning, SceneNAT employs a dedicated triplet predictor for modeling the scene’s layout and object relationships by mapping a set of learnable relation queries to a sparse set of symbolic triplets (subject, predicate, object). Extensive experiments on the 3D-FRONT dataset demonstrate that SceneNAT achieves superior performance compared to state-of-the-art autoregressive and diffusion baselines in both semantic compliance and spatial arrangement accuracy, while operating with substantially lower computational cost.

Method: VENUS employs a split prompt conditioning strategy to disentangle target objects from background context, uses noise inversion to preserve unedited regions, and integrates scene graphs from multimodal LLMs with diffusion backbones without additional training.

Result: VENUS significantly improves background preservation (PSNR: 24.80 vs 22.45, SSIM: 0.84 vs 0.79, LPIPS: 0.070 vs 0.100) and semantic consistency (CLIP: 24.97 vs 24.19) on PIE-Bench, while reducing runtime from 6-10 minutes to 20-30 seconds per image.

Conclusion: VENUS provides an efficient, training-free solution for scene graph-guided image editing that outperforms both scene graph-based and text-based editing methods in preserving background fidelity while achieving better semantic alignment.

Abstract: State-of-the-art text-based image editing models often struggle to balance background preservation with semantic consistency, frequently resulting either in the synthesis of entirely new images or in outputs that fail to realize the intended edits. In contrast, scene graph-based image editing addresses this limitation by providing a structured representation of semantic entities and their relations, thereby offering improved controllability. However, existing scene graph editing methods typically depend on model fine-tuning, which incurs high computational cost and limits scalability. To this end, we introduce VENUS (Visual Editing with Noise inversion Using Scene graphs), a training-free framework for scene graph-guided image editing. Specifically, VENUS employs a split prompt conditioning strategy that disentangles the target object of the edit from its background context, while simultaneously leveraging noise inversion to preserve fidelity in unedited regions. Moreover, our proposed approach integrates scene graphs extracted from multimodal large language models with diffusion backbones, without requiring any additional training. Empirically, VENUS substantially improves both background preservation and semantic alignment on PIE-Bench, increasing PSNR from 22.45 to 24.80, SSIM from 0.79 to 0.84, and reducing LPIPS from 0.100 to 0.070 relative to the state-of-the-art scene graph editing model (SGEdit). In addition, VENUS enhances semantic consistency as measured by CLIP similarity (24.97 vs. 24.19). On EditVal, VENUS achieves the highest fidelity with a 0.87 DINO score and, crucially, reduces per-image runtime from 6-10 minutes to only 20-30 seconds. Beyond scene graph-based editing, VENUS also surpasses strong text-based editing baselines such as LEDIT++ and P2P+DirInv, thereby demonstrating consistent improvements across both paradigms.

Method: LACE consists of two components: (1) GLIP-Adapter that fuses global semantics with local structural features for consistency preservation, and (2) Multi-Domain Control Guidance that grounds semantic differences between source and target prompts into per-attribute translation vectors.

Result: Experiments on CelebA(Dialog) and BDD100K show LACE achieves high visual fidelity, structural preservation, and interpretable domain-specific control, surpassing prior baselines.

Conclusion: LACE serves as a cross-modal content generation framework that effectively bridges language semantics with controllable visual translation for multi-domain image transformation.

Abstract: Multi-domain image-to-image translation re quires grounding semantic differences ex pressed in natural language prompts into corresponding visual transformations, while preserving unrelated structural and seman tic content. Existing methods struggle to maintain structural integrity and provide fine grained, attribute-specific control, especially when multiple domains are involved. We propose LACE (Language-grounded Attribute Controllable Translation), built on two compo nents: (1) a GLIP-Adapter that fuses global semantics with local structural features to pre serve consistency, and (2) a Multi-Domain Control Guidance mechanism that explicitly grounds the semantic delta between source and target prompts into per-attribute translation vec tors, aligning linguistic semantics with domain level visual changes. Together, these modules enable compositional multi-domain control with independent strength modulation for each attribute. Experiments on CelebA(Dialog) and BDD100K demonstrate that LACE achieves high visual fidelity, structural preservation, and interpretable domain-specific control, surpass ing prior baselines. This positions LACE as a cross-modal content generation framework bridging language semantics and controllable visual translation.

Method: UDAP leverages distinct reconstruction behaviors of clean vs adversarial images during DDIM inversion, minimizes DDIM metric loss to remove adversarial noise, and uses dynamic epoch adjustment strategy that adapts optimization iterations based on reconstruction errors.

Result: UDAP demonstrates robustness against diverse adversarial methods (PID, Anti-DreamBooth, MIST, Anti-DF, MetaCloak), generalizes well across SD versions and text prompts, and improves efficiency without sacrificing purification quality.

Conclusion: UDAP effectively addresses the security gap in Stable Diffusion models by providing a specialized adversarial purification framework that is practical for real-world applications and robust against various SD-specific attacks.

Abstract: Stable Diffusion (SD) often produces degraded outputs when the training dataset contains adversarial noise. Adversarial purification offers a promising solution by removing adversarial noise from contaminated data. However, existing purification methods are primarily designed for classification tasks and fail to address SD-specific adversarial strategies, such as attacks targeting the VAE encoder, UNet denoiser, or both. To address the gap in SD security, we propose Universal Diffusion Adversarial Purification (UDAP), a novel framework tailored for defending adversarial attacks targeting SD models. UDAP leverages the distinct reconstruction behaviors of clean and adversarial images during Denoising Diffusion Implicit Models (DDIM) inversion to optimize the purification process. By minimizing the DDIM metric loss, UDAP can effectively remove adversarial noise. Additionally, we introduce a dynamic epoch adjustment strategy that adapts optimization iterations based on reconstruction errors, significantly improving efficiency without sacrificing purification quality. Experiments demonstrate UDAP’s robustness against diverse adversarial methods, including PID (VAE-targeted), Anti-DreamBooth (UNet-targeted), MIST (hybrid), and robustness-enhanced variants like Anti-Diffusion (Anti-DF) and MetaCloak. UDAP also generalizes well across SD versions and text prompts, showcasing its practical applicability in real-world scenarios.

Method: Compared two AE representations (principal components and full 64-band embeddings) with conventional LCFs across three study areas using three deep learning models (CNN1D, CNN2D, Vision Transformer) with multiple evaluation metrics.

Result: AE-based models consistently outperformed LCFs with 4-15% higher F1-scores and 0.04-0.11 higher AUC values, showing clearer spatial correspondence with observed landslides.

Conclusion: AE embeddings show strong potential as standardized, information-rich alternatives to conventional LCFs for landslide susceptibility mapping, with performance improvements linked to temporal alignment with landslide inventories.

Abstract: Data-driven landslide susceptibility mapping (LSM) typically relies on landslide conditioning factors (LCFs), whose availability, heterogeneity, and preprocessing-related uncertainties can constrain mapping reliability. Recently, Google AlphaEarth (AE) embeddings, derived from multi-source geospatial observations, have emerged as a unified representation of Earth surface conditions. This study evaluated the potential of AE embeddings as alternative predictors for LSM. Two AE representations, including retained principal components and the full set of 64 embedding bands, were systematically compared with conventional LCFs across three study areas (Nantou County, Taiwan; Hong Kong; and part of Emilia-Romagna, Italy) using three deep learning models (CNN1D, CNN2D, and Vision Transformer). Performance was assessed using multiple evaluation metrics, ROC-AUC analysis, error statistics, and spatial pattern assessment. Results showed that AE-based models consistently outperformed LCFs across all regions and models, yielding higher F1-scores, AUC values, and more stable error distributions. Such improvement was most pronounced when using the full 64-band AE representation, with F1-score improvements of approximately 4% to 15% and AUC increased ranging from 0.04 to 0.11, depending on the study area and model. AE-based susceptibility maps also exhibited clearer spatial correspondence with observed landslide occurrences and enhanced sensitivity to localised landslide-prone conditions. Performance improvements were more evident in Nantou and Emilia than in Hong Kong, revealing that closer temporal alignment between AE embeddings and landslide inventories may lead to more effective LSM outcomes. These findings highlight the strong potential of AE embeddings as a standardised and information-rich alternative to conventional LCFs for LSM.

Method: PALUM learns common motion representations across diverse skeleton topologies by: 1) partitioning joints into semantic body parts, 2) applying attention mechanisms to capture spatio-temporal relationships, 3) leveraging skeleton-agnostic representations with target-specific structural information, and 4) introducing a cycle consistency mechanism to maintain semantic coherence.

Result: Extensive experiments demonstrate superior performance in handling diverse skeletal structures while maintaining motion realism and semantic fidelity. The method generalizes well to previously unseen skeleton-motion combinations.

Conclusion: PALUM provides an effective solution for motion retargeting across characters with different skeleton structures, addressing the fundamental challenge of maintaining motion semantics and quality. The approach will be made publicly available to support future research.

Abstract: Retargeting motion between characters with different skeleton structures is a fundamental challenge in computer animation. When source and target characters have vastly different bone arrangements, maintaining the original motion’s semantics and quality becomes increasingly difficult. We present PALUM, a novel approach that learns common motion representations across diverse skeleton topologies by partitioning joints into semantic body parts and applying attention mechanisms to capture spatio-temporal relationships. Our method transfers motion to target skeletons by leveraging these skeleton-agnostic representations alongside target-specific structural information. To ensure robust learning and preserve motion fidelity, we introduce a cycle consistency mechanism that maintains semantic coherence throughout the retargeting process. Extensive experiments demonstrate superior performance in handling diverse skeletal structures while maintaining motion realism and semantic fidelity, even when generalizing to previously unseen skeleton-motion combinations. We will make our implementation publicly available to support future research.

Method: Uses a conditional generation architecture built on pre-trained Stable Diffusion, formulating detection as semantic constraints in latent space. Conditions on input images to directly generate bounding boxes with semantic annotations in original image space, enabling precise control over positions and categories.

Result: Achieves competitive accuracy compared to discriminative detectors while retaining the flexibility characteristic of generative methods, demonstrating effective bridging between generative models and discriminative tasks.

Conclusion: GenDet successfully redefines object detection as a generation task, providing a novel methodology that bridges generative and discriminative approaches and offers new perspectives for unified visual understanding systems.

Abstract: This paper presents GenDet, a novel framework that redefines object detection as an image generation task. In contrast to traditional approaches, GenDet adopts a pioneering approach by leveraging generative modeling: it conditions on the input image and directly generates bounding boxes with semantic annotations in the original image space. GenDet establishes a conditional generation architecture built upon the large-scale pre-trained Stable Diffusion model, formulating the detection task as semantic constraints within the latent space. It enables precise control over bounding box positions and category attributes, while preserving the flexibility of the generative model. This novel methodology effectively bridges the gap between generative models and discriminative tasks, providing a fresh perspective for constructing unified visual understanding systems. Systematic experiments demonstrate that GenDet achieves competitive accuracy compared to discriminative detectors, while retaining the flexibility characteristic of generative methods.

Method: Focal Guidance (FG) with two mechanisms: 1) Fine-grained Semantic Guidance (FSG) uses CLIP to identify key regions in reference frames as anchors for Semantic-Weak Layers; 2) Attention Cache transfers attention maps from semantically responsive layers to Semantic-Weak Layers to inject explicit semantic signals.

Result: FG improves instruction following in I2V models, raising total score on Wan2.1-I2V to 0.7250 (+3.97%) and boosting MMDiT-based HunyuanVideo-I2V to 0.5571 (+7.44%). The paper also introduces a benchmark for assessing instruction following in I2V models.

Conclusion: Focal Guidance effectively addresses Condition Isolation in DiT-based I2V models by enhancing controllability of Semantic-Weak Layers, improving adherence to textual instructions while maintaining visual consistency, with demonstrated effectiveness and generalizability across different models.

Abstract: The task of Image-to-Video (I2V) generation aims to synthesize a video from a reference image and a text prompt. This requires diffusion models to reconcile high-frequency visual constraints and low-frequency textual guidance during the denoising process. However, while existing I2V models prioritize visual consistency, how to effectively couple this dual guidance to ensure strong adherence to the text prompt remains underexplored. In this work, we observe that in Diffusion Transformer (DiT)-based I2V models, certain intermediate layers exhibit weak semantic responses (termed Semantic-Weak Layers), as indicated by a measurable drop in text-visual similarity. We attribute this to a phenomenon called Condition Isolation, where attention to visual features becomes partially detached from text guidance and overly relies on learned visual priors. To address this, we propose Focal Guidance (FG), which enhances the controllability from Semantic-Weak Layers. FG comprises two mechanisms: (1) Fine-grained Semantic Guidance (FSG) leverages CLIP to identify key regions in the reference frame and uses them as anchors to guide Semantic-Weak Layers. (2) Attention Cache transfers attention maps from semantically responsive layers to Semantic-Weak Layers, injecting explicit semantic signals and alleviating their over-reliance on the model’s learned visual priors, thereby enhancing adherence to textual instructions. To further validate our approach and address the lack of evaluation in this direction, we introduce a benchmark for assessing instruction following in I2V models. On this benchmark, Focal Guidance proves its effectiveness and generalizability, raising the total score on Wan2.1-I2V to 0.7250 (+3.97%) and boosting the MMDiT-based HunyuanVideo-I2V to 0.5571 (+7.44%).

Method: Developed VideoLoom, a unified Video Large Language Model. Created LoomData-8.7k dataset with human-centric videos featuring temporally grounded and spatially localized captions. Introduced LoomBench benchmark for comprehensive evaluation of spatial-temporal capabilities.

Result: Achieved state-of-the-art performance: 63.1 J&F on ReVOS for referring video object segmentation, and 48.3 R1@0.7 on Charades-STA for temporal grounding. Highly competitive across various spatial and temporal benchmarks.

Conclusion: VideoLoom provides a universal and effective suite for joint spatial-temporal video understanding, setting a new standard in multimodal intelligence through unified modeling, curated dataset, and comprehensive benchmark.

Abstract: This paper presents VideoLoom, a unified Video Large Language Model (Video LLM) for joint spatial-temporal understanding. To facilitate the development of fine-grained spatial and temporal localization capabilities, we curate LoomData-8.7k, a human-centric video dataset with temporally grounded and spatially localized captions. With this, VideoLoom achieves state-of-the-art or highly competitive performance across a variety of spatial and temporal benchmarks (e.g., 63.1 J&F on ReVOS for referring video object segmentation, and 48.3 R1@0.7 on Charades-STA for temporal grounding). In addition, we introduce LoomBench, a novel benchmark consisting of temporal, spatial, and compositional video-question pairs, enabling a comprehensive evaluation of Video LLMs from diverse aspects. Collectively, these contributions offer a universal and effective suite for joint spatial-temporal video understanding, setting a new standard in multimodal intelligence.

Method: Uses a lightweight prefix-tuner to extract dynamic Visual-Evidence Weights that quantify evidentiary support for candidate tokens based on visual input. These weights guide adaptive vocabulary partitioning and logits perturbation, concentrating watermark strength specifically on visually-supported tokens to align watermark with visual evidence.

Result: Outperforms conventional methods with 7.8% improvement in visual consistency (Chair-I) and superior semantic fidelity. Maintains highly competitive detection accuracy (96.88% AUC) and robust attack resilience (99.3%) without sacrificing inference efficiency.

Conclusion: VISA-Mark effectively establishes a new standard for reliability-preserving multimodal watermarking by actively aligning watermark signals with visual evidence, maintaining both visual fidelity and detection performance while preserving inference efficiency.

Abstract: Watermarking has emerged as a pivotal solution for content traceability and intellectual property protection in Large Vision-Language Models (LVLMs). However, vision-agnostic watermarks introduce visually irrelevant tokens and disrupt visual grounding by enforcing indiscriminate pseudo-random biases, while some semantic-aware methods incur prohibitive inference latency due to rejection sampling. In this paper, we propose the VIsual Semantic Adaptive Watermark (VISA-Mark), a novel framework that embeds detectable signals while strictly preserving visual fidelity. Our approach employs a lightweight, efficiently trained prefix-tuner to extract dynamic Visual-Evidence Weights, which quantify the evidentiary support for candidate tokens based on the visual input. These weights guide an adaptive vocabulary partitioning and logits perturbation mechanism, concentrating watermark strength specifically on visually-supported tokens. By actively aligning the watermark with visual evidence, VISA-Mark effectively maintains visual fidelity. Empirical results confirm that VISA-Mark outperforms conventional methods with a 7.8% improvement in visual consistency (Chair-I) and superior semantic fidelity. The framework maintains highly competitive detection accuracy (96.88% AUC) and robust attack resilience (99.3%) without sacrificing inference efficiency, effectively establishing a new standard for reliability-preserving multimodal watermarking.

Method: Map discrete sampling spaces to quasi-continuous feature spaces using kernel density estimation, then use density-adaptive hybrid sampling: Top-k for high-density regions (quality) and Random-k for low-density areas (diversity).

Result: Experiments in class-conditional and text-to-image evaluations show significant improvements in inference process quality.

Conclusion: VAR-Scaling successfully enables inference-time scaling for VQ visual autoregressive models by overcoming discrete space limitations, optimizing sample fidelity at critical scales to enhance output quality.

Abstract: While inference-time scaling has significantly enhanced generative quality in large language and diffusion models, its application to vector-quantized (VQ) visual autoregressive modeling (VAR) remains unexplored. We introduce VAR-Scaling, the first general framework for inference-time scaling in VAR, addressing the critical challenge of discrete latent spaces that prohibit continuous path search. We find that VAR scales exhibit two distinct pattern types: general patterns and specific patterns, where later-stage specific patterns conditionally optimize early-stage general patterns. To overcome the discrete latent space barrier in VQ models, we map sampling spaces to quasi-continuous feature spaces via kernel density estimation (KDE), where high-density samples approximate stable, high-quality solutions. This transformation enables effective navigation of sampling distributions. We propose a density-adaptive hybrid sampling strategy: Top-k sampling focuses on high-density regions to preserve quality near distribution modes, while Random-k sampling explores low-density areas to maintain diversity and prevent premature convergence. Consequently, VAR-Scaling optimizes sample fidelity at critical scales to enhance output quality. Experiments in class-conditional and text-to-image evaluations demonstrate significant improvements in inference process. The code is available at https://github.com/WD7ang/VAR-Scaling.

Method: Proposes CINEMA framework with five cognitive meta-actions (Global, Focus, Hint, Think, Answer), uses Retrieval-Based Tree Sampling for cold-start training, and implements two-stage RL with diversity-preserving exploration and annealed exploitation.

Result: Achieves competitive SOTA performance on multi-image reasoning benchmarks (surpasses GPT-4o on MUIR and MVMath), outperforms specialized video reasoning models, and demonstrates strong generalizability across multi-image, multi-frame, and single-image tasks.

Conclusion: The human cognition-inspired meta-action framework effectively addresses multi-image reasoning challenges and shows strong performance across diverse multimodal reasoning tasks.

Abstract: While Multimodal Large Language Models (MLLMs) excel at single-image understanding, they exhibit significantly degraded performance in multi-image reasoning scenarios. Multi-image reasoning presents fundamental challenges including complex inter-relationships between images and scattered critical information across image sets. Inspired by human cognitive processes, we propose the Cognition-Inspired Meta-Action Framework (CINEMA), a novel approach that decomposes multi-image reasoning into five structured meta-actions: Global, Focus, Hint, Think, and Answer which explicitly modeling the sequential cognitive steps humans naturally employ. For cold-start training, we introduce a Retrieval-Based Tree Sampling strategy that generates high-quality meta-action trajectories to bootstrap the model with reasoning patterns. During reinforcement learning, we adopt a two-stage paradigm: an exploration phase with Diversity-Preserving Strategy to avoid entropy collapse, followed by an annealed exploitation phase with DAPO to gradually strengthen exploitation. To train our model, we construct a dataset of 57k cold-start and 58k reinforcement learning instances spanning multi-image, multi-frame, and single-image tasks. We conduct extensive evaluations on multi-image reasoning benchmarks, video understanding benchmarks, and single-image benchmarks, achieving competitive state-of-the-art performance on several key benchmarks. Our model surpasses GPT-4o on the MUIR and MVMath benchmarks and notably outperforms specialized video reasoning models on video understanding benchmarks, demonstrating the effectiveness and generalizability of our human cognition-inspired reasoning framework.

Method: Built an evaluation suite of 18 attention topologies across four classes (sequential, parallel, multi-scale, residual) and systematically compared them across two vision and nine medical datasets.

Result: Discovered a “data scale-method-performance” coupling law: different attention structures perform best at different data scales (few-shot, medium-scale, large-scale). Also found that “Spatial-Channel” order is more stable for fine-grained classification, and residual connections help mitigate vanishing gradients.

Conclusion: Proposed scenario-based guidelines for building future attention modules based on data scale, with code open-sourced to facilitate further research and application.

Abstract: Attention mechanisms have become a core component of deep learning models, with Channel Attention and Spatial Attention being the two most representative architectures. Current research on their fusion strategies primarily bifurcates into sequential and parallel paradigms, yet the selection process remains largely empirical, lacking systematic analysis and unified principles. We systematically compare channel-spatial attention combinations under a unified framework, building an evaluation suite of 18 topologies across four classes: sequential, parallel, multi-scale, and residual. Across two vision and nine medical datasets, we uncover a “data scale-method-performance” coupling law: (1) in few-shot tasks, the “Channel-Multi-scale Spatial” cascaded structure achieves optimal performance; (2) in medium-scale tasks, parallel learnable fusion architectures demonstrate superior results; (3) in large-scale tasks, parallel structures with dynamic gating yield the best performance. Additionally, experiments indicate that the “Spatial-Channel” order is more stable and effective for fine-grained classification, while residual connections mitigate vanishing gradient problems across varying data scales. We thus propose scenario-based guidelines for building future attention modules. Code is open-sourced at https://github.com/DWlzm.

Method: Two-stage retrieval: 1) Text-based filtering with CLIP using captions from multi-view renderings, 2) Image-based refinement with DINOv2 for visual similarity. Uses GroundedSAM for object detection and image captioning for database annotation.

Result: Outperforms all SOTA methods on MI3DOR benchmark. Achieves 90.48% average precision on YCB-V dataset. Enables pose estimation with similar models when exact instances aren’t available, beating reconstruction-based approaches.

Conclusion: OSCAR provides effective training-free CAD model retrieval for open-set scenarios, automating model sourcing for 6D pose estimation and handling continuously changing object sets without requiring object-specific training.

Abstract: 6D object pose estimation plays a crucial role in scene understanding for applications such as robotics and augmented reality. To support the needs of ever-changing object sets in such context, modern zero-shot object pose estimators were developed to not require object-specific training but only rely on CAD models. Such models are hard to obtain once deployed, and a continuously changing and growing set of objects makes it harder to reliably identify the instance model of interest. To address this challenge, we introduce an Open-Set CAD Retrieval from a Language Prompt and a Single Image (OSCAR), a novel training-free method that retrieves a matching object model from an unlabeled 3D object database. During onboarding, OSCAR generates multi-view renderings of database models and annotates them with descriptive captions using an image captioning model. At inference, GroundedSAM detects the queried object in the input image, and multi-modal embeddings are computed for both the Region-of-Interest and the database captions. OSCAR employs a two-stage retrieval: text-based filtering using CLIP identifies candidate models, followed by image-based refinement using DINOv2 to select the most visually similar object. In our experiments we demonstrate that OSCAR outperforms all state-of-the-art methods on the cross-domain 3D model retrieval benchmark MI3DOR. Furthermore, we demonstrate OSCAR’s direct applicability in automating object model sourcing for 6D object pose estimation. We propose using the most similar object model for pose estimation if the exact instance is not available and show that OSCAR achieves an average precision of 90.48% during object retrieval on the YCB-V object dataset. Moreover, we demonstrate that the most similar object model can be utilized for pose estimation using Megapose achieving better results than a reconstruction-based approach.

Method: Proposes RGFS-Net with a masked image reconstruction task where parts of input images are occluded and reconstructed to encourage semantically rich feature learning, enhancing spatial understanding and class discrimination in low-data settings.

Result: Outperforms existing baselines on EuroSAT and PatternNet datasets under both 1-shot and 5-shot protocols, demonstrating consistent improvement in few-shot classification performance.

Conclusion: RGFS-Net provides a simple, effective, and backbone-compatible solution for few-shot remote sensing classification, offering robust performance through reconstruction-guided feature learning.

Abstract: Few-shot remote sensing image classification is challenging due to limited labeled samples and high variability in land-cover types. We propose a reconstruction-guided few-shot network (RGFS-Net) that enhances generalization to unseen classes while preserving consistency for seen categories. Our method incorporates a masked image reconstruction task, where parts of the input are occluded and reconstructed to encourage semantically rich feature learning. This auxiliary task strengthens spatial understanding and improves class discrimination under low-data settings. We evaluated the efficacy of EuroSAT and PatternNet datasets under 1-shot and 5-shot protocols, our approach consistently outperforms existing baselines. The proposed method is simple, effective, and compatible with standard backbones, offering a robust solution for few-shot remote sensing classification. Codes are available at https://github.com/stark0908/RGFS.

Method: 1) Construct MediScope dataset with 98,000 real-world multi-turn consultations and 601,500 medical images across 10+ clinical departments; 2) Develop PulseMind Benchmark with four-dimensional evaluation protocol (proactiveness, accuracy, usefulness, language quality); 3) Design training framework with Comparison-based Reinforcement Policy Optimization (CRPO) using relative preference signals.

Result: PulseMind achieves competitive performance on both the diagnostic consultation benchmark and public medical benchmarks through extensive experiments.

Conclusion: PulseMind bridges the gap between specialized medical image analysis and real-world clinical diagnostics by providing a comprehensive framework for multi-modal diagnostic models that better reflect the complexity of clinical practice.

Abstract: Recent advances in medical multi-modal models focus on specialized image analysis like dermatology, pathology, or radiology. However, they do not fully capture the complexity of real-world clinical diagnostics, which involve heterogeneous inputs and require ongoing contextual understanding during patient-physician interactions. To bridge this gap, we introduce PulseMind, a new family of multi-modal diagnostic models that integrates a systematically curated dataset, a comprehensive evaluation benchmark, and a tailored training framework. Specifically, we first construct a diagnostic dataset, MediScope, which comprises 98,000 real-world multi-turn consultations and 601,500 medical images, spanning over 10 major clinical departments and more than 200 sub-specialties. Then, to better reflect the requirements of real-world clinical diagnosis, we develop the PulseMind Benchmark, a multi-turn diagnostic consultation benchmark with a four-dimensional evaluation protocol comprising proactiveness, accuracy, usefulness, and language quality. Finally, we design a training framework tailored for multi-modal clinical diagnostics, centered around a core component named Comparison-based Reinforcement Policy Optimization (CRPO). Compared to absolute score rewards, CRPO uses relative preference signals from multi-dimensional com-parisons to provide stable and human-aligned training guidance. Extensive experiments demonstrate that PulseMind achieves competitive performance on both the diagnostic consultation benchmark and public medical benchmarks.

Method: DualPD has two components: 1) Layer-wise attention-guided contrastive logits module compares output logits between layers with largest attention shifts to track belief evolution; 2) Head-wise information filtering module suppresses low-contribution attention heads focusing on irrelevant regions to improve attention quality.

Result: Experiments on LLaVA and Qwen-VL model families across multiple multimodal benchmarks show DualPD consistently improves accuracy without any training, confirming effectiveness and generalizability.

Conclusion: DualPD effectively addresses the “seeing it right but saying it wrong” problem in MLLMs through dual-perspective decoding refinement, enhancing visual understanding without additional training.

Abstract: Multimodal Large Language Models (MLLMs) have demonstrated strong capabilities across a variety of vision-language tasks. However, their internal reasoning often exhibits a critical inconsistency: although deeper layers may attend to the correct visual regions, final predictions are frequently misled by noisy attention from earlier layers. This results in a disconnect between what the model internally understands and what it ultimately expresses, a phenomenon we describe as seeing it right but saying it wrong. To address this issue, we propose DualPD, a dual-perspective decoding refinement strategy that enhances the visual understanding without any additional training. DualPD consists of two components. (1) The layer-wise attention-guided contrastive logits module captures how the belief in the correct answer evolves by comparing output logits between layers that exhibit the largest attention shift. (2) The head-wise information filtering module suppresses low-contribution attention heads that focus on irrelevant regions, thereby improving attention quality within each layer. Experiments conducted on both the LLaVA and Qwen-VL model families across multiple multimodal benchmarks demonstrate that DualPD consistently improves accuracy without training, confirming its effectiveness and generalizability. The code will be released upon publication.

Method: 1) Create E-HVC dataset with Temporal Chain-of-Thought (event-level) and Chapter Summary annotations via staged construction; 2) Develop SPA-Compressor to compress multimodal tokens into hierarchical representations using ASR semantic cues; 3) Build HiVid-Narrator framework for efficient training.

Result: HiVid-Narrator achieves superior narrative quality with fewer input tokens compared to existing methods, effectively handling e-commerce video characteristics.

Conclusion: The proposed dual-granularity annotation approach and hierarchical compression framework successfully address the challenges of generating structured narrations for complex e-commerce videos.

Abstract: Generating structured narrations for real-world e-commerce videos requires models to perceive fine-grained visual details and organize them into coherent, high-level stories–capabilities that existing approaches struggle to unify. We introduce the E-commerce Hierarchical Video Captioning (E-HVC) dataset with dual-granularity, temporally grounded annotations: a Temporal Chain-of-Thought that anchors event-level observations and Chapter Summary that compose them into concise, story-centric summaries. Rather than directly prompting chapters, we adopt a staged construction that first gathers reliable linguistic and visual evidence via curated ASR and frame-level descriptions, then refines coarse annotations into precise chapter boundaries and titles conditioned on the Temporal Chain-of-Thought, yielding fact-grounded, time-aligned narratives. We also observe that e-commerce videos are fast-paced and information-dense, with visual tokens dominating the input sequence. To enable efficient training while reducing input tokens, we propose the Scene-Primed ASR-anchored Compressor (SPA-Compressor), which compresses multimodal tokens into hierarchical scene and event representations guided by ASR semantic cues. Built upon these designs, our HiVid-Narrator framework achieves superior narrative quality with fewer input tokens compared to existing methods.

Method: Proposes a dynamic collaborative network where two models dynamically switch teacher-student roles, includes a multi-view integration module to capture various input perspectives, and uses adversarial supervision to constrain vessel shape in unlabeled data. Projects 3D volumes to 2D views to mitigate label inconsistencies.

Result: DiCo achieves state-of-the-art performance on three 3D vessel segmentation benchmarks.

Conclusion: The dynamic role-switching approach combined with multi-view integration and adversarial supervision effectively addresses limitations of conventional mean teacher methods for 3D vessel segmentation.

Abstract: In this paper, we present a new dynamic collaborative network for semi-supervised 3D vessel segmentation, termed DiCo. Conventional mean teacher (MT) methods typically employ a static approach, where the roles of the teacher and student models are fixed. However, due to the complexity of 3D vessel data, the teacher model may not always outperform the student model, leading to cognitive biases that can limit performance. To address this issue, we propose a dynamic collaborative network that allows the two models to dynamically switch their teacher-student roles. Additionally, we introduce a multi-view integration module to capture various perspectives of the inputs, mirroring the way doctors conduct medical analysis. We also incorporate adversarial supervision to constrain the shape of the segmented vessels in unlabeled data. In this process, the 3D volume is projected into 2D views to mitigate the impact of label inconsistencies. Experiments demonstrate that our DiCo method sets new state-of-the-art performance on three 3D vessel segmentation benchmarks. The code repository address is https://github.com/xujiaommcome/DiCo

Method: SVD-Cache decomposes diffusion features via Singular Value Decomposition (SVD) into principal (low-rank) and residual subspaces. It applies exponential moving average (EMA) prediction to the smooth principal components while directly reusing the volatile residual subspace.

Result: Achieves near-lossless acceleration across diverse models, including 5.55× speedup on FLUX and HunyuanVideo. Compatible with other acceleration techniques like distillation, quantization, and sparse attention.

Conclusion: By exploiting the distinct temporal behaviors of principal vs. residual subspaces in DiT features, SVD-Cache provides an effective caching framework that significantly accelerates inference while maintaining quality.

Abstract: Diffusion Transformer (DiT) models have achieved unprecedented quality in image and video generation, yet their iterative sampling process remains computationally prohibitive. To accelerate inference, feature caching methods have emerged by reusing intermediate representations across timesteps. However, existing caching approaches treat all feature components uniformly. We reveal that DiT feature spaces contain distinct principal and residual subspaces with divergent temporal behavior: the principal subspace evolves smoothly and predictably, while the residual subspace exhibits volatile, low-energy oscillations that resist accurate prediction. Building on this insight, we propose SVD-Cache, a subspace-aware caching framework that decomposes diffusion features via Singular Value Decomposition (SVD), applies exponential moving average (EMA) prediction to the dominant low-rank components, and directly reuses the residual subspace. Extensive experiments demonstrate that SVD-Cache achieves near-lossless across diverse models and methods, including 5.55$\times$ speedup on FLUX and HunyuanVideo, and compatibility with model acceleration techniques including distillation, quantization and sparse attention. Our code is in supplementary material and will be released on Github.

Method: Apply self-distillation to HSI classification by treating earlier network outputs as soft targets, enforcing consistency between intermediate and final predictions to improve intra-class compactness and inter-class separability in the learned feature space.

Result: Validated on two benchmark HSI datasets, the approach demonstrates significant improvements in classification accuracy and robustness, highlighting SD’s effectiveness for spectral-spatial learning.

Conclusion: Self-distillation is an effective strategy for HSI classification that enhances model performance without requiring external teacher networks, improving feature space properties and overall classification results.

Abstract: Hyperspectral image (HSI) classification presents unique challenges due to its high spectral dimensionality and limited labeled data. Traditional deep learning models often suffer from overfitting and high computational costs. Self-distillation (SD), a variant of knowledge distillation where a network learns from its own predictions, has recently emerged as a promising strategy to enhance model performance without requiring external teacher networks. In this work, we explore the application of SD to HSI by treating earlier outputs as soft targets, thereby enforcing consistency between intermediate and final predictions. This process improves intra-class compactness and inter-class separability in the learned feature space. Our approach is validated on two benchmark HSI datasets and demonstrates significant improvements in classification accuracy and robustness, highlighting the effectiveness of SD for spectral-spatial learning. Codes are available at https://github.com/Prachet-Dev-Singh/SDHSI.

Method: Integrates pre-trained SAM encoder modified to output multi-stage features. Introduces spatio-modal fusion module to select relevant modalities and best features for different areas. Uses semantic decoder with spherical attention and dual view fusion to handle panoramic distortions and edge discontinuities.

Result: Achieves state-of-the-art results on Stanford2D3DS for RGB, RGB-D, and RGB-D-N modalities, and on Matterport3D for RGB and RGB-D modalities.

Conclusion: PanoSAMic successfully adapts foundation models for panoramic image segmentation, demonstrating effective handling of spherical distortions and multi-modal fusion for superior performance on panoramic datasets.

Abstract: Existing image foundation models are not optimized for spherical images having been trained primarily on perspective images. PanoSAMic integrates the pre-trained Segment Anything (SAM) encoder to make use of its extensive training and integrate it into a semantic segmentation model for panoramic images using multiple modalities. We modify the SAM encoder to output multi-stage features and introduce a novel spatio-modal fusion module that allows the model to select the relevant modalities and best features from each modality for different areas of the input. Furthermore, our semantic decoder uses spherical attention and dual view fusion to overcome the distortions and edge discontinuity often associated with panoramic images. PanoSAMic achieves state-of-the-art (SotA) results on Stanford2D3DS for RGB, RGB-D, and RGB-D-N modalities and on Matterport3D for RGB and RGB-D modalities. https://github.com/dfki-av/PanoSAMic

Method: Proposes query-based frame selection using submodular mutual information (SMI) functions to select frames relevant to questions, ensuring complementary and essential visual information for accurate VideoQA.

Result: Achieves up to 4% accuracy improvement on MVBench dataset compared to uniform sampling when tested with Video-LLaVA and LLaVA-NeXT models. Qualitative analysis shows selected frames are better aligned with questions.

Conclusion: Query-based frame selection using SMI functions significantly improves VideoQA accuracy and can benefit various tasks relying on video frame subsets, offering a smarter alternative to uniform sampling.

Abstract: Video Question Answering (VideoQA) models enhance understanding and interaction with audiovisual content, making it more accessible, searchable, and useful for a wide range of fields such as education, surveillance, entertainment, and content creation. Due to heavy compute requirements, most large visual language models (VLMs) for VideoQA rely on a fixed number of frames by uniformly sampling the video. However, this process does not pick important frames or capture the context of the video. We present a novel query-based selection of frames relevant to the questions based on the submodular mutual Information (SMI) functions. By replacing uniform frame sampling with query-based selection, our method ensures that the chosen frames provide complementary and essential visual information for accurate VideoQA. We evaluate our approach on the MVBench dataset, which spans a diverse set of multi-action video tasks. VideoQA accuracy on this dataset was assessed using two VLMs, namely Video-LLaVA and LLaVA-NeXT, both of which originally employed uniform frame sampling. Experiments were conducted using both uniform and query-based sampling strategies. An accuracy improvement of up to \textbf{4%} was observed when using query-based frame selection over uniform sampling. Qualitative analysis further highlights that query-based selection, using SMI functions, consistently picks frames better aligned with the question. We opine that such query-based frame selection can enhance accuracy in a wide range of tasks that rely on only a subset of video frames.

Method: Fresco proposes a dynamic resolution framework that unifies re-noise and global structure across stages with progressive upsampling. It preserves both the efficiency of low-resolution drafting and the fidelity of high-resolution refinement, with all stages aligned toward the same final target.

Result: Fresco achieves near-lossless acceleration across diverse domains and models, including 10x speedup on FLUX and 5x on HunyuanVideo. It remains orthogonal to other acceleration techniques like distillation, quantization, and feature caching, reaching 22x speedup when combined with distilled models.

Conclusion: Fresco provides an effective dynamic resolution framework that addresses the limitations of existing methods, offering significant acceleration while maintaining quality, and is compatible with other optimization techniques for even greater speed improvements.

Abstract: Diffusion Transformers achieve impressive generative quality but remain computationally expensive due to iterative sampling. Recently, dynamic resolution sampling has emerged as a promising acceleration technique by reducing the resolution of early sampling steps. However, existing methods rely on heuristic re-noising at every resolution transition, injecting noise that breaks cross-stage consistency and forces the model to relearn global structure. In addition, these methods indiscriminately upsample the entire latent space at once without checking which regions have actually converged, causing accumulated errors, and visible artifacts. Therefore, we propose \textbf{Fresco}, a dynamic resolution framework that unifies re-noise and global structure across stages with progressive upsampling, preserving both the efficiency of low-resolution drafting and the fidelity of high-resolution refinement, with all stages aligned toward the same final target. Fresco achieves near-lossless acceleration across diverse domains and models, including 10$\times$ speedup on FLUX, and 5$\times$ on HunyuanVideo, while remaining orthogonal to distillation, quantization and feature caching, reaching 22$\times$ speedup when combined with distilled models. Our code is in supplementary material and will be released on Github.

Method: Proposes FocalOrder framework with Focal Preference Optimization (FPO): 1) Adaptive difficulty discovery using exponential moving average to dynamically identify hard-to-learn transitions, 2) Difficulty-calibrated pairwise ranking objective to enforce global logical consistency.

Result: Establishes new state-of-the-art results on OmniDocBench v1.0 and Comp-HRDoc benchmarks. Compact model outperforms specialized baselines and significantly surpasses large-scale general VLMs.

Conclusion: Aligning optimization with intrinsic structural ambiguity of documents is critical for mastering complex document structures. FocalOrder’s focus on difficult layout transitions addresses the positional disparity problem effectively.

Abstract: Reading order detection is the foundation of document understanding. Most existing methods rely on uniform supervision, implicitly assuming a constant difficulty distribution across layout regions. In this work, we challenge this assumption by revealing a critical flaw: \textbf{Positional Disparity}, a phenomenon where models demonstrate mastery over the deterministic start and end regions but suffer a performance collapse in the complex intermediate sections. This degradation arises because standard training allows the massive volume of easy patterns to drown out the learning signals from difficult layouts. To address this, we propose \textbf{FocalOrder}, a framework driven by \textbf{Focal Preference Optimization (FPO)}. Specifically, FocalOrder employs adaptive difficulty discovery with exponential moving average mechanism to dynamically pinpoint hard-to-learn transitions, while introducing a difficulty-calibrated pairwise ranking objective to enforce global logical consistency. Extensive experiments demonstrate that FocalOrder establishes new state-of-the-art results on OmniDocBench v1.0 and Comp-HRDoc. Our compact model not only outperforms competitive specialized baselines but also significantly surpasses large-scale general VLMs. These results demonstrate that aligning the optimization with intrinsic structural ambiguity of documents is critical for mastering complex document structures.

Method: Proposes Anatomy Aware Cascade Network (AACNet) with two key mechanisms: 1) Ambiguity Gated Boundary Refiner (AGBR) using entropy-based gating for targeted feature rectification in high uncertainty zones, and 2) Signed Distance Map guided Anatomical Attention (SDMAA) integrating implicit geometric constraints to enforce topological consistency and preserve spatial details.

Result: Achieves Dice Similarity Coefficient of 90.17% and 95% Hausdorff Distance of 3.63 mm on 125 CBCT volumes, significantly outperforming state-of-the-art methods. Strong generalization on external dataset with HD95 of 2.19 mm.

Conclusion: AACNet effectively resolves boundary ambiguity while maintaining global structural consistency, demonstrating reliability for clinical applications like surgical planning. Code is publicly available.

Abstract: Accurate three-dimensional (3D) tooth segmentation from Cone-Beam Computed Tomography (CBCT) is a prerequisite for digital dental workflows. However, achieving high-fidelity segmentation remains challenging due to adhesion artifacts in naturally occluded scans, which are caused by low contrast and indistinct inter-arch boundaries. To address these limitations, we propose the Anatomy Aware Cascade Network (AACNet), a coarse-to-fine framework designed to resolve boundary ambiguity while maintaining global structural consistency. Specifically, we introduce two mechanisms: the Ambiguity Gated Boundary Refiner (AGBR) and the Signed Distance Map guided Anatomical Attention (SDMAA). The AGBR employs an entropy based gating mechanism to perform targeted feature rectification in high uncertainty transition zones. Meanwhile, the SDMAA integrates implicit geometric constraints via signed distance map to enforce topological consistency, preventing the loss of spatial details associated with standard pooling. Experimental results on a dataset of 125 CBCT volumes demonstrate that AACNet achieves a Dice Similarity Coefficient of 90.17 % and a 95% Hausdorff Distance of 3.63 mm, significantly outperforming state-of-the-art methods. Furthermore, the model exhibits strong generalization on an external dataset with an HD95 of 2.19 mm, validating its reliability for downstream clinical applications such as surgical planning. Code for AACNet is available at https://github.com/shiliu0114/AACNet.

Method: Two-phase approach: 1) Offline multi-view reconstruction to build user-specific 3DGS-based avatar, 2) Real-time monocular inference using single RGB camera to capture motion/expressions. Features transmitted via WebRTC data channel, with lightweight 3DGS attribute deformation network on receiver side for dynamic adjustments.

Result: State-of-the-art performance: PSNR >28 dB for novel poses, end-to-end latency ~80 ms, >1000x bandwidth reduction vs point-cloud streaming, operates at ~60 FPS on Meta Quest 3 with <0.2 Mbps bandwidth.

Conclusion: Mon3tr enables practical, high-quality immersive telepresence on mobile devices by combining 3DGS parametric modeling with efficient monocular capture and transmission, overcoming hardware and bandwidth limitations of existing systems.

Abstract: Immersive telepresence aims to transform human interaction in AR/VR applications by enabling lifelike full-body holographic representations for enhanced remote collaboration. However, existing systems rely on hardware-intensive multi-camera setups and demand high bandwidth for volumetric streaming, limiting their real-time performance on mobile devices. To overcome these challenges, we propose Mon3tr, a novel Monocular 3D telepresence framework that integrates 3D Gaussian splatting (3DGS) based parametric human modeling into telepresence for the first time. Mon3tr adopts an amortized computation strategy, dividing the process into a one-time offline multi-view reconstruction phase to build a user-specific avatar and a monocular online inference phase during live telepresence sessions. A single monocular RGB camera is used to capture body motions and facial expressions in real time to drive the 3DGS-based parametric human model, significantly reducing system complexity and cost. The extracted motion and appearance features are transmitted at < 0.2 Mbps over WebRTC’s data channel, allowing robust adaptation to network fluctuations. On the receiver side, e.g., Meta Quest 3, we develop a lightweight 3DGS attribute deformation network to dynamically generate corrective 3DGS attribute adjustments on the pre-built avatar, synthesizing photorealistic motion and appearance at ~ 60 FPS. Extensive experiments demonstrate the state-of-the-art performance of our method, achieving a PSNR of > 28 dB for novel poses, an end-to-end latency of ~ 80 ms, and > 1000x bandwidth reduction compared to point-cloud streaming, while supporting real-time operation from monocular inputs across diverse scenarios. Our demos can be found at https://mon3tr3d.github.io.

Method: ViewMorpher3D uses image diffusion models to jointly process multiple rendered views conditioned on camera poses, 3D geometric priors, and temporally adjacent/spatially overlapping reference views. It accommodates variable camera numbers and flexible view configurations.

Result: Experiments on real-world driving datasets show substantial improvements in image quality metrics, effectively reducing artifacts while preserving geometric fidelity.

Conclusion: ViewMorpher3D provides an effective framework for enhancing multi-view image quality in driving simulators, addressing artifacts from 3D reconstruction methods and improving photorealism and cross-view consistency for autonomous driving development.

Abstract: Autonomous driving systems rely heavily on multi-view images to ensure accurate perception and robust decision-making. To effectively develop and evaluate perception stacks and planning algorithms, realistic closed-loop simulators are indispensable. While 3D reconstruction techniques such as Gaussian Splatting offer promising avenues for simulator construction, the rendered novel views often exhibit artifacts, particularly in extrapolated perspectives or when available observations are sparse. We introduce ViewMorpher3D, a multi-view image enhancement framework based on image diffusion models, designed to elevate photorealism and multi-view coherence in driving scenes. Unlike single-view approaches, ViewMorpher3D jointly processes a set of rendered views conditioned on camera poses, 3D geometric priors, and temporally adjacent or spatially overlapping reference views. This enables the model to infer missing details, suppress rendering artifacts, and enforce cross-view consistency. Our framework accommodates variable numbers of cameras and flexible reference/target view configurations, making it adaptable to diverse sensor setups. Experiments on real-world driving datasets demonstrate substantial improvements in image quality metrics, effectively reducing artifacts while preserving geometric fidelity.

Method: Created BenchSeg dataset aggregating 55 dish scenes from existing datasets with 360° camera motion annotations. Evaluated 20 state-of-the-art segmentation models (SAM-based, transformer, CNN, large multimodal) on FoodSeg103 and BenchSeg, testing them alone and combined with video-memory modules.

Result: Standard image segmenters degrade sharply under novel viewpoints, but memory-augmented methods maintain temporal consistency. Best model (SeTR-MLA+XMem2) outperforms prior work (improving ~2.63% mAP over FoodMem).

Conclusion: BenchSeg provides a valuable benchmark for multi-view food segmentation, showing memory augmentation is crucial for temporal consistency in dietary analysis applications. Dataset and models are publicly released.

Abstract: Food image segmentation is a critical task for dietary analysis, enabling accurate estimation of food volume and nutrients. However, current methods suffer from limited multi-view data and poor generalization to new viewpoints. We introduce BenchSeg, a novel multi-view food video segmentation dataset and benchmark. BenchSeg aggregates 55 dish scenes (from Nutrition5k, Vegetables & Fruits, MetaFood3D, and FoodKit) with 25,284 meticulously annotated frames, capturing each dish under free 360° camera motion. We evaluate a diverse set of 20 state-of-the-art segmentation models (e.g., SAM-based, transformer, CNN, and large multimodal) on the existing FoodSeg103 dataset and evaluate them (alone and combined with video-memory modules) on BenchSeg. Quantitative and qualitative results demonstrate that while standard image segmenters degrade sharply under novel viewpoints, memory-augmented methods maintain temporal consistency across frames. Our best model based on a combination of SeTR-MLA+XMem2 outperforms prior work (e.g., improving over FoodMem by ~2.63% mAP), offering new insights into food segmentation and tracking for dietary analysis. We release BenchSeg to foster future research. The project page including the dataset annotations and the food segmentation models can be found at https://amughrabi.github.io/benchseg.

Method: Developed foundation model-based AI pipeline using UMedPT (best performing pretrained model) with MLP head for classification and FCOS-based head for lesion detection, integrating uncertainty quantification and explainability (Grad-CAM).

Result: Classification: 0.90 AUC, 0.82 sensitivity on combined test set; 0.85 sensitivity on external cohort. Detection: 69.1% overall lesion detection, ranging from 30% to 98% across lesion size quartiles. Uncertainty filtering improved AUC to 0.91.

Conclusion: Foundation model-based pipelines can support robust and interpretable CRLM detection and classification across heterogeneous CT data, with clinical benefit demonstrated for threshold probabilities between 0.30-0.40.

Abstract: Colorectal liver metastases (CRLM) are a major cause of cancer-related mortality, and reliable detection on CT remains challenging in multi-centre settings. We developed a foundation model-based AI pipeline for patient-level classification and lesion-level detection of CRLM on contrast-enhanced CT, integrating uncertainty quantification and explainability. CT data from the EuCanImage consortium (n=2437) and an external TCIA cohort (n=197) were used. Among several pretrained models, UMedPT achieved the best performance and was fine-tuned with an MLP head for classification and an FCOS-based head for lesion detection. The classification model achieved an AUC of 0.90 and a sensitivity of 0.82 on the combined test set, with a sensitivity of 0.85 on the external cohort. Excluding the most uncertain 20 percent of cases improved AUC to 0.91 and balanced accuracy to 0.86. Decision curve analysis showed clinical benefit for threshold probabilities between 0.30 and 0.40. The detection model identified 69.1 percent of lesions overall, increasing from 30 percent to 98 percent across lesion size quartiles. Grad-CAM highlighted lesion-corresponding regions in high-confidence cases. These results demonstrate that foundation model-based pipelines can support robust and interpretable CRLM detection and classification across heterogeneous CT data.

Method: Derive likelihood function for SPAD raw signals given photon flux, then derive score function, and apply diffusion models to express image priors for solving inverse problems.

Result: Developed statistical framework for SPAD signals, enabling use of diffusion models for image reconstruction, demonstrating impact of photon counts and timing information utilization.

Conclusion: Proper statistical modeling of SPAD signals enables effective inverse problem solving using diffusion models, with timing information providing significant benefits in reconstruction quality.

Abstract: We derive the likelihood of a raw signal in a single photon avalanche diode (SPAD), given a fixed photon flux. The raw signal comprises timing of detection events, which are nonlinearly related to the flux. Moreover, they are naturally stochastic. We then derive a score function of the signal. This is a key for solving inverse problems based on SPAD signals. We focus on deriving solutions involving a diffusion model, to express image priors. We demonstrate the effect of low or high photon counts, and the consequence of exploiting timing of detection events.

Method: 1) UV-guided avatar modeling with pixel-wise facial correspondence estimation to reproject screen pixels to UV space (camera/expression independent). 2) Learnable UV tokens with attention mechanisms at screen and UV levels. 3) Decoding learned UV tokens into canonical Gaussian attributes using aggregated UV information. 4) Training on large-scale synthetic identity-rich dataset.

Result: Significantly outperforms existing approaches in both monocular and multi-view settings. The method works with arbitrary number of unposed inputs including single images, multi-view captures, and smartphone videos.

Conclusion: UIKA presents an efficient, feed-forward approach to animatable Gaussian head modeling that eliminates the need for studio capture systems and lengthy optimization, making high-quality avatar creation more accessible from diverse input sources.

Abstract: We present UIKA, a feed-forward animatable Gaussian head model from an arbitrary number of unposed inputs, including a single image, multi-view captures, and smartphone-captured videos. Unlike the traditional avatar method, which requires a studio-level multi-view capture system and reconstructs a human-specific model through a long-time optimization process, we rethink the task through the lenses of model representation, network design, and data preparation. First, we introduce a UV-guided avatar modeling strategy, in which each input image is associated with a pixel-wise facial correspondence estimation. Such correspondence estimation allows us to reproject each valid pixel color from screen space to UV space, which is independent of camera pose and character expression. Furthermore, we design learnable UV tokens on which the attention mechanism can be applied at both the screen and UV levels. The learned UV tokens can be decoded into canonical Gaussian attributes using aggregated UV information from all input views. To train our large avatar model, we additionally prepare a large-scale, identity-rich synthetic training dataset. Our method significantly outperforms existing approaches in both monocular and multi-view settings. Project page: https://zijian-wu.github.io/uika-page/

Method: PARL uses two key components: 1) Bidirectional Spatial Position-Guided Deformable Attention module to embed explicit positional dependencies among layout elements into visual features, and 2) Graph Refinement Classifier (GRC) to refine predictions by modeling contextual relationships through a dynamically constructed layout graph.

Result: PARL achieves state-of-the-art results, establishing a new benchmark for vision-only methods on DocLayNet and surpassing even strong multimodal models on M6Doc. It’s highly efficient with 65M parameters (4x fewer than typical 256M multimodal models).

Conclusion: Sophisticated visual structure modeling can be both more efficient and robust than multimodal fusion for document layout analysis, demonstrating that OCR-free vision-only approaches can outperform text-dependent methods while being computationally lighter.

Abstract: Document layout analysis aims to detect and categorize structural elements (e.g., titles, tables, figures) in scanned or digital documents. Popular methods often rely on high-quality Optical Character Recognition (OCR) to merge visual features with extracted text. This dependency introduces two major drawbacks: propagation of text recognition errors and substantial computational overhead, limiting the robustness and practical applicability of multimodal approaches. In contrast to the prevailing multimodal trend, we argue that effective layout analysis depends not on text-visual fusion, but on a deep understanding of documents’ intrinsic visual structure. To this end, we propose PARL (Position-Aware Relation Learning Network), a novel OCR-free, vision-only framework that models layout through positional sensitivity and relational structure. Specifically, we first introduce a Bidirectional Spatial Position-Guided Deformable Attention module to embed explicit positional dependencies among layout elements directly into visual features. Second, we design a Graph Refinement Classifier (GRC) to refine predictions by modeling contextual relationships through a dynamically constructed layout graph. Extensive experiments show PARL achieves state-of-the-art results. It establishes a new benchmark for vision-only methods on DocLayNet and, notably, surpasses even strong multimodal models on M6Doc. Crucially, PARL (65M) is highly efficient, using roughly four times fewer parameters than large multimodal models (256M), demonstrating that sophisticated visual structure modeling can be both more efficient and robust than multimodal fusion.

Method: 1) Decoder-only quantizer with Gumbel-Softmax for differentiable training and balanced codebook usage; 2) Sparse projection that maps motion codes into LLM embedding space while preserving orthogonality; 3) Two-stage orthonormal regularization schedule for soft constraints during tokenizer training and LLM fine-tuning.

Result: Extensive experiments on HumanML3D demonstrate 20% performance improvement over current state-of-the-art methods.

Conclusion: A unified geometric basis between motion codebook and LLM embedding space effectively empowers LLMs for nuanced motion reasoning, validating the importance of geometric alignment in motion-language models.

Abstract: Discrete motion tokenization has recently enabled Large Language Models (LLMs) to serve as versatile backbones for motion understanding and motion-language reasoning. However, existing pipelines typically decouple motion quantization from semantic embedding learning, linking them solely via token IDs. This approach fails to effectively align the intrinsic geometry of the motion space with the embedding space, thereby hindering the LLM’s capacity for nuanced motion reasoning. We argue that alignment is most effective when both modalities share a unified geometric basis. Therefore, instead of forcing the LLM to reconstruct the complex geometry among motion tokens from scratch, we present a novel framework that explicitly enforces orthogonality on both the motion codebook and the LLM embedding space, ensuring that their relational structures naturally mirror each other. Specifically, we employ a decoder-only quantizer with Gumbel-Softmax for differentiable training and balanced codebook usage. To bridge the modalities, we use a sparse projection that maps motion codes into the LLM embedding space while preserving orthogonality. Finally, a two-stage orthonormal regularization schedule enforces soft constraints during tokenizer training and LLM fine-tuning to maintain geometric alignment without hindering semantic adaptation. Extensive experiments on HumanML3D demonstrate that our framework achieves a 20% performance improvement over current state-of-the-art methods, validating that a unified geometric basis effectively empowers the LLM for nuanced motion reasoning.

Method: Uses Dual-branch Semantic-aware Large Reconstruction Model (Dual-Branch S-LRM) for joint geometry, color, and per-component semantic reconstruction. Introduces semantic surface extraction with coarse-to-fine proposal scheme for memory efficiency, and video-diffusion-based texture decomposition for editable appearance layers.

Result: Achieves state-of-the-art performance in geometric accuracy and semantic disentanglement, significantly outperforming existing methods. Enables structural independence for downstream applications.

Conclusion: StdGEN++ provides a robust solution for automated character asset production with advanced capabilities including non-destructive editing, physics-compliant animation, and gaze tracking, making it suitable for industrial pipelines.

Abstract: We present StdGEN++, a novel and comprehensive system for generating high-fidelity, semantically decomposed 3D characters from diverse inputs. Existing 3D generative methods often produce monolithic meshes that lack the structural flexibility required by industrial pipelines in gaming and animation. Addressing this gap, StdGEN++ is built upon a Dual-branch Semantic-aware Large Reconstruction Model (Dual-Branch S-LRM), which jointly reconstructs geometry, color, and per-component semantics in a feed-forward manner. To achieve production-level fidelity, we introduce a novel semantic surface extraction formalism compatible with hybrid implicit fields. This mechanism is accelerated by a coarse-to-fine proposal scheme, which significantly reduces memory footprint and enables high-resolution mesh generation. Furthermore, we propose a video-diffusion-based texture decomposition module that disentangles appearance into editable layers (e.g., separated iris and skin), resolving semantic confusion in facial regions. Experiments demonstrate that StdGEN++ achieves state-of-the-art performance, significantly outperforming existing methods in geometric accuracy and semantic disentanglement. Crucially, the resulting structural independence unlocks advanced downstream capabilities, including non-destructive editing, physics-compliant animation, and gaze tracking, making it a robust solution for automated character asset production.

Method: Proposes a variational contrastive learning framework that integrates probabilistic latent modeling with contrastive self-supervised learning, enabling learning of structured and semantically meaningful representations.

Result: Outperforms existing approaches on three skeleton-based action recognition benchmarks, especially in low-label regimes, with features that are more motion-relevant and focus on important skeleton joints.

Conclusion: The proposed variational contrastive learning framework effectively addresses motion uncertainty and variability, producing superior representations for skeleton-based action recognition that generalize well across datasets and supervision levels.

Abstract: In recent years, self-supervised representation learning for skeleton-based action recognition has advanced with the development of contrastive learning methods. However, most of contrastive paradigms are inherently discriminative and often struggle to capture the variability and uncertainty intrinsic to human motion. To address this issue, we propose a variational contrastive learning framework that integrates probabilistic latent modeling with contrastive self-supervised learning. This formulation enables the learning of structured and semantically meaningful representations that generalize across different datasets and supervision levels. Extensive experiments on three widely used skeleton-based action recognition benchmarks show that our proposed method consistently outperforms existing approaches, particularly in low-label regimes. Moreover, qualitative analyses show that the features provided by our method are more relevant given the motion and sample characteristics, with more focus on important skeleton joints, when compared to the other methods.

Method: Benchmarked 16 OCR models on 12 public datasets from various regions. Explored three synthetic data generation methods: template-based generation, character permutation, and GAN models. Investigated combined use of these methodologies and trade-offs between accuracy and speed.

Result: Massive incorporation of synthetic data substantially boosts model performance in both intra- and cross-dataset scenarios. Combined methodologies show synergistic effects, surpassing state-of-the-art methods and commercial systems. Synthetic data also mitigates limited training data challenges, enabling remarkable results with small fractions of original data.

Conclusion: Synthetic data integration significantly enhances LPR performance, with combined generation methods showing synergistic effects. The approach addresses training data limitations and identifies optimal accuracy-speed trade-offs for different scenarios, outperforming existing methods and commercial systems.

Abstract: Automatic License Plate Recognition is a frequent research topic due to its wide-ranging practical applications. While recent studies use synthetic images to improve License Plate Recognition (LPR) results, there remain several limitations in these efforts. This work addresses these constraints by comprehensively exploring the integration of real and synthetic data to enhance LPR performance. We subject 16 Optical Character Recognition (OCR) models to a benchmarking process involving 12 public datasets acquired from various regions. Several key findings emerge from our investigation. Primarily, the massive incorporation of synthetic data substantially boosts model performance in both intra- and cross-dataset scenarios. We examine three distinct methodologies for generating synthetic data: template-based generation, character permutation, and utilizing a Generative Adversarial Network (GAN) model, each contributing significantly to performance enhancement. The combined use of these methodologies demonstrates a notable synergistic effect, leading to end-to-end results that surpass those reached by state-of-the-art methods and established commercial systems. Our experiments also underscore the efficacy of synthetic data in mitigating challenges posed by limited training data, enabling remarkable results to be achieved even with small fractions of the original training data. Finally, we investigate the trade-off between accuracy and speed among different models, identifying those that strike the optimal balance in each intra-dataset and cross-dataset settings.

Method: R3DPA introduces three key innovations: (1) aligning intermediate features of the generative model with self-supervised 3D features to improve quality, (2) transferring knowledge from large-scale image-pretrained generative models to LiDAR generation, and (3) enabling point cloud control at inference for object inpainting and scene mixing using only an unconditional model.

Result: Achieves state-of-the-art performance on the KITTI-360 benchmark for LiDAR scene generation, demonstrating superior quality compared to existing approaches.

Conclusion: R3DPA successfully bridges the gap between 2D image priors and 3D LiDAR generation, providing an effective solution to data scarcity in 3D perception tasks while enabling flexible scene manipulation capabilities.

Abstract: LiDAR scene synthesis is an emerging solution to scarcity in 3D data for robotic tasks such as autonomous driving. Recent approaches employ diffusion or flow matching models to generate realistic scenes, but 3D data remains limited compared to RGB datasets with millions of samples. We introduce R3DPA, the first LiDAR scene generation method to unlock image-pretrained priors for LiDAR point clouds, and leverage self-supervised 3D representations for state-of-the-art results. Specifically, we (i) align intermediate features of our generative model with self-supervised 3D features, which substantially improves generation quality; (ii) transfer knowledge from large-scale image-pretrained generative models to LiDAR generation, mitigating limited LiDAR datasets; and (iii) enable point cloud control at inference for object inpainting and scene mixing with solely an unconditional model. On the KITTI-360 benchmark R3DPA achieves state of the art performance. Code and pretrained models are available at https://github.com/valeoai/R3DPA.

Method: Introduces Smooth Numerical Reward Activation (SNRA) operator using dynamically parameterized Sigmoid function to transform raw feedback into dense continuous rewards, and Absolute-Preserving GRPO (AP-GRPO) framework that integrates absolute scalar gradients to prevent numerical information loss from relative-ranking mechanisms.

Result: Created Numerical3D-50k dataset with 50,000 verifiable 3D subtasks. AP-GRPO achieves performance parity with large-scale supervised methods while maintaining higher data efficiency, effectively activating latent 3D reasoning in VLMs without architectural modifications.

Conclusion: The AP-GRPO framework with SNRA operator successfully addresses reward sparsity and gradient instability in VLMs for 3D scene understanding, enabling precise numerical prediction while maintaining data efficiency and requiring no architectural changes.

Abstract: Vision-Language Models (VLMs) face a critical bottleneck in achieving precise numerical prediction for 3D scene understanding. Traditional reinforcement learning (RL) approaches, primarily based on relative ranking, often suffer from severe reward sparsity and gradient instability, failing to effectively exploit the verifiable signals provided by 3D physical constraints. Notably, in standard GRPO frameworks, relative normalization causes “near-miss” samples (characterized by small but non-zero errors) to suffer from advantage collapse. This leads to a severe data utilization bottleneck where valuable boundary samples are discarded during optimization. To address this, we introduce the Smooth Numerical Reward Activation (SNRA) operator and the Absolute-Preserving GRPO (AP-GRPO) framework. SNRA employs a dynamically parameterized Sigmoid function to transform raw feedback into a dense, continuous reward continuum. Concurrently, AP-GRPO integrates absolute scalar gradients to mitigate the numerical information loss inherent in conventional relative-ranking mechanisms. By leveraging this approach, we constructed Numerical3D-50k, a dataset comprising 50,000 verifiable 3D subtasks. Empirical results indicate that AP-GRPO achieves performance parity with large-scale supervised methods while maintaining higher data efficiency, effectively activating latent 3D reasoning in VLMs without requiring architectural modifications.

Method: Two approaches: 1) Adaptation of decomposition of trained ReLU networks into two monotone and convex parts, overcoming numerical weight blowup issues. 2) Training models as the difference between two monotone neural networks.

Result: Proposed saliency methods (SplitCAM and SplitLRP) improve state-of-the-art results on VGG16 and ResNet18 networks on ImageNet-S across all Quantus saliency metric categories. The second approach yields systems with strong self-explainability properties.

Conclusion: Monotonicity can be effectively leveraged to boost explainability through decomposition techniques and architectural designs, even when full monotonic approximation isn’t possible, leading to improved saliency methods and self-explainable systems.

Abstract: It has been demonstrated in various contexts that monotonicity leads to better explainability in neural networks. However, not every function can be well approximated by a monotone neural network. We demonstrate that monotonicity can still be used in two ways to boost explainability. First, we use an adaptation of the decomposition of a trained ReLU network into two monotone and convex parts, thereby overcoming numerical obstacles from an inherent blowup of the weights in this procedure. Our proposed saliency methods – SplitCAM and SplitLRP – improve on state of the art results on both VGG16 and Resnet18 networks on ImageNet-S across all Quantus saliency metric categories. Second, we exhibit that training a model as the difference between two monotone neural networks results in a system with strong self-explainability properties.

Method: Developed physically-based ray tracing rendering software that uses standard camera calibration coefficients, reproduces distortions/blur, and applies sub-pixel sampling. Uses low-discrepancy sampling of 6DOF space to systematically analyze pose errors across 36 parameter combinations.

Result: The method enables systematic evaluation of pose estimation accuracy, revealing strengths and weaknesses of well-known markers. The open-source code provides reproducible comparisons.

Conclusion: The proposed framework enables fair and comprehensive comparisons of fiducial marker pose estimation accuracy through high-fidelity synthetic data generation and systematic 6DOF analysis.

Abstract: This paper presents a method for carrying fair comparisons of the accuracy of pose estimation using fiducial markers. These comparisons rely on large sets of high-fidelity synthetic images enabling deep exploration of the 6 degrees of freedom. A low-discrepancy sampling of the space allows to check the correlations between each degree of freedom and the pose errors by plotting the 36 pairs of combinations. The images are rendered using a physically based ray tracing code that has been specifically developed to use the standard calibration coefficients of any camera directly. The software reproduces image distortions, defocus and diffraction blur. Furthermore, sub-pixel sampling is applied to sharp edges to enhance the fidelity of the rendered image. After introducing the rendering algorithm and its experimental validation, the paper proposes a method for evaluating the pose accuracy. This method is applied to well-known markers, revealing their strengths and weaknesses for pose estimation. The code is open source and available on GitHub.

Method: Semi-automated process using large language models, few-shot prompt engineering, and text-to-image generation to synthesize high-quality uncommon-sense image-text pairs, each with multiple-choice questions for fine-grained reasoning evaluation.

Result: All evaluated VLMs perform significantly worse than humans in semantic judgment, especially in distinguishing grammatical correctness from semantic rationality. Lightweight models show accuracy improvement after fine-tuning, demonstrating directional adaptation potential.

Conclusion: The study reveals key weaknesses in VLMs’ semantic reasoning capabilities and provides diagnostic tools and research directions for developing robust models with genuine visual semantic understanding.

Abstract: We propose UAIT (Uncommon-sense Action Image-Text) dataset, a new evaluation benchmark designed to test the semantic understanding ability of visual language models (VLMs) in uncommon-sense action scenes. Unlike previous datasets that focus on common visual scenes with statistical frequency advantages, UAIT challenges models with grammatically reasonable but semantically counter-common sense image-text pairs. Such tasks require models to go beyond superficial pattern recognition and demonstrate a deep understanding of agent-patient relationships and physical feasibility. To build UAIT, we designed a semi-automated process to synthesize high-quality uncommon-sense image-text samples using large language models, few-shot prompt engineering, and text-to-image generation. Each sample is accompanied by a carefully designed multiple-choice question to test the model’s competence in fine-grained reasoning. We evaluate multiple state-of-the-art visual language models and compare them with models based on contrastive learning. Experiments show that all models perform significantly worse than humans in semantic judgment, especially in distinguishing grammatical correctness from semantic rationality. Further experiments show that even the lightweight model can improve its accuracy after fine-tuning, demonstrating the great potential of directional adaptation. This study not only reveals the key weaknesses of VLMs, but also provides diagnostic tools and research directions for the development of robust models with real visual semantic reasoning capabilities.

Method: Develops variants of Wasserstein distance that incorporate discrete probability measures to weight different fragments of 2D curves, allowing selective focus on prescribed parts during classification.

Result: The approach is tested through experiments on clustering analysis of 2D curves using archaeological data, demonstrating the effectiveness of fragment-focused classification.

Conclusion: The proposed Wasserstein distance variants successfully enable targeted classification of specific curve fragments, providing a valuable tool for archaeological data analysis where partial curve information is most relevant.

Abstract: In this work we analyse a number of variants of the Wasserstein distance which allow to focus the classification on the prescribed parts (fragments) of classified 2D curves. These variants are based on the use of a number of discrete probability measures which reflect the importance of given fragments of curves. The performance of this approach is tested through a series of experiments related to the clustering analysis of 2D curves performed on data coming from the field of archaeology.

Method: 1. Evidence Grounding Module (EGM): Lightweight query-guided filter that extracts compact visual evidence. 2. Evidence-Anchoring Protocol: RL-optimized mechanism with composite reward enforcing process alignment, making models reference temporal anchors during deduction. 3. CoE-Instruct dataset: 164k samples with dual-annotation schema for separate perception and reasoning supervision.

Result: CoE-enhanced models achieve state-of-the-art performance on five benchmarks (Video-MME, MVBench, VSI-Bench, etc.), significantly outperforming existing methods in accuracy while reducing hallucinations.

Conclusion: CoE provides a powerful and practical paradigm for reliable video understanding by architecturally decoupling and co-optimizing perceptual grounding and reasoning efficiency, resolving the fundamental dilemma in video reasoning.

Abstract: Large Vision-Language Models (LVLMs) face a fundamental dilemma in video reasoning: they are caught between the prohibitive computational costs of verbose reasoning and the hallucination risks of efficient, ungrounded approaches. To resolve this, we introduce the Chain of Evidence (CoE), a novel framework that architecturally decouples and co-optimizes perceptual grounding and reasoning efficiency. CoE incorporates two core innovations: (1) A lightweight Evidence Grounding Module (EGM) that acts as a query-guided filter, dynamically identifying and extracting a compact set of high-fidelity visual evidence; and (2) An Evidence-Anchoring Protocol optimized via Reinforcement Learning. Crucially, we design a composite reward mechanism that enforces process alignment, compelling the model to strictly reference identified temporal anchors during deduction, thereby mitigating hallucinations. To enable this, we construct CoE-Instruct, a large-scale dataset (164k samples) featuring a novel dual-annotation schema for separate perception and reasoning supervision. Extensive experiments on five benchmarks, including Video-MME, MVBench, and VSI-Bench, demonstrate that CoE-enhanced models establish a new state-of-the-art. They significantly outperform existing methods in accuracy, proving CoE to be a powerful and practical paradigm for reliable video understanding.

Method: Two-phase approach: 1) Initially train DiT by aligning shallow features with VAE latent representations for short phase (e.g., 40 epochs), 2) Apply classifier-free guidance to intermediate features to enhance discriminative capability, then use these enriched internal features as supervision signals to guide new DiT training.

Result: Significant performance boost compared to existing self-contained methods. Can surpass REPA in generation quality and convergence speed without needing external pretrained models. More flexible for different backbones and potentially applicable to wider range of diffusion-based generative tasks.

Conclusion: Self-Transcendence demonstrates that DiTs can effectively guide their own training using internal feature supervision, achieving fast convergence without external dependencies while maintaining or improving performance over methods that rely on pretrained external networks.

Abstract: Recent works such as REPA have shown that guiding diffusion models with external semantic features (e.g., DINO) can significantly accelerate the training of diffusion transformers (DiTs). However, this requires the use of pretrained external networks, introducing additional dependencies and reducing flexibility. In this work, we argue that DiTs actually have the power to guide the training of themselves, and propose \textbf{Self-Transcendence}, a simple yet effective method that achieves fast convergence using internal feature supervision only. It is found that the slow convergence in DiT training primarily stems from the difficulty of representation learning in shallow layers. To address this, we initially train the DiT model by aligning its shallow features with the latent representations from the pretrained VAE for a short phase (e.g., 40 epochs), then apply classifier-free guidance to the intermediate features, enhancing their discriminative capability and semantic expressiveness. These enriched internal features, learned entirely within the model, are used as supervision signals to guide a new DiT training. Compared to existing self-contained methods, our approach brings a significant performance boost. It can even surpass REPA in terms of generation quality and convergence speed, but without the need for any external pretrained models. Our method is not only more flexible for different backbones but also has the potential to be adopted for a wider range of diffusion-based generative tasks. The source code of our method can be found at https://github.com/csslc/Self-Transcendence.

Method: Uses OWLv2 Vision Transformer model, fine-tuned on manually labeled IMPACT dataset of high-resolution lunar images. Employs parameter-efficient fine-tuning with Low-Rank Adaptation (LoRA) and optimizes combined CIoU loss for localization and contrastive loss for classification.

Result: Achieves maximum recall of 94.0% and maximum precision of 73.1% on IMPACT test dataset, with satisfactory visual results across challenging lunar imaging conditions.

Conclusion: The proposed method provides reliable crater detection for lunar exploration, enabling robust crater analysis and supporting safe lunar landings for ESA’s future missions.

Abstract: The European Space Agency (ESA), driven by its ambitions on planned lunar missions with the Argonaut lander, has a profound interest in reliable crater detection, since craters pose a risk to safe lunar landings. This task is usually addressed with automated crater detection algorithms (CDA) based on deep learning techniques. It is non-trivial due to the vast amount of craters of various sizes and shapes, as well as challenging conditions such as varying illumination and rugged terrain. Therefore, we propose a deep-learning CDA based on the OWLv2 model, which is built on a Vision Transformer, that has proven highly effective in various computer vision tasks. For fine-tuning, we utilize a manually labeled dataset fom the IMPACT project, that provides crater annotations on high-resolution Lunar Reconnaissance Orbiter Camera Calibrated Data Record images. We insert trainable parameters using a parameter-efficient fine-tuning strategy with Low-Rank Adaptation, and optimize a combined loss function consisting of Complete Intersection over Union (CIoU) for localization and a contrastive loss for classification. We achieve satisfactory visual results, along with a maximum recall of 94.0% and a maximum precision of 73.1% on a test dataset from IMPACT. Our method achieves reliable crater detection across challenging lunar imaging conditions, paving the way for robust crater analysis in future lunar exploration.

Method: SEED uses weight-shared Siamese encoders and decoders with a parameter-free feature exchange mechanism instead of explicit differencing. The feature exchange is formalized as an orthogonal permutation operator that preserves mutual information and Bayes optimal risk under pixel consistency. The approach can transform standard semantic segmentation models into change detectors (SEG2CD) by simply inserting the exchange mechanism.

Result: Extensive experiments across five benchmarks (SYSU-CD, LEVIR-CD, PX-CLCD, WaterCD, CDD) and three backbones (SwinT, EfficientNet, ResNet) show SEED matches or surpasses state-of-the-art methods despite its simplicity. The approach demonstrates that simple feature exchange is sufficient for high-performance information fusion.

Conclusion: SEED provides a robust, unified, and interpretable framework for change detection, proving that parameter-free feature exchange can effectively replace complex explicit differencing modules while maintaining or improving performance.

Abstract: Remote sensing change detection fundamentally relies on the effective fusion and discrimination of bi-temporal features. Prevailing paradigms typically utilize Siamese encoders bridged by explicit difference computation modules, such as subtraction or concatenation, to identify changes. In this work, we challenge this complexity with SEED (Siamese Encoder-Exchange-Decoder), a streamlined paradigm that replaces explicit differencing with parameter-free feature exchange. By sharing weights across both Siamese encoders and decoders, SEED effectively operates as a single parameter set model. Theoretically, we formalize feature exchange as an orthogonal permutation operator and prove that, under pixel consistency, this mechanism preserves mutual information and Bayes optimal risk, whereas common arithmetic fusion methods often introduce information loss. Extensive experiments across five benchmarks, including SYSU-CD, LEVIR-CD, PX-CLCD, WaterCD, and CDD, and three backbones, namely SwinT, EfficientNet, and ResNet, demonstrate that SEED matches or surpasses state of the art methods despite its simplicity. Furthermore, we reveal that standard semantic segmentation models can be transformed into competitive change detectors solely by inserting this exchange mechanism, referred to as SEG2CD. The proposed paradigm offers a robust, unified, and interpretable framework for change detection, demonstrating that simple feature exchange is sufficient for high performance information fusion. Code and full training and evaluation protocols will be released at https://github.com/dyzy41/open-rscd.

Method: Introduces MIMIC benchmark for rigorous multi-image evaluation; proposes two remedies: 1) procedural data generation composing single-image annotations into targeted multi-image training examples, 2) attention-masking scheme derived from layer-wise attention pattern analysis for multi-image inputs.

Result: Diagnostic experiments reveal LVLMs fail to aggregate information across images and struggle with multiple concept tracking; proposed solutions substantially improve cross-image aggregation and outperform prior state-of-the-art on existing multi-image benchmarks.

Conclusion: The MIMIC benchmark identifies critical weaknesses in LVLMs’ multi-image capabilities, and the proposed data generation and attention optimization methods effectively address these limitations, advancing multi-image reasoning performance.

Abstract: Large Vision Language Models (LVLMs) have demonstrated remarkable capabilities, yet their proficiency in understanding and reasoning over multiple images remains largely unexplored. While existing benchmarks have initiated the evaluation of multi-image models, a comprehensive analysis of their core weaknesses and their causes is still lacking. In this work, we introduce MIMIC (Multi-Image Model Insights and Challenges), a new benchmark designed to rigorously evaluate the multi-image capabilities of LVLMs. Using MIMIC, we conduct a series of diagnostic experiments that reveal pervasive issues: LVLMs often fail to aggregate information across images and struggle to track or attend to multiple concepts simultaneously. To address these failures, we propose two novel complementary remedies. On the data side, we present a procedural data-generation strategy that composes single-image annotations into rich, targeted multi-image training examples. On the optimization side, we analyze layer-wise attention patterns and derive an attention-masking scheme tailored for multi-image inputs. Experiments substantially improved cross-image aggregation, while also enhancing performance on existing multi-image benchmarks, outperforming prior state of the art across tasks. Data and code will be made available at https://github.com/anurag-198/MIMIC.

Method: Proposes Multi-Head Linear Attention (MHLA) which preserves representational diversity by computing attention within divided heads along the token dimension. This maintains linear complexity while recovering expressive power of softmax attention.

Result: Achieves significant improvements across multiple domains: 3.6% improvement on ImageNet classification, 6.3% gain on NLP, 12.6% improvement on image generation, and 41% enhancement on video generation under the same time complexity.

Conclusion: MHLA effectively addresses the global context collapse problem in linear attention methods, maintaining computational efficiency while recovering much of the performance of standard softmax attention across diverse domains.

Abstract: While the Transformer architecture dominates many fields, its quadratic self-attention complexity hinders its use in large-scale applications. Linear attention offers an efficient alternative, but its direct application often degrades performance, with existing fixes typically re-introducing computational overhead through extra modules (e.g., depthwise separable convolution) that defeat the original purpose. In this work, we identify a key failure mode in these methods: global context collapse, where the model loses representational diversity. To address this, we propose Multi-Head Linear Attention (MHLA), which preserves this diversity by computing attention within divided heads along the token dimension. We prove that MHLA maintains linear complexity while recovering much of the expressive power of softmax attention, and verify its effectiveness across multiple domains, achieving a 3.6% improvement on ImageNet classification, a 6.3% gain on NLP, a 12.6% improvement on image generation, and a 41% enhancement on video generation under the same time complexity.

Method: 1) Created a large-scale dataset of triplets (reference effect video, input, output) using an automated pipeline for video-to-video effects. 2) Augmented with image-to-video effects from LoRA adapters and code-based temporal effects. 3) Trained a reference-conditioned model using recent text-to-video backbones.

Result: RefVFX produces visually consistent and temporally coherent edits, generalizes across unseen effect categories, and outperforms prompt-only baselines in both quantitative metrics and human preference.

Conclusion: The framework successfully enables transfer of complex temporal effects from reference videos to target content, addressing a key limitation in current video editing methods.

Abstract: We present RefVFX, a new framework that transfers complex temporal effects from a reference video onto a target video or image in a feed-forward manner. While existing methods excel at prompt-based or keyframe-conditioned editing, they struggle with dynamic temporal effects such as dynamic lighting changes or character transformations, which are difficult to describe via text or static conditions. Transferring a video effect is challenging, as the model must integrate the new temporal dynamics with the input video’s existing motion and appearance. % To address this, we introduce a large-scale dataset of triplets, where each triplet consists of a reference effect video, an input image or video, and a corresponding output video depicting the transferred effect. Creating this data is non-trivial, especially the video-to-video effect triplets, which do not exist naturally. To generate these, we propose a scalable automated pipeline that creates high-quality paired videos designed to preserve the input’s motion and structure while transforming it based on some fixed, repeatable effect. We then augment this data with image-to-video effects derived from LoRA adapters and code-based temporal effects generated through programmatic composition. Building on our new dataset, we train our reference-conditioned model using recent text-to-video backbones. Experimental results demonstrate that RefVFX produces visually consistent and temporally coherent edits, generalizes across unseen effect categories, and outperforms prompt-only baselines in both quantitative metrics and human preference. See our website $\href{https://tuningfreevisualeffects-maker.github.io/Tuning-free-Visual-Effect-Transfer-across-Videos-Project-Page/}{at\ this\ URL}$.

Method: Uses a bank of independently trained GANs as priors (one per intrinsic component), jointly inverts their latent codes to reproduce input image, and allows fine-tuning of generators during real image inference.

Result: Successfully decomposes both synthetic and real images, achieves excellent generalization on real images using only synthetic training data, and demonstrates modularity with various forward imaging models.

Conclusion: JoIN provides stronger, more disentangled priors through independent GAN training, effectively addresses cross-contamination and Sim-to-Real issues, and offers a flexible framework for intrinsic image decomposition.

Abstract: Intrinsic Image Decomposition (IID) is a challenging inverse problem that seeks to decompose a natural image into its underlying intrinsic components such as albedo and shading. While recent image decomposition methods rely on learning-based priors on these components, they often suffer from component cross-contamination owing to joint training of priors; or from Sim-to-Real gap since the priors trained on synthetic data are kept frozen during the inference on real images. In this work, we propose to solve the intrinsic image decomposition problem using a bank of Generative Adversarial Networks (GANs) as priors where each GAN is independently trained only on a single intrinsic component, providing stronger and more disentangled priors. At the core of our approach is the idea that the latent space of a GAN is a well-suited optimization domain to solve inverse problems. Given an input image, we propose to jointly invert the latent codes of a set of GANs and combine their outputs to reproduce the input. Contrary to all existing GAN inversion methods that are limited to inverting only a single GAN, our proposed approach, JoIN, is able to jointly invert multiple GANs using only a single image as supervision while still maintaining distribution priors of each intrinsic component. We show that our approach is modular, allowing various forward imaging models, and that it can successfully decompose both synthetic and real images. Further, taking inspiration from existing GAN inversion approaches, we allow for careful fine-tuning of the generator priors during the inference on real images. This way, our method is able to achieve excellent generalization on real images even though it uses only synthetic data to train the GAN priors. We demonstrate the success of our approach through exhaustive qualitative and quantitative evaluations and ablation studies on various datasets.

Method: The research uses a specialized dataset of 923 images (438 healthy, 485 bleached) collected from Flickr via API, resized to max 300 pixels. It employs convolutional neural networks (CNNs), specifically ResNet architecture, comparing models trained from scratch versus pretrained models for classification.

Result: The ResNet model trained from scratch outperformed pretrained models in both precision and accuracy for distinguishing between healthy and bleached corals.

Conclusion: The developed accurate classification model provides substantial benefits for researchers and marine biologists, enabling better understanding of coral reef health and supporting conservation efforts through effective monitoring of environmental changes in coral reef ecosystems.

Abstract: The rich biodiversity of coral reefs in Indonesian waters represents a valuable asset that must be preserved. Rapid climate change and uncontrolled human activities have caused significant degradation of coral reef ecosystems, including coral bleaching, which is a critical indicator of declining reef health. Therefore, this study aims to develop an accurate classification model to distinguish between healthy corals and bleached corals. This research utilizes a specialized dataset consisting of 923 images collected from Flickr using the Flickr API. The dataset comprises two distinct classes: healthy corals (438 images) and bleached corals (485 images). All images were resized so that the maximum width or height does not exceed 300 pixels, ensuring consistent image dimensions across the dataset. The proposed approach employs machine learning techniques, particularly convolutional neural networks (CNNs), to identify and differentiate visual patterns associated with healthy and bleached corals. The dataset can be used to train and evaluate various classification models in order to achieve optimal performance. Using the ResNet architecture, the results indicate that a ResNet model trained from scratch outperforms pretrained models in terms of both precision and accuracy. The successful development of an accurate classification model provides substantial benefits for researchers and marine biologists by enabling a deeper understanding of coral reef health. Furthermore, these models can be applied to monitor environmental changes in coral reef ecosystems, thereby contributing meaningfully to conservation and restoration efforts that are vital to sustaining marine life.

Method: Generates multiple alternative modifications for each anomaly that make it appear normal to the detector. Each modification captures different concepts of anomalousness and is trained to be perceived as normal by the anomaly detector.

Result: The method provides high-quality semantic explanations across various image datasets when applied to state-of-the-art detectors, allowing users to explore “what-if scenarios” for anomaly understanding.

Conclusion: The proposed method successfully addresses the interpretability challenge in deep learning-based anomaly detection by generating diverse, semantically meaningful explanations of why instances are flagged as anomalous.

Abstract: Deep learning-based methods have achieved a breakthrough in image anomaly detection, but their complexity introduces a considerable challenge to understanding why an instance is predicted to be anomalous. We introduce a novel explanation method that generates multiple alternative modifications for each anomaly, capturing diverse concepts of anomalousness. Each modification is trained to be perceived as normal by the anomaly detector. The method provides a semantic explanation of the mechanism that triggered the detector, allowing users to explore ``what-if scenarios.’’ Qualitative and quantitative analyses across various image datasets demonstrate that applying this method to state-of-the-art detectors provides high-quality semantic explanations.

Method: Proposes hyperspectral point object detection as a new task framework and introduces SpecDETR, a specialized Transformer network with self-excited subpixel-scale attention modules to directly extract deep spatial-spectral joint features from hyperspectral cubes without pre-trained backbones.

Result: SpecDETR outperforms state-of-the-art visual object detection networks and HTD methods on the proposed SPOD benchmark, demonstrating superior performance for hyperspectral point object detection.

Conclusion: The paper successfully rethinks HTD from a spatial-spectral synergistic perspective, introduces a new detection framework, and provides a specialized network (SpecDETR) that eliminates dependency on pre-trained backbones while achieving superior performance.

Abstract: Hyperspectral target detection (HTD) aims to identify specific materials based on spectral information in hyperspectral imagery and can detect extremely small-sized objects, some of which occupy a smaller than one-pixel area. However, existing HTD methods are developed based on per-pixel binary classification, neglecting the three-dimensional cube structure of hyperspectral images (HSIs) that integrates both spatial and spectral dimensions. The synergistic existence of spatial and spectral features in HSIs enable objects to simultaneously exhibit both, yet the per-pixel HTD framework limits the joint expression of these features. In this paper, we rethink HTD from the perspective of spatial-spectral synergistic representation and propose hyperspectral point object detection as an innovative task framework. We introduce SpecDETR, the first specialized network for hyperspectral multi-class point object detection, which eliminates dependence on pre-trained backbone networks commonly required by vision-based object detectors. SpecDETR uses a multi-layer Transformer encoder with self-excited subpixel-scale attention modules to directly extract deep spatial-spectral joint features from hyperspectral cubes. We develop a simulated hyperspectral point object detection benchmark termed SPOD, and for the first time, evaluate and compare the performance of visual object detection networks and HTD methods on hyperspectral point object detection. Extensive experiments demonstrate that our proposed SpecDETR outperforms SOTA visual object detection networks and HTD methods. Our code and dataset are available at https://github.com/ZhaoxuLi123/SpecDETR.

Method: Introduces Multi-Attribute Composition (MAC) dataset with 22,838 images and 17,627 compositions. Proposes MVP-Integrator method that disentangles semantic primitives and performs effective visual-primitive association for multi-attribute CZSL.

Result: MAC shows complex relationships: each attribute type linked to average 82.2 object types, each object type associated with 31.4 attribute types. MVP-Integrator significantly outperforms existing CZSL methods on MAC with improved inference efficiency.

Conclusion: The work establishes a more realistic and challenging benchmark for CZSL, requiring deeper semantic understanding and advanced attribute associations. The proposed approach effectively addresses limitations of single-attribute CZSL.

Abstract: Compositional Zero-Shot Learning (CZSL) aims to learn semantic primitives (attributes and objects) from seen compositions and recognize unseen attribute-object compositions. Existing CZSL datasets focus on single attributes, neglecting the fact that objects naturally exhibit multiple interrelated attributes. Their narrow attribute scope and single attribute labeling introduce annotation biases, misleading the learning of attributes and causing inaccurate evaluation. To address these issues, we introduce the Multi-Attribute Composition (MAC) dataset, encompassing 22,838 images and 17,627 compositions with comprehensive and representative attribute annotations. MAC shows complex relationship between attributes and objects, with each attribute type linked to an average of 82.2 object types, and each object type associated with 31.4 attribute types. Based on MAC, we propose multi-attribute compositional zero-shot learning that requires deeper semantic understanding and advanced attribute associations, establishing a more realistic and challenging benchmark for CZSL. We also propose Multi-attribute Visual-Primitive Integrator (MVP-Integrator), a robust baseline for multi-attribute CZSL, which disentangles semantic primitives and performs effective visual-primitive association. Experimental results demonstrate that MVP-Integrator significantly outperforms existing CZSL methods on MAC with improved inference efficiency.

Method: Two key components: (1) SliceRestore adapter reconstructs sliced patches to original form using down-up-sampling and convolution layers to extract global and local features; (2) Self-Mining Sampler compresses vision tokens while preserving original context and positional information.

Result: Superior performance on existing benchmarks and new EntityGrid-QA benchmark (edge-related and position-related tasks), especially on document-oriented tasks, establishing new standards for high-resolution input processing.

Conclusion: HiRes-LLaVA efficiently handles any size of high-resolution input without losing contextual and geometric information, overcoming fragmentation issues of sliding window approaches while reducing training overhead.

Abstract: High-resolution inputs enable Large Vision-Language Models (LVLMs) to discern finer visual details, enhancing their comprehension capabilities. To reduce the training and computation costs caused by high-resolution input, one promising direction is to use sliding windows to slice the input into uniform patches, each matching the input size of the well-trained vision encoder. Although efficient, this slicing strategy leads to the fragmentation of original input, i.e., the continuity of contextual information and spatial geometry is lost across patches, adversely affecting performance in cross-patch context perception and position-specific tasks. To overcome these shortcomings, we introduce HiRes-LLaVA, a novel framework designed to efficiently process any size of high-resolution input without altering the original contextual and geometric information. HiRes-LLaVA comprises two innovative components: (i) a SliceRestore adapter that reconstructs sliced patches into their original form, efficiently extracting both global and local features via down-up-sampling and convolution layers, and (ii) a Self-Mining Sampler to compresses the vision tokens based on themselves, preserving the original context and positional information while reducing training overhead. To assess the ability of handling context fragmentation, we construct a new benchmark, EntityGrid-QA, consisting of edge-related and position-related tasks. Our comprehensive experiments demonstrate the superiority of HiRes-LLaVA on both existing public benchmarks and on EntityGrid-QA, particularly on document-oriented tasks, establishing new standards for handling high-resolution inputs.

Method: Proposes ORB-guided visual odometry with selective online adaptation. Uses ORB features for learning-based ego-motion estimation, introduces cross-attention mechanism to enhance PoseNet explainability, and implements selective online adaptation for domain adaptation.

Result: Outperforms previous state-of-the-art deep visual odometry methods on KITTI and vKITTI datasets in terms of ego-motion accuracy and generalizability.

Conclusion: ORB-SfMLearner successfully addresses accuracy and generalizability limitations in deep visual odometry through ORB feature guidance, cross-attention mechanisms, and selective online adaptation, enabling broader application.

Abstract: Deep visual odometry, despite extensive research, still faces limitations in accuracy and generalizability that prevent its broader application. To address these challenges, we propose an Oriented FAST and Rotated BRIEF (ORB)-guided visual odometry with selective online adaptation named ORB-SfMLearner. We present a novel use of ORB features for learning-based ego-motion estimation, leading to more robust and accurate results. We also introduce the cross-attention mechanism to enhance the explainability of PoseNet and have revealed that driving direction of the vehicle can be explained through the attention weights. To improve generalizability, our selective online adaptation allows the network to rapidly and selectively adjust to the optimal parameters across different domains. Experimental results on KITTI and vKITTI datasets show that our method outperforms previous state-of-the-art deep visual odometry methods in terms of ego-motion accuracy and generalizability. Code is available at https://github.com/PeaceNeil/ORB-SfMLearner

Method: Uses Dynamic Attention Transformer (DATrans) to simulate central difference convolutions for gradient features, plus a Global Feature Extraction Module (GFEM) to balance local details with global context.

Result: Demonstrates effective performance compared to state-of-the-art approaches in infrared small target detection.

Conclusion: DATransNet successfully addresses ISTD challenges by extracting detailed gradient features while maintaining global perspective, outperforming existing methods.

Abstract: Infrared small target detection (ISTD) is widely used in civilian and military applications. However, ISTD encounters several challenges, including the tendency for small and dim targets to be obscured by complex backgrounds. To address this issue, we propose the Dynamic Attention Transformer Network (DATransNet), which aims to extract and preserve detailed information vital for small targets. DATransNet employs the Dynamic Attention Transformer (DATrans), simulating central difference convolutions (CDC) to extract gradient features. Furthermore, we propose a global feature extraction module (GFEM) that offers a comprehensive perspective to prevent the network from focusing solely on details while neglecting the global information. We compare the network with state-of-the-art (SOTA) approaches and demonstrate that our method performs effectively. Our source code is available at https://github.com/greekinRoma/DATransNet.

Method: TRASE learns a 4D segmentation feature field in a weakly-supervised manner using soft-mined contrastive learning guided by SAM masks. The resulting feature space is semantically coherent and well-separated, with final object-level segmentation obtained via unsupervised clustering.

Result: TRASE achieves state-of-the-art segmentation performance on five dynamic benchmarks from unseen viewpoints and demonstrates effectiveness across various interactive editing tasks like object removal, composition, and style transfer.

Conclusion: TRASE enables fast editing of dynamic scenes by directly manipulating scene Gaussians, providing a practical solution for interactive applications in XR and autonomous driving through tracking-free 4D segmentation.

Abstract: Understanding dynamic 3D scenes is crucial for extended reality (XR) and autonomous driving. Incorporating semantic information into 3D reconstruction enables holistic scene representations, unlocking immersive and interactive applications. To this end, we introduce TRASE, a novel tracking-free 4D segmentation method for dynamic scene understanding. TRASE learns a 4D segmentation feature field in a weakly-supervised manner, leveraging a soft-mined contrastive learning objective guided by SAM masks. The resulting feature space is semantically coherent and well-separated, and final object-level segmentation is obtained via unsupervised clustering. This enables fast editing, such as object removal, composition, and style transfer, by directly manipulating the scene’s Gaussians. We evaluate TRASE on five dynamic benchmarks, demonstrating state-of-the-art segmentation performance from unseen viewpoints and its effectiveness across various interactive editing tasks. Our project page is available at: https://yunjinli.github.io/project-sadg/

Method: Uses neural Gaussians to learn instance/hierarchical segmentation from multi-view images with 2D masks, creates sparse Super-Gaussians, and distills 2D language features into 3D space for efficient rendering.

Result: Outperforms prior works on both open-vocabulary object localization and semantic segmentation tasks while avoiding extreme GPU memory increases.

Conclusion: SuperGSeg enables cohesive, context-aware scene representation through disentangled segmentation and language field distillation, advancing 3D Gaussian Splatting for scene understanding.

Abstract: 3D Gaussian Splatting has recently gained traction for its efficient training and real-time rendering. While the vanilla Gaussian Splatting representation is mainly designed for view synthesis, more recent works investigated how to extend it with scene understanding and language features. However, existing methods lack a detailed comprehension of scenes, limiting their ability to segment and interpret complex structures. To this end, We introduce SuperGSeg, a novel approach that fosters cohesive, context-aware scene representation by disentangling segmentation and language field distillation. SuperGSeg first employs neural Gaussians to learn instance and hierarchical segmentation features from multi-view images with the aid of off-the-shelf 2D masks. These features are then leveraged to create a sparse set of what we call Super-Gaussians. Super-Gaussians facilitate the distillation of 2D language features into 3D space. Through Super-Gaussians, our method enables high-dimensional language feature rendering without extreme increases in GPU memory. Extensive experiments demonstrate that SuperGSeg outperforms prior works on both open-vocabulary object localization and semantic segmentation tasks.

Method: Introduces GS Tokenizer to generate serialized gaussian tokens from 3DGS, processes them through transformer layers pre-initialized with point cloud model weights to get 3DGS embeddings. Uses contrastive loss between 3DGS and CLIP visual-text embeddings, plus image voting loss for gradient optimization guidance. Develops efficient triplet generation (3DGS, images, text) for unified multimodal representation learning.

Result: CLIP-GS demonstrates versatility and outperforms point cloud-based models on various 3D tasks including multimodal retrieval, zero-shot, and few-shot classification, leveraging well-aligned multimodal representations.

Conclusion: The proposed CLIP-GS framework successfully bridges 3D Gaussian Splatting with multimodal learning, overcoming limitations of point cloud-based approaches and achieving superior performance across multiple 3D understanding tasks through unified representation learning.

Abstract: Recent works in 3D multimodal learning have made remarkable progress. However, typically 3D multimodal models are only capable of handling point clouds. Compared to the emerging 3D representation technique, 3D Gaussian Splatting (3DGS), the spatially sparse point cloud cannot depict the texture information of 3D objects, resulting in inferior reconstruction capabilities. This limitation constrains the potential of point cloud-based 3D multimodal representation learning. In this paper, we present CLIP-GS, a novel multimodal representation learning framework grounded in 3DGS. We introduce the GS Tokenizer to generate serialized gaussian tokens, which are then processed through transformer layers pre-initialized with weights from point cloud models, resulting in the 3DGS embeddings. CLIP-GS leverages contrastive loss between 3DGS and the visual-text embeddings of CLIP, and we introduce an image voting loss to guide the directionality and convergence of gradient optimization. Furthermore, we develop an efficient way to generate triplets of 3DGS, images, and text, facilitating CLIP-GS in learning unified multimodal representations. Leveraging the well-aligned multimodal representations, CLIP-GS demonstrates versatility and outperforms point cloud-based models on various 3D tasks, including multimodal retrieval, zero-shot, and few-shot classification.

Method: GS-LoRA uses Low-Rank Adaptation (LoRA) modules to fine-tune FFN layers in Transformer blocks for each forgetting task independently, with group sparse regularization for automatic selection of specific LoRA groups. GS-LoRA++ extends this with prototype supervision: moving logits away from forgotten class prototypes and pulling them closer to remaining class prototypes.

Result: Extensive experiments on face recognition, object detection, and image classification show the method successfully forgets specific classes with minimal impact on other classes.

Conclusion: The proposed GS-LoRA and GS-LoRA++ effectively address continual forgetting challenges in vision models, providing practical solutions for real-world scenarios where selective information needs to be continuously removed while maintaining model performance on remaining knowledge.

Abstract: For privacy and security concerns, the need to erase unwanted information from pre-trained vision models is becoming evident nowadays. In real-world scenarios, erasure requests originate at any time from both users and model owners, and these requests usually form a sequence. Therefore, under such a setting, selective information is expected to be continuously removed from a pre-trained model while maintaining the rest. We define this problem as continual forgetting and identify three key challenges. (i) For unwanted knowledge, efficient and effective deleting is crucial. (ii) For remaining knowledge, the impact brought by the forgetting procedure should be minimal. (iii) In real-world scenarios, the training samples may be scarce or partially missing during the process of forgetting. To address them, we first propose Group Sparse LoRA (GS-LoRA). Specifically, towards (i), we introduce Low-Rank Adaptation (LoRA) modules to fine-tune the Feed-Forward Network (FFN) layers in Transformer blocks for each forgetting task independently, and towards (ii), a simple group sparse regularization is adopted, enabling automatic selection of specific LoRA groups and zeroing out the others. To further extend GS-LoRA to more practical scenarios, we incorporate prototype information as additional supervision and introduce a more practical approach, GS-LoRA++. For each forgotten class, we move the logits away from its original prototype. For the remaining classes, we pull the logits closer to their respective prototypes. We conduct extensive experiments on face recognition, object detection, and image classification and demonstrate that our method manages to forget specific classes with minimal impact on other classes. Codes have been released on https://github.com/bjzhb666/GS-LoRA.

Method: Created MomentSeeker benchmark with: 1) Long videos averaging 1200+ seconds from diverse domains (movie, anomaly, egocentric, sports), 2) Three-level scenarios (global, event, object) covering various tasks, 3) Multiple query types (text, image, video-conditioned).

Result: Comprehensive experiments show significant challenges in long-video moment retrieval for both generation-based (MLLMs) and retrieval-based approaches, with accuracy and efficiency issues persisting despite latest advancements.

Conclusion: MomentSeeker addresses the gap in evaluating long-video moment retrieval, reveals current limitations, and provides a public benchmark to advance research in this crucial area of video understanding.

Abstract: Accurately locating key moments within long videos is crucial for solving long video understanding (LVU) tasks. However, existing benchmarks are either severely limited in terms of video length and task diversity, or they focus solely on the end-to-end LVU performance, making them inappropriate for evaluating whether key moments can be accurately accessed. To address this challenge, we propose MomentSeeker, a novel benchmark for long-video moment retrieval (LMVR), distinguished by the following features. First, it is created based on long and diverse videos, averaging over 1200 seconds in duration and collected from various domains, e.g., movie, anomaly, egocentric, and sports. Second, it covers a variety of real-world scenarios in three levels: global-level, event-level, object-level, covering common tasks like action recognition, object localization, and causal reasoning, etc. Third, it incorporates rich forms of queries, including text-only queries, image-conditioned queries, and video-conditioned queries. On top of MomentSeeker, we conduct comprehensive experiments for both generation-based approaches (directly using MLLMs) and retrieval-based approaches (leveraging video retrievers). Our results reveal the significant challenges in long-video moment retrieval in terms of accuracy and efficiency, despite improvements from the latest long-video MLLMs and task-specific fine-tuning. We have publicly released MomentSeeker(https://yhy-2000.github.io/MomentSeeker/) to facilitate future research in this area.

Method: FlexVAR uses a flexible visual autoregressive paradigm with ground-truth prediction at each step, trained solely on low-resolution images (≤256px). This allows each step to independently produce plausible images and adapt to different generation requirements.

Result: The 1.0B model outperforms VAR counterparts on ImageNet 256×256, achieves 2.08 FID with 13-step zero-shot transfer (beating AiM/VAR and LDM/DiT), and shows competitive zero-shot performance on ImageNet 512×512 compared to larger supervised models.

Conclusion: FlexVAR demonstrates that moving beyond residual prediction to ground-truth prediction enables more flexible and efficient visual autoregressive modeling with strong performance across resolutions and tasks, even with limited training data.

Abstract: This work challenges the residual prediction paradigm in visual autoregressive modeling and presents FlexVAR, a new Flexible Visual AutoRegressive image generation paradigm. FlexVAR facilitates autoregressive learning with ground-truth prediction, enabling each step to independently produce plausible images. This simple, intuitive approach swiftly learns visual distributions and makes the generation process more flexible and adaptable. Trained solely on low-resolution images ($\leq$ 256px), FlexVAR can: (1) Generate images of various resolutions and aspect ratios, even exceeding the resolution of the training images. (2) Support various image-to-image tasks, including image refinement, in/out-painting, and image expansion. (3) Adapt to various autoregressive steps, allowing for faster inference with fewer steps or enhancing image quality with more steps. Our 1.0B model outperforms its VAR counterpart on the ImageNet 256$\times$256 benchmark. Moreover, when zero-shot transfer the image generation process with 13 steps, the performance further improves to 2.08 FID, outperforming state-of-the-art autoregressive models AiM/VAR by 0.25/0.28 FID and popular diffusion models LDM/DiT by 1.52/0.19 FID, respectively. When transferring our 1.0B model to the ImageNet 512$\times$512 benchmark in a zero-shot manner, FlexVAR achieves competitive results compared to the VAR 2.3B model, which is a fully supervised model trained at 512$\times$512 resolution.

Method: Proposes Gated-Mechanism Mixture-of-Experts (GM-MoE) with dynamic gated weight conditioning network and three specialized sub-expert networks, using a gating mechanism to dynamically adjust weights for different data domains, plus local-global feature fusion for multi-scale feature capture.

Result: Achieves superior generalization compared to 25 approaches, reaching state-of-the-art performance on PSNR on 5 benchmarks and SSIM on 4 benchmarks.

Conclusion: GM-MoE is the first mixture-of-experts framework for low-light enhancement that effectively addresses generalization limitations and achieves top performance across multiple evaluation metrics.

Abstract: Low-light enhancement has wide applications in autonomous driving, 3D reconstruction, remote sensing, surveillance, and so on, which can significantly improve information utilization. However, most existing methods lack generalization and are limited to specific tasks such as image recovery. To address these issues, we propose Gated-Mechanism Mixture-of-Experts (GM-MoE), the first framework to introduce a mixture-of-experts network for low-light image enhancement. GM-MoE comprises a dynamic gated weight conditioning network and three sub-expert networks, each specializing in a distinct enhancement task. Combining a self-designed gated mechanism that dynamically adjusts the weights of the sub-expert networks for different data domains. Additionally, we integrate local and global feature fusion within sub-expert networks to enhance image quality by capturing multi-scale features. Experimental results demonstrate that the GM-MoE achieves superior generalization with respect to 25 compared approaches, reaching state-of-the-art performance on PSNR on 5 benchmarks and SSIM on 4 benchmarks, respectively.

Method: Unsafe Weights Manipulation (UWM) uses calibration sets to compare activations between safe/unsafe content, identifies key parameters for unsafe processing, and manipulates them via negation without training.

Result: UWM achieves best tradeoff between safety and knowledge preservation, improving safety on unsafe queries while outperforming training-based methods on safe inputs.

Conclusion: UWM provides effective non-training approach to VLM safety that addresses limitations of current methods by maintaining performance on safe content while improving safety.

Abstract: Vision-language models (VLMs) often inherit the biases and unsafe associations present within their large-scale training dataset. While recent approaches mitigate unsafe behaviors, their evaluation focuses on how safe the model is on unsafe inputs, ignoring potential shortcomings on safe ones. In this paper, we first revise safety evaluation by introducing SafeGround, a new set of metrics that evaluate safety at different levels of granularity. With this metric, we uncover a surprising issue of training-based methods: they make the model less safe on safe inputs. From this finding, we take a different direction and explore whether it is possible to make a model safer without training, introducing Unsafe Weights Manipulation (UWM). UWM uses a calibration set of safe and unsafe instances to compare activations between safe and unsafe content, identifying the most important parameters for processing the latter. Their values are then manipulated via negation. Experiments show that UWM achieves the best tradeoff between safety and knowledge preservation, consistently improving VLMs on unsafe queries while outperforming even training-based state-of-the-art methods on safe ones.

Method: Organizes 3D Gaussians into groups corresponding to Laplacian pyramid subbands of input images. Each group trained with spatial frequency regularization to confine to target frequencies. Higher-frequency bands use signed residual colors for fine details. Progressive coarse-to-fine training schedule stabilizes decomposition.

Result: Achieves state-of-the-art reconstruction quality and rendering speed among all LOD-capable methods. Enables dynamic level-of-detail rendering, progressive streaming, foveated rendering, promptable 3D focus, and artistic filtering.

Conclusion: Frequency-aware decomposition of 3D Gaussians improves interpretability and enables practical applications like LOD rendering and progressive streaming while maintaining high quality and speed.

Abstract: 3D Gaussian Splatting (3D-GS) enables efficient novel view synthesis, but treats all frequencies uniformly, making it difficult to separate coarse structure from fine detail. Recent works have started to exploit frequency signals, but lack explicit frequency decomposition of the 3D representation itself. We propose a frequency-aware decomposition that organizes 3D Gaussians into groups corresponding to Laplacian-pyramid subbands of the input images. Each group is trained with spatial frequency regularization to confine it to its target frequency, while higher-frequency bands use signed residual colors to capture fine details that may be missed by lower-frequency reconstructions. A progressive coarse-to-fine training schedule stabilizes the decomposition. Our method achieves state-of-the-art reconstruction quality and rendering speed among all LOD-capable methods. In addition to improved interpretability, our method enables dynamic level-of-detail rendering, progressive streaming, foveated rendering, promptable 3D focus, and artistic filtering. Our code will be made publicly available.

Method: Introduces StarFlow framework that uses vision-language models to generate structured workflows from sketches. Curates a diverse dataset of workflow diagrams (synthetic, manually annotated, real-world samples) for training and evaluation. Finetunes and benchmarks multiple VLMs, conducting ablation studies to analyze approach strengths and limitations.

Result: Finetuning significantly enhances structured workflow generation, with the approach outperforming large vision-language models on this specific task.

Conclusion: The StarFlow framework demonstrates that vision-language models can effectively automate workflow generation from visual inputs, with finetuning being crucial for achieving good performance on this structured output task.

Abstract: Workflows are a fundamental component of automation in enterprise platforms, enabling the orchestration of tasks, data processing, and system integrations. Despite being widely used, building workflows can be complex, often requiring manual configuration through low-code platforms or visual programming tools. To simplify this process, we explore the use of generative foundation models, particularly vision-language models (VLMs), to automatically generate structured workflows from visual inputs. Translating hand-drawn sketches or computer-generated diagrams into executable workflows is challenging due to the ambiguity of free-form drawings, variations in diagram styles, and the difficulty of inferring execution logic from visual elements. To address this, we introduce StarFlow, a framework for generating structured workflow outputs from sketches using vision-language models. We curate a diverse dataset of workflow diagrams – including synthetic, manually annotated, and real-world samples – to enable robust training and evaluation. We finetune and benchmark multiple vision-language models, conducting a series of ablation studies to analyze the strengths and limitations of our approach. Our results show that finetuning significantly enhances structured workflow generation, outperforming large vision-language models on this task.

Method: Proposes Dynamic Diffusion Transformer (DyDiT) that dynamically adjusts computation along timestep and spatial dimensions, with extended DyDiT++ supporting flow matching, video/text-to-image generation, and parameter-efficient training via timestep-based dynamic LoRA (TD-LoRA).

Result: Reduces FLOPs of DiT-XL by 51% with <3% additional fine-tuning, achieves 1.73x realistic hardware speedup, and maintains competitive FID score of 2.07 on ImageNet. Works across DiT, SiT, Latte, and FLUX models.

Conclusion: DyDiT++ effectively addresses computational inefficiency in diffusion models through dynamic computation, extends to flow matching and complex generation tasks, and enables parameter-efficient training, making advanced visual generation more accessible.

Abstract: Diffusion Transformer (DiT), an emerging diffusion model for visual generation, has demonstrated superior performance but suffers from substantial computational costs. Our investigations reveal that these costs primarily stem from the static inference paradigm, which inevitably introduces redundant computation in certain diffusion timesteps and spatial regions. To overcome this inefficiency, we propose Dynamic Diffusion Transformer (DyDiT), an architecture that dynamically adjusts its computation along both timestep and spatial dimensions. Building on these designs, we present an extended version, DyDiT++, with improvements in three key aspects. First, it extends the generation mechanism of DyDiT beyond diffusion to flow matching, demonstrating that our method can also accelerate flow-matching-based generation, enhancing its versatility. Furthermore, we enhance DyDiT to tackle more complex visual generation tasks, including video generation and text-to-image generation, thereby broadening its real-world applications. Finally, to address the high cost of full fine-tuning and democratize technology access, we investigate the feasibility of training DyDiT in a parameter-efficient manner and introduce timestep-based dynamic LoRA (TD-LoRA). Extensive experiments on diverse visual generation models, including DiT, SiT, Latte, and FLUX, demonstrate the effectiveness of DyDiT++. Remarkably, with <3% additional fine-tuning iterations, our approach reduces the FLOPs of DiT-XL by 51%, yielding 1.73x realistic speedup on hardware, and achieves a competitive FID score of 2.07 on ImageNet. The code is available at https://github.com/alibaba-damo-academy/DyDiT.

Method: CLOC uses margin-based contrastive learning with a novel multi-margin n-pair loss (MMNP) that learns ordered representations by optimizing multiple margins, allowing flexible decision boundaries across key adjacent categories and smooth transitions between classes.

Result: CLOC outperforms existing ordinal classification methods on five real-world image datasets (Adience, Historical Colour Image Dating, Knee Osteoarthritis, Indian Diabetic Retinopathy Image, Breast Carcinoma Subtyping) and one synthetic dataset simulating clinical decision bias.

Conclusion: CLOC provides interpretable and controllable ordered representations that align with clinical needs, reduces overfitting to training biases, and effectively handles varying importance of different classification boundaries in ordinal tasks.

Abstract: In ordinal classification, misclassifying neighboring ranks is common, yet the consequences of these errors are not the same. For example, misclassifying benign tumor categories is less consequential, compared to an error at the pre-cancerous to cancerous threshold, which could profoundly influence treatment choices. Despite this, existing ordinal classification methods do not account for the varying importance of these margins, treating all neighboring classes as equally significant. To address this limitation, we propose CLOC, a new margin-based contrastive learning method for ordinal classification that learns an ordered representation based on the optimization of multiple margins with a novel multi-margin n-pair loss (MMNP). CLOC enables flexible decision boundaries across key adjacent categories, facilitating smooth transitions between classes and reducing the risk of overfitting to biases present in the training data. We provide empirical discussion regarding the properties of MMNP and show experimental results on five real-world image datasets (Adience, Historical Colour Image Dating, Knee Osteoarthritis, Indian Diabetic Retinopathy Image, and Breast Carcinoma Subtyping) and one synthetic dataset simulating clinical decision bias. Our results demonstrate that CLOC outperforms existing ordinal classification methods and show the interpretability and controllability of CLOC in learning meaningful, ordered representations that align with clinical and practical needs.

Method: Orthogonal Residual Update decomposes the module’s output relative to the input stream and adds only the component orthogonal to this stream, guiding modules to contribute primarily new representational directions.

Result: The orthogonal update strategy improves generalization accuracy and training stability across diverse architectures (ResNetV2, Vision Transformers) and datasets (CIFARs, TinyImageNet, ImageNet-1k), achieving +3.78 pp top-1 accuracy gain for ViT-B on ImageNet-1k.

Conclusion: Orthogonal residual updates enable richer feature learning by encouraging modules to contribute novel representational directions, leading to improved generalization and more efficient training across various neural network architectures.

Abstract: Residual connections are pivotal for deep neural networks, enabling greater depth by mitigating vanishing gradients. However, in standard residual updates, the module’s output is directly added to the input stream. This can lead to updates that predominantly reinforce or modulate the existing stream direction, potentially underutilizing the module’s capacity for learning entirely novel features. In this work, we introduce Orthogonal Residual Update: we decompose the module’s output relative to the input stream and add only the component orthogonal to this stream. This design aims to guide modules to contribute primarily new representational directions, fostering richer feature learning while promoting more efficient training. We demonstrate that our orthogonal update strategy improves generalization accuracy and training stability across diverse architectures (ResNetV2, Vision Transformers) and datasets (CIFARs, TinyImageNet, ImageNet-1k), achieving, for instance, a +3.78 pp top-1 accuracy gain for ViT-B on ImageNet-1k.

Method: TEA (Targeted Edge-informed Attack) uses edge information from the target image to carefully perturb it, producing an adversarial image that is closer to the source image while still achieving the desired target classification.

Result: TEA consistently outperforms current state-of-the-art methods across different models in low query settings, using nearly 70% fewer queries. It also provides improved target initialization for established geometry-based attacks.

Conclusion: TEA demonstrates that incorporating edge information from target images enables more efficient targeted adversarial attacks in black-box settings with limited queries, offering practical advantages for real-world applications.

Abstract: Deep neural networks for image classification remain vulnerable to adversarial examples – small, imperceptible perturbations that induce misclassifications. In black-box settings, where only the final prediction is accessible, crafting targeted attacks that aim to misclassify into a specific target class is particularly challenging due to narrow decision regions. Current state-of-the-art methods often exploit the geometric properties of the decision boundary separating a source image and a target image rather than incorporating information from the images themselves. In contrast, we propose Targeted Edge-informed Attack (TEA), a novel attack that utilizes edge information from the target image to carefully perturb it, thereby producing an adversarial image that is closer to the source image while still achieving the desired target classification. Our approach consistently outperforms current state-of-the-art methods across different models in low query settings (nearly 70% fewer queries are used), a scenario especially relevant in real-world applications with limited queries and black-box access. Furthermore, by efficiently generating a suitable adversarial example, TEA provides an improved target initialization for established geometry-based attacks.

Method: Built survival models using handcrafted radiomic and deep features from multiple thoracic regions (lung, tumor, mediastinal nodes, coronary arteries, CAC scores) from 876 patients across 5 centres. Used ComBat, RKN, and RKN-ComBat for harmonization. Employed ROI-level and ensemble strategies with regularized Cox models. Assessed performance via C-index, t-AUC, hazard ratios, and used SHAP for feature interpretation.

Result: TNM staging showed baseline prognostic value (C-index=0.67). Clinical+tumor texture radiomics with ComBat achieved C-index=0.76. FM deep features from 50-voxel cubes also showed C-index=0.76. Ensemble model combining multiple features achieved C-index=0.71. Consensus analysis identified high-confidence patient subset with 5-year t-AUC=0.92, sensitivity=96.8%, specificity=70.0%.

Conclusion: Harmonization and multi-region feature integration enhance survival prediction in NSCLC patients using CT imaging, supporting individualized risk stratification in multicentre settings.

Abstract: Purpose: This study evaluates the impact of harmonization and multi-region feature integration on survival prediction in non-small cell lung cancer (NSCLC) patients. We assess the prognostic utility of handcrafted radiomics and pretrained deep features from thoracic CT images, integrating them with clinical data using a multicentre dataset. Methods: Survival models were built using handcrafted radiomic and deep features from lung, tumor, mediastinal nodes, coronary arteries, and coronary artery calcium (CAC) scores from 876 patients across five centres. CT features were harmonized using ComBat, reconstruction kernel normalization (RKN), and RKN-ComBat. Models were constructed at the region of interest (ROI) level and through ensemble strategies. Regularized Cox models estimated overall survival, with performance assessed via the concordance index (C-index), 5-year time-dependent area under the curve (t-AUC), and hazard ratios. SHAP values interpreted feature contributions, while consensus analysis categorized predicted survival probabilities at fixed time points. Results: TNM staging showed prognostic value (C-index = 0.67; hazard ratio = 2.70; t-AUC = 0.85). The clinical and tumor texture radiomics model with ComBat yielded high performance (C-index = 0.76; t-AUC = 0.88). FM deep features from 50 voxel cubes also showed predictive value (C-index = 0.76; t-AUC = 0.89). An ensemble model combining tumor, lung, mediastinal node, CAC, and FM features achieved a C-index of 0.71 and t-AUC of 0.79. Consensus analysis identified a high-confidence patient subset, resulting in a model with a 5-year t-AUC of 0.92, sensitivity of 96.8%, and specificity of 70.0%. Conclusion: Harmonization and multi-region feature integration enhance survival prediction in NSCLC patients using CT imaging, supporting individualized risk stratification in multicentre settings.

Method: 1) Voxel-based progressive 3D Gaussian mapping with multiple submaps for compact scene representation; 2) 2D-3D fusion camera tracking for robust pose estimation; 3) 2D-3D Gaussian loop closure to eliminate pose drift; 4) Submap fusion with online distillation for global consistency.

Result: Experiments on various indoor and outdoor datasets demonstrate superior performance and generalizability, with the framework scaling to arbitrary scenes while maintaining robustness even under pose drifts.

Conclusion: VPGS-SLAM successfully addresses the scalability limitations of 3DGS-based SLAM, enabling large-scale indoor/outdoor applications through novel voxel-based mapping, fusion tracking, and loop closure techniques.

Abstract: 3D Gaussian Splatting has recently shown promising results in dense visual SLAM. However, existing 3DGS-based SLAM methods are all constrained to small-room scenarios and struggle with memory explosion in large-scale scenes and long sequences. To this end, we propose VPGS-SLAM, the first 3DGS-based large-scale RGBD SLAM framework for both indoor and outdoor scenarios. We design a novel voxel-based progressive 3D Gaussian mapping method with multiple submaps for compact and accurate scene representation in large-scale and long-sequence scenes. This allows us to scale up to arbitrary scenes and improves robustness (even under pose drifts). In addition, we propose a 2D-3D fusion camera tracking method to achieve robust and accurate camera tracking in both indoor and outdoor large-scale scenes. Furthermore, we design a 2D-3D Gaussian loop closure method to eliminate pose drift. We further propose a submap fusion method with online distillation to achieve global consistency in large-scale scenes when detecting a loop. Experiments on various indoor and outdoor datasets demonstrate the superiority and generalizability of the proposed framework. The code will be open source on https://github.com/dtc111111/vpgs-slam.

Method: Uses pixel-wise joint re-parametrization of reflectance and surface normals as radiance vectors under varying simulated illumination. This enables integration into both traditional MVS and modern neural volume rendering pipelines.

Result: Achieves state-of-the-art performance on MVPS benchmarks (DiLiGenT-MV, LUCES-MV, Skoltech3D), excels at reconstructing fine-grained details and handling challenging visibility conditions.

Conclusion: The framework provides versatile integration of photometric information into surface reconstruction, offering improved detail preservation and robustness. Extended version includes accelerated algorithm and broader evaluation.

Abstract: Achieving high-fidelity 3D surface reconstruction while preserving fine details remains challenging, especially in the presence of materials with complex reflectance properties and without a dense-view setup. In this paper, we introduce a versatile framework that incorporates multi-view normal and optionally reflectance maps into radiance-based surface reconstruction. Our approach employs a pixel-wise joint re-parametrization of reflectance and surface normals, representing them as a vector of radiances under simulated, varying illumination. This formulation enables seamless incorporation into standard surface reconstruction pipelines, such as traditional multi-view stereo (MVS) frameworks or modern neural volume rendering (NVR) ones. Combined with the latter, our approach achieves state-of-the-art performance on multi-view photometric stereo (MVPS) benchmark datasets, including DiLiGenT-MV, LUCES-MV and Skoltech3D. In particular, our method excels in reconstructing fine-grained details and handling challenging visibility conditions. The present paper is an extended version of the earlier conference paper by Brument et al. (in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024), featuring an accelerated and more robust algorithm as well as a broader empirical evaluation. The code and data relative to this article is available at https://github.com/RobinBruneau/RNb-NeuS2.

Method: MiCo uses context-aware clustering to distill discriminative morphological patterns with cluster centroids as semantic anchors. It employs a Cluster Route module to dynamically link instances of the same tissue type across distant regions via feature similarity, and a Cluster Reducer module to consolidate redundant anchors while enhancing information exchange between distinct semantic groups.

Result: Extensive experiments on two challenging tasks across nine large-scale public cancer datasets demonstrate MiCo’s effectiveness and superiority over state-of-the-art methods.

Conclusion: MiCo successfully addresses the limitations of conventional MIL methods in handling spatial heterogeneity in WSIs by enhancing cross-regional intra-tissue correlations and strengthening inter-tissue semantic associations, showing promising results for cancer diagnosis and prognosis.

Abstract: Multiple instance learning (MIL) has shown significant promise in histopathology whole slide image (WSI) analysis for cancer diagnosis and prognosis. However, the inherent spatial heterogeneity of WSIs presents critical challenges, as morphologically similar tissue types are often dispersed across distant anatomical regions. Conventional MIL methods struggle to model these scattered tissue distributions and capture cross-regional spatial interactions effectively. To address these limitations, we propose a novel Multiple instance learning framework with Context-Aware Clustering (MiCo), designed to enhance cross-regional intra-tissue correlations and strengthen inter-tissue semantic associations in WSIs. MiCo begins by clustering instances to distill discriminative morphological patterns, with cluster centroids serving as semantic anchors. To enhance cross-regional intra-tissue correlations, MiCo employs a Cluster Route module, which dynamically links instances of the same tissue type across distant regions via feature similarity. These semantic anchors act as contextual hubs, propagating semantic relationships to refine instance-level representations. To eliminate semantic fragmentation and strengthen inter-tissue semantic associations, MiCo integrates a Cluster Reducer module, which consolidates redundant anchors while enhancing information exchange between distinct semantic groups. Extensive experiments on two challenging tasks across nine large-scale public cancer datasets demonstrate the effectiveness of MiCo, showcasing its superiority over state-of-the-art methods. The code is available at https://github.com/junjianli106/MiCo.

Method: Two key components: 1) BMPDS - an automatic data-synthesis pipeline for position-annotated multi-subject datasets providing spatial supervision, and 2) a lightweight layout-aware diffusion framework with visibility-aware attention mechanism using NeRF-inspired volumetric weight regulation to decouple spatial embeddings from identity features.

Result: Achieves state-of-the-art performance on public benchmarks, setting new records for spatial precision and identity consistency in multi-subject image customization.

Conclusion: The work represents a significant step towards truly controllable, high-fidelity image customization in multi-entity scenarios, with code and data to be publicly released.

Abstract: Recent subject-driven image customization excels in fidelity, yet fine-grained instance-level spatial control remains an elusive challenge, hindering real-world applications. This limitation stems from two factors: a scarcity of scalable, position-annotated datasets, and the entanglement of identity and layout by global attention mechanisms. To this end, we introduce \modelname{}, a unified framework for high-fidelity, spatially controllable multi-subject customization. First, we present BMPDS, the first automatic data-synthesis pipeline for position-annotated multi-subject datasets, effectively providing crucial spatial supervision. Second, we design a lightweight, layout-aware diffusion framework that integrates a novel visibility-aware attention mechanism. This mechanism explicitly models spatial relationships via an NeRF-inspired volumetric weight regulation to effectively decouple instance-level spatial embeddings from semantic identity features, enabling precise, occlusion-aware placement of multiple subjects. Extensive experiments demonstrate \modelname{} achieves state-of-the-art performance on public benchmarks, setting new records for spatial precision and identity consistency. Our work represents a significant step towards truly controllable, high-fidelity image customization in multi-entity scenarios. Code and data will be publicly released.

Method: Polymorph uses context-aware framework with lightweight Low Rank Adapters (LoRA) specialized in subsets of classes derived from co-occurrence patterns. At runtime, it dynamically selects and composes only needed adapters per frame, avoiding full-model switching and weight merging.

Result: Polymorph achieves 40% lower energy consumption and improves mAP by 9 points over strong baselines on the TAO dataset.

Conclusion: The modular adapter strategy improves scalability while reducing latency and energy overhead for real-time multi-label video classification on resource-constrained embedded devices.

Abstract: Real-time multi-label video classification on embedded devices is constrained by limited compute and energy budgets. Yet, video streams exhibit structural properties such as label sparsity, temporal continuity, and label co-occurrence that can be leveraged for more efficient inference. We introduce Polymorph, a context-aware framework that activates a minimal set of lightweight Low Rank Adapters (LoRA) per frame. Each adapter specializes in a subset of classes derived from co-occurrence patterns and is implemented as a LoRA weight over a shared backbone. At runtime, Polymorph dynamically selects and composes only the adapters needed to cover the active labels, avoiding full-model switching and weight merging. This modular strategy improves scalability while reducing latency and energy overhead. Polymorph achieves 40% lower energy consumption and improves mAP by 9 points over strong baselines on the TAO dataset. Polymorph is open source at https://github.com/inference-serving/polymorph/.

Method: Proposes taxonomy of hallucination based on faithfulness and factuality, reviews existing evaluation benchmarks for T2I and I2T tasks, and summarizes hallucination detection methods at instance level.

Result: Comprehensive review of hallucination evaluation benchmarks (construction, objectives, metrics) and detection methods, identifying current limitations and research gaps.

Conclusion: Highlights key limitations in current benchmarks and detection methods, outlines future research directions to improve hallucination evaluation and mitigation in MLLMs.

Abstract: Multi-modal Large Language Models (MLLMs) have emerged as a powerful paradigm for integrating visual and textual information, supporting a wide range of multi-modal tasks. However, these models often suffer from hallucination, producing content that appears plausible but contradicts the input content or established world knowledge. This survey offers an in-depth review of hallucination evaluation benchmarks and detection methods across Image-to-Text (I2T) and Text-to-image (T2I) generation tasks. Specifically, we first propose a taxonomy of hallucination based on faithfulness and factuality, incorporating the common types of hallucinations observed in practice. Then we provide an overview of existing hallucination evaluation benchmarks for both T2I and I2T tasks, highlighting their construction process, evaluation objectives, and employed metrics. Furthermore, we summarize recent advances in hallucination detection methods, which aims to identify hallucinated content at the instance level and serve as a practical complement of benchmark-based evaluation. Finally, we highlight key limitations in current benchmarks and detection methods, and outline potential directions for future research.

Method: GEMS uses a multimodal architecture combining swin-transformer and S3Attention to process input scenes, group members, and context information, generating joint predictions for basic discrete/continuous emotions (valence/arousal) at individual, group, and event levels.

Result: The framework is evaluated on VGAF-GEMS benchmark (extended from VGAF dataset), showing effectiveness through quantitative and qualitative comparisons with adapted state-of-the-art models.

Conclusion: GEMS provides a holistic approach to emotion comprehension in social situations, linking individual, group, and situational emotional responses, and paves the way for further research in this area.

Abstract: Understanding individual, group and event level emotions along with contextual information is crucial for analyzing a multi-person social situation. To achieve this, we frame emotion comprehension as the task of predicting fine-grained individual emotion to coarse grained group and event level emotion. We introduce GEMS that leverages a multimodal swin-transformer and S3Attention based architecture, which processes an input scene, group members, and context information to generate joint predictions. Existing multi-person emotion related benchmarks mainly focus on atomic interactions primarily based on emotion perception over time and group level. To this end, we extend and propose VGAF-GEMS to provide more fine grained and holistic analysis on top of existing group level annotation of VGAF dataset. GEMS aims to predict basic discrete and continuous emotions (including valence and arousal) as well as individual, group and event level perceived emotions. Our benchmarking effort links individual, group and situational emotional responses holistically. The quantitative and qualitative comparisons with adapted state-of-the-art models demonstrate the effectiveness of GEMS framework on VGAF-GEMS benchmarking. We believe that it will pave the way of further research. The code and data is available at: https://github.com/katariaak579/GEMS

Method: Developed a scalable data engine to generate complex LaTeX sequences, creating Tex80M (80M+ formulas). Mixed this with limited handwritten formulas to train TexTeller, the first large-scale HMER model.

Result: TexTeller achieves state-of-the-art performance across nearly all HMER benchmarks, demonstrating the effectiveness of large-scale training with synthetic data.

Conclusion: The approach successfully bridges the data scarcity gap in HMER through scalable synthetic data generation, enabling SOTA performance. The model, dataset, and code will be openly released to advance the field.

Abstract: Large foundation models have achieved significant performance gains through scalable training on massive datasets. However, the field of \textbf{H}andwritten \textbf{M}athematical \textbf{E}xpression \textbf{R}ecognition (HMER) has been impeded by the scarcity of data, primarily due to the arduous and costly process of manual annotation. To bridge this gap, we propose a novel method integrating limited handwritten formulas with large-scale LaTeX-rendered formulas by developing a scalable data engine to generate complex and consistent LaTeX sequences. With this engine, we built the largest formula dataset to date, termed \texttt{Tex80M}, comprising over 80 million high-quality training instances. Then we propose \texttt{TexTeller}, the first HMER model trained at scale, by mix-training \texttt{Tex80M} with a relatively small HME dataset. The expansive training dataset and our refined pipeline have equipped \texttt{TexTeller} with state-of-the-art (SOTA) performance across nearly all benchmarks. To advance the field, we will openly release our complete model, entire dataset, and full codebase, enabling further research building upon our contributions.

Method: SegDAC uses Segment Anything (SAM) for object-centric decomposition and YOLO-World to ground segmentation via text inputs. It features a novel transformer-based architecture supporting dynamic number of segments per time step, learning which segments to focus on using online RL without human labels.

Result: On the challenging Maniskill3 benchmark with diverse manipulation tasks under strong visual perturbations, SegDAC achieves significantly better visual generalization, doubling prior performance on hardest settings, and matches or surpasses prior methods in sample efficiency across all tasks.

Conclusion: SegDAC successfully integrates segmentation models into RL for improved visual generalization and sample efficiency, demonstrating the value of object-centric representations learned through online RL without human supervision.

Abstract: Visual reinforcement learning (RL) is challenging due to the need to extract useful representations from high-dimensional inputs while learning effective control from sparse and noisy rewards. Although large perception models exist, integrating them effectively into RL for visual generalization and improved sample efficiency remains difficult. We propose SegDAC, a Segmentation-Driven Actor-Critic method. SegDAC uses Segment Anything (SAM) for object-centric decomposition and YOLO-World to ground the image segmentation process via text inputs. It includes a novel transformer-based architecture that supports a dynamic number of segments at each time step and effectively learns which segments to focus on using online RL, without using human labels. By evaluating SegDAC over a challenging visual generalization benchmark using Maniskill3, which covers diverse manipulation tasks under strong visual perturbations, we demonstrate that SegDAC achieves significantly better visual generalization, doubling prior performance on the hardest setting and matching or surpassing prior methods in sample efficiency across all evaluated tasks. Project Page: https://segdac.github.io/

Method: Uses a simple rotating unbalanced mass to induce periodic vibrational motion for persistent event generation, combined with a motion-compensation pipeline that removes injected motion to yield clean, motion-corrected events.

Result: Hardware prototype demonstrates reliable recovery of motion parameters and improves both image reconstruction and edge detection compared to event-based sensing without motion induction.

Conclusion: The lightweight vibrational approach effectively addresses the limitation of event cameras in static scenes, providing a practical solution for sustained event generation with improved downstream perception performance.

Abstract: Event cameras are a bio-inspired class of sensors that asynchronously measure per-pixel intensity changes. Under fixed illumination conditions in static or low-motion scenes, rigidly mounted event cameras are unable to generate any events and become unsuitable for most computer vision tasks. To address this limitation, recent work has investigated motion-induced event stimulation, which often requires complex hardware or additional optical components. In contrast, we introduce a lightweight approach to sustain persistent event generation by employing a simple rotating unbalanced mass to induce periodic vibrational motion. This is combined with a motion-compensation pipeline that removes the injected motion and yields clean, motion-corrected events for downstream perception tasks. We develop a hardware prototype to demonstrate our approach and evaluate it on real-world datasets. Our method reliably recovers motion parameters and improves both image reconstruction and edge detection compared to event-based sensing without motion induction.

Method: Benchmark DINOv3 across common medical vision tasks (2D/3D classification and segmentation) on various medical imaging modalities, systematically analyzing scalability by varying model sizes and input resolutions.

Result: DINOv3 shows impressive performance and establishes a formidable new baseline, even outperforming medical-specific foundation models like BiomedCLIP and CT-Net on several tasks. However, it has clear limitations in deep domain specialization scenarios (WSIs, EM, PET) and doesn’t consistently obey scaling laws in the medical domain.

Conclusion: DINOv3 serves as a strong baseline for medical vision tasks with powerful visual features that can act as a robust prior, opening promising future directions like leveraging its features for multiview consistency in 3D reconstruction.

Abstract: The advent of large-scale vision foundation models, pre-trained on diverse natural images, has marked a paradigm shift in computer vision. However, how the frontier vision foundation models’ efficacies transfer to specialized domains remains such as medical imaging remains an open question. This report investigates whether DINOv3, a state-of-the-art self-supervised vision transformer (ViT) that features strong capability in dense prediction tasks, can directly serve as a powerful, unified encoder for medical vision tasks without domain-specific pre-training. To answer this, we benchmark DINOv3 across common medical vision tasks, including 2D/3D classification and segmentation on a wide range of medical imaging modalities. We systematically analyze its scalability by varying model sizes and input image resolutions. Our findings reveal that DINOv3 shows impressive performance and establishes a formidable new baseline. Remarkably, it can even outperform medical-specific foundation models like BiomedCLIP and CT-Net on several tasks, despite being trained solely on natural images. However, we identify clear limitations: The model’s features degrade in scenarios requiring deep domain specialization, such as in Whole-Slide Pathological Images (WSIs), Electron Microscopy (EM), and Positron Emission Tomography (PET). Furthermore, we observe that DINOv3 does not consistently obey scaling law in the medical domain; performance does not reliably increase with larger models or finer feature resolutions, showing diverse scaling behaviors across tasks. Ultimately, our work establishes DINOv3 as a strong baseline, whose powerful visual features can serve as a robust prior for multiple complex medical tasks. This opens promising future directions, such as leveraging its features to enforce multiview consistency in 3D reconstruction.

Method: 1) Generate point cloud from multi-view images using 3D reconstruction; 2) Parallel 2D encoder aggregates view-aligned features; 3) Fusion module aligns 3D geometry with 2D embeddings; 4) Graph-based decoder regresses volume, surface area, and uncertainties.

Result: Reliable performance across diverse scenarios (corals, food items, human bodies), demonstrating versatility, adaptability, robustness to sparse/noisy data, and providing a scalable, fast solution for quantitative shape analysis.

Conclusion: The framework successfully couples 3D reconstruction with neural regression and 2D features to provide accurate volumetric and surface measurements from visual data, addressing an important open challenge with practical applications.

Abstract: Accurate estimation of object volume and surface area from visual data is an open challenge with broad implications across various domains. We propose a unified framework that predicts volumetric and surface metrics directly from a set of 2D multi-view images. Our approach first generates a point cloud from the captured multi-view images using recent 3D reconstruction techniques, while a parallel 2D encoder aggregates view-aligned features. A fusion module then aligns and merges 3D geometry with 2D visual embeddings, followed by a graph-based decoder that regresses volume, surface area, and their corresponding uncertainties. This proposed architecture maintains robustness against sparse or noisy data. We evaluate the framework across multiple application domains: corals, where precise geometric measurements support growth monitoring; food items, where volume prediction relates to dietary tracking and portion analysis; and human bodies, where volumetric cues are crucial for anthropometric and medical applications. Experimental results demonstrate the reliable performance of our framework across diverse scenarios, highlighting its versatility and adaptability. Furthermore, by coupling 3D reconstruction with neural regression and 2D features, our model provides a scalable and fast solution for quantitative shape analysis from visual data.

Method: Developed a lightweight open-source large multimodal model (LMM) for reference-free evaluation. Introduced systematic approach for generating captions with controllable errors, paired with graded quality scores and explanatory annotations for robust training and interpretable evaluation.

Result: VC-Inspector achieves state-of-the-art correlation with human judgments, generalizes across diverse domains (VATEX-Eval, Flickr8K-Expert, Flickr8K-CF benchmarks), and reveals potential for caption improvement.

Conclusion: VC-Inspector provides a reproducible, fact-aware alternative to existing video caption evaluation metrics, offering strong alignment with human judgments and interpretable evaluation capabilities across multiple domains.

Abstract: We propose VC-Inspector, a lightweight, open-source large multimodal model (LMM) for reference-free evaluation of video captions, with a focus on factual accuracy. Unlike existing metrics that suffer from limited context handling, weak factuality assessment, or reliance on proprietary services, VC-Inspector offers a reproducible, fact-aware alternative that aligns closely with human judgments. To enable robust training and interpretable evaluation, we introduce a systematic approach for generating captions with controllable errors, paired with graded quality scores and explanatory annotations. Experiments show that VC-Inspector achieves state-of-the-art correlation with human judgments, generalizing across diverse domains (e.g., VATEX-Eval, Flickr8K-Expert, and Flickr8K-CF benchmarks) and revealing the potential for caption improvement.

Method: Three-stage framework: (1) generates diverse 3D meshes, (2) textures them with diffusion models, (3) applies AU-driven face rigging to synthesize multi-view faces with paired neutral/pain images, AU configurations, PSPI scores, and pain-region heatmaps. Dataset includes 82,500 samples across 25,000 pain expression heatmaps and 2,500 synthetic identities balanced by age, gender, and ethnicity. ViTPain: Vision Transformer based cross-modal distillation framework where a heatmap-trained teacher guides an RGB-trained student.

Result: Created 3DPain dataset with unprecedented annotation richness and demographic diversity. ViTPain framework enhances accuracy, interpretability, and clinical reliability through cross-modal distillation.

Conclusion: 3DPain and ViTPain together establish a controllable, diverse, and clinically grounded foundation for generalizable automated pain assessment, addressing key limitations in existing approaches.

Abstract: Automated pain assessment from facial expressions is crucial for non-communicative patients, such as those with dementia. Progress has been limited by two challenges: (i) existing datasets exhibit severe demographic and label imbalance due to ethical constraints, and (ii) current generative models cannot precisely control facial action units (AUs), facial structure, or clinically validated pain levels. We present 3DPain, a large-scale synthetic dataset specifically designed for automated pain assessment, featuring unprecedented annotation richness and demographic diversity. Our three-stage framework generates diverse 3D meshes, textures them with diffusion models, and applies AU-driven face rigging to synthesize multi-view faces with paired neutral and pain images, AU configurations, PSPI scores, and the first dataset-level annotations of pain-region heatmaps. The dataset comprises 82,500 samples across 25,000 pain expression heatmaps and 2,500 synthetic identities balanced by age, gender, and ethnicity. We further introduce ViTPain, a Vision Transformer based cross-modal distillation framework in which a heatmap-trained teacher guides a student trained on RGB images, enhancing accuracy, interpretability, and clinical reliability. Together, 3DPain and ViTPain establish a controllable, diverse, and clinically grounded foundation for generalizable automated pain assessment.

Method: Used large vision foundation models (Virchow, Virchow2, UNI) with Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning. Conducted experiments with different LoRA ranks (including rank 8) and evaluated using both random and group-based data splits to assess robustness.

Result: Best approach: Virchow with LoRA rank 8 and ensemble of three-fold cross-validation achieved 88.37% balanced accuracy on preliminary test set, ranking joint 9th in the MIDOG 2025 challenge leaderboard.

Conclusion: Foundation models with efficient adaptation strategies like LoRA show promise for atypical mitosis classification, but improvements are needed in specificity and domain generalization for clinical deployment.

Abstract: Atypical mitotic figures (AMFs) are rare abnormal cell divisions associated with tumor aggressiveness and poor prognosis. Their detection remains a significant challenge due to subtle morphological cues, class imbalance, and inter-observer variability among pathologists. The MIDOG 2025 challenge introduced a dedicated track for atypical mitosis classification, enabling systematic evaluation of deep learning methods. In this study, we investigated the use of large vision foundation models, including Virchow, Virchow2, and UNI, with Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning. We conducted extensive experiments with different LoRA ranks, as well as random and group-based data splits, to analyze robustness under varied conditions. Our best approach, Virchow with LoRA rank 8 and ensemble of three-fold cross-validation, achieved a balanced accuracy of 88.37% on the preliminary test set, ranking joint 9th in the challenge leaderboard. These results highlight the promise of foundation models with efficient adaptation strategies for the classification of atypical mitosis, while underscoring the need for improvements in specificity and domain generalization.

Method: KG-SAM incorporates: (1) medical knowledge graph for fine-grained anatomical relationships, (2) energy-based Conditional Random Field for anatomically consistent predictions, and (3) uncertainty-aware fusion module for clinical reliability.

Result: Achieves 82.69% average Dice score on prostate segmentation, 78.05% on abdominal MRI segmentation, and 79.68% on abdominal CT segmentation across multi-center medical datasets.

Conclusion: KG-SAM establishes a robust and generalizable framework that advances medical image segmentation by synergistically integrating anatomical priors with boundary refinement and uncertainty estimation.

Abstract: While the Segment Anything Model (SAM) has achieved remarkable success in image segmentation, its direct application to medical imaging remains hindered by fundamental challenges, including ambiguous boundaries, insufficient modeling of anatomical relationships, and the absence of uncertainty quantification. To address these limitations, we introduce KG-SAM, a knowledge-guided framework that synergistically integrates anatomical priors with boundary refinement and uncertainty estimation. Specifically, KG-SAM incorporates (i) a medical knowledge graph to encode fine-grained anatomical relationships, (ii) an energy-based Conditional Random Field (CRF) to enforce anatomically consistent predictions, and (iii) an uncertainty-aware fusion module to enhance reliability in high-stakes clinical scenarios. Extensive experiments across multi-center medical datasets demonstrate the effectiveness of our approach: KG-SAM achieves an average Dice score of 82.69% on prostate segmentation and delivers substantial gains in abdominal segmentation, reaching 78.05% on MRI and 79.68% on CT. These results establish KG-SAM as a robust and generalizable framework for advancing medical image segmentation.

Method: The method uses point-of-gaze tracking by analyzing what the deepfake is “seeing” (the screen displayed to the malicious actor) combined with estimated gaze from the streamed video. The model is built on explainable features selected from research on gaze patterns during dyadic conversations, focusing on subtle nonverbal communication cues that deepfakes cannot mimic.

Result: The model achieves 82% accuracy on a novel dataset created specifically for this research. This represents the first reported method to utilize point-of-gaze tracking for deepfake detection.

Conclusion: Point-of-gaze tracking provides an effective approach for real-time deepfake detection in video meetings, leveraging previously unavailable biometric information to identify subtle nonverbal communication patterns that deepfakes cannot replicate, offering a novel defense against real-time phishing attacks.

Abstract: With recent advancements in deepfake technology, it is now possible to generate convincing deepfakes in real-time. Unfortunately, malicious actors have started to use this new technology to perform real-time phishing attacks during video meetings. The nature of a video call allows access to what the deepfake is ``seeing,’’ that is, the screen displayed to the malicious actor. Using this with the estimated gaze from the malicious actors streamed video enables us to estimate where the deepfake is looking on screen, the point of gaze. Because the point of gaze during conversations is not random and is instead used as a subtle nonverbal communicator, it can be used to detect deepfakes, which are not capable of mimicking this subtle nonverbal communication. This paper proposes a real-time deepfake detection method adapted to this genre of attack, utilizing previously unavailable biometric information. We built our model based on explainable features selected after careful review of research on gaze patterns during dyadic conversations. We then test our model on a novel dataset of our creation, achieving an accuracy of 82%. This is the first reported method to utilize point-of-gaze tracking for deepfake detection.

Method: REN uses anatomical priors to train seven specialized experts for distinct lung lobes and bilateral combinations, with multi-modal gating that integrates radiomics biomarkers and DL features (CNN, ViT, Mamba) to weight expert contributions.

Result: Achieved average AUC of 0.8646 ± 0.0467 (12.5% improvement over SwinUNETR baseline), with lower-lobe experts reaching AUCs of 0.88-0.90, outperforming DL counterparts and aligning with known disease progression patterns.

Conclusion: REN demonstrates strong generalizability and clinical interpretability, presenting a scalable, anatomically-guided approach extensible to other structured medical imaging applications.

Abstract: Mixture-of-Experts (MoE) architectures have significantly contributed to scalable machine learning by enabling specialized subnetworks to tackle complex tasks efficiently. However, traditional MoE systems lack domain-specific constraints essential for medical imaging, where anatomical structure and regional disease heterogeneity strongly influence pathological patterns. Here, we introduce \textit{Regional Expert Networks (REN)}, the first anatomically-informed MoE framework tailored specifically for medical image classification. REN leverages anatomical priors to train seven specialized experts, each dedicated to distinct lung lobes and bilateral lung combinations, enabling precise modeling of region-specific pathological variations. Multi-modal gating mechanisms dynamically integrate radiomics biomarkers and deep learning (DL) features (CNN, ViT, Mamba) to weight expert contributions optimally. Applied to interstitial lung disease (ILD) classification, REN achieves consistently superior performance: the radiomics-guided ensemble reached an average AUC of 0.8646 +- 0.0467, a +12.5% improvement over the SwinUNETR baseline (AUC 0.7685, p=0.031). Region-specific experts further revealed that lower-lobe models achieved AUCs of 0.88-0.90, surpassing DL counterparts (CNN: 0.76-0.79) and aligning with known disease progression patterns. Through rigorous patient-level cross-validation, REN demonstrates strong generalizability and clinical interpretability, presenting a scalable, anatomically-guided approach readily extensible to other structured medical imaging applications. Code is available on our GitHub: https://github.com/NUBagciLab/MoE-REN.

Method: Systematically classifies image-to-video transfer learning techniques into two main categories (frozen features and adapted features) with fine-grained subcategories, and analyzes their applications across various video-text learning tasks from fine-grained to coarse-grained settings.

Result: Provides comprehensive experimental analysis of different transfer learning paradigms on downstream video understanding tasks, establishing a structured roadmap for advancing video-text learning based on existing ILFMs.

Conclusion: Identifies current challenges and promising future directions, aiming to inspire further research in this rapidly evolving field of extending image-based models to video understanding while maintaining efficiency and performance.

Abstract: Image-Language Foundation Models (ILFMs) have demonstrated remarkable success in vision-language understanding, providing transferable multimodal representations that generalize across diverse downstream image-based tasks. The advancement of video-text research has spurred growing interest in extending image-based models to the video domain. This paradigm, termed as image-to-video transfer learning, effectively mitigates the substantial data and computational demands compared to training video-language models from scratch while achieves comparable or even stronger model performance. This survey provides the first comprehensive review of this emerging field, which begins by summarizing the widely used ILFMs and their capabilities. We then systematically classify existing image-to-video transfer learning techniques into two broad root categories (frozen features and adapted features), along with numerous fine-grained subcategories, based on the paradigm for transferring image understanding capability to video tasks. Building upon the task-specific nature of image-to-video transfer, this survey methodically elaborates these strategies and details their applications across a spectrum of video-text learning tasks, ranging from fine-grained settings (e.g., spatio-temporal video grounding) to coarse-grained ones (e.g., video question answering). We further present a detailed experimental analysis to investigate the efficacy of different image-to-video transfer learning paradigms on a range of downstream video understanding tasks. Finally, we identify prevailing challenges and highlight promising directions for future research. By offering a comprehensive and structured overview, this survey aims to establish a structured roadmap for advancing video-text learning based on existing ILFM, and to inspire future research directions in this rapidly evolving domain. Github repository is available.

Method: The authors explored landmark subset selection strategies to optimize recognition performance when using lightweight MediaPipe instead of OpenPose. They also implemented spline-based imputation to handle missing landmarks effectively.

Result: Proper landmark subset selection achieved comparable or superior performance to state-of-the-art methods while reducing processing time by more than 5x compared to previous approaches. Spline-based imputation also provided substantial accuracy gains by mitigating missing landmark issues.

Conclusion: Careful landmark selection combined with simple imputation techniques enables efficient and accurate isolated sign recognition, paving the way for scalable Sign Language Recognition systems that balance both speed and accuracy.

Abstract: This paper investigates the feasibility of using lightweight body landmark detection for the recognition of isolated signs in Brazilian Sign Language (LIBRAS). Although the skeleton-based approach by Alves et al. (2024) enabled substantial improvements in recognition performance, the use of OpenPose for landmark extraction hindered time performance. In a preliminary investigation, we observed that simply replacing OpenPose with the lightweight MediaPipe, while improving processing speed, significantly reduced accuracy. To overcome this limitation, we explored landmark subset selection strategies aimed at optimizing recognition performance. Experimental results showed that a proper landmark subset achieves comparable or superior performance to state-of-the-art methods while reducing processing time by more than 5X compared to Alves et al. (2024). As an additional contribution, we demonstrated that spline-based imputation effectively mitigates missing landmark issues, leading to substantial accuracy gains. These findings highlight that careful landmark selection, combined with simple imputation techniques, enables efficient and accurate isolated sign recognition, paving the way for scalable Sign Language Recognition systems.

Method: LangHOPS integrates Multimodal Large Language Models into the object-part parsing pipeline to ground object-part hierarchies in language space. It uses MLLM-driven part query refinement strategy to leverage the model’s knowledge and reasoning capabilities for linking hierarchical concepts.

Result: State-of-the-art performance: 5.5% AP improvement (in-domain) and 4.8% AP improvement (cross-dataset) on PartImageNet; 2.5% mIOU improvement on unseen object parts in ADE20K for zero-shot semantic segmentation.

Conclusion: Grounding object-part hierarchies in language space using MLLMs is effective for open-vocabulary object-part instance segmentation, outperforming previous visual grouping approaches and demonstrating strong generalization across domains and zero-shot scenarios.

Abstract: We propose LangHOPS, the first Multimodal Large Language Model (MLLM) based framework for open-vocabulary object-part instance segmentation. Given an image, LangHOPS can jointly detect and segment hierarchical object and part instances from open-vocabulary candidate categories. Unlike prior approaches that rely on heuristic or learnable visual grouping, our approach grounds object-part hierarchies in language space. It integrates the MLLM into the object-part parsing pipeline to leverage its rich knowledge and reasoning capabilities, and link multi-granularity concepts within the hierarchies. We evaluate LangHOPS across multiple challenging scenarios, including in-domain and cross-dataset object-part instance segmentation, and zero-shot semantic segmentation. LangHOPS achieves state-of-the-art results, surpassing previous methods by 5.5% Average Precision (AP) (in-domain) and 4.8% (cross-dataset) on the PartImageNet dataset and by 2.5% mIOU on unseen object parts in ADE20K (zero-shot). Ablation studies further validate the effectiveness of the language-grounded hierarchy and MLLM driven part query refinement strategy. The code will be released here.

Method: Two-part framework: (1) DermBench - a curated benchmark with 4,000 real-world dermatology images paired with expert-certified diagnostic narratives, using LLM-based judges for scoring; (2) DermEval - a reference-free multimodal evaluator trained to produce structured critiques with overall scores and per-dimension ratings.

Result: Experiments on 4,500 cases show DermBench and DermEval achieve close alignment with expert ratings, with mean deviations of 0.251 and 0.117 (out of 5), respectively, providing reliable measurement across different multimodal LLMs.

Conclusion: The proposed framework enables clinically meaningful, reproducible, and scalable assessment of multimodal LLMs in dermatology, addressing the critical need for reliable evaluation before clinical deployment.

Abstract: Multimodal large language models (LLMs) are increasingly used to generate dermatology diagnostic narratives directly from images. However, reliable evaluation remains the primary bottleneck for responsible clinical deployment. We introduce a novel evaluation framework that combines DermBench, a meticulously curated benchmark, with DermEval, a robust automatic evaluator, to enable clinically meaningful, reproducible, and scalable assessment. We build DermBench, which pairs 4,000 real-world dermatology images with expert-certified diagnostic narratives and uses an LLM-based judge to score candidate narratives across clinically grounded dimensions, enabling consistent and comprehensive evaluation of multimodal models. For individual case assessment, we train DermEval, a reference-free multimodal evaluator. Given an image and a generated narrative, DermEval produces a structured critique along with an overall score and per-dimension ratings. This capability enables fine-grained, per-case analysis, which is critical for identifying model limitations and biases. Experiments on a diverse dataset of 4,500 cases demonstrate that DermBench and DermEval achieve close alignment with expert ratings, with mean deviations of 0.251 and 0.117 (out of 5), respectively, providing reliable measurement of diagnostic ability and trustworthiness across different multimodal LLMs.

Method: Used Delphi process to achieve consensus on workflow descriptors, developed ColoWorkflow tool based on consensus framework, then applied it to 54 operative videos from 5 centers. Evaluated applicability and inter-rater reliability through independent raters.

Result: Achieved consensus for 10 procedure-agnostic phases and 34 procedure-specific steps. Tool demonstrated broad applicability (all but one label used) with moderate inter-rater reliability (mean Cohen’s K of 0.71 for phases, 0.66 for steps). Most discrepancies occurred at phase transitions and step boundaries.

Conclusion: ColoWorkflow is the first consensus-based, validated VBA tool for comprehensive workflow analysis in minimally invasive colorectal surgery. It provides a reproducible framework for performance assessment, enabling benchmarking and supporting AI-driven workflow recognition to standardize training and improve surgical quality.

Abstract: Minimally invasive colorectal surgery is characterized by procedural variability, a difficult learning curve, and complications that impact quality and outcomes. Video-based assessment (VBA) offers an opportunity to generate data-driven insights to reduce variability, optimize training, and improve surgical performance. However, existing tools for workflow analysis remain difficult to standardize and implement. This study aims to develop and validate a VBA tool for workflow analysis across minimally invasive colorectal procedures. A Delphi process was conducted to achieve consensus on generalizable workflow descriptors. The resulting framework informed the development of a new VBA tool, ColoWorkflow. Independent raters then applied ColoWorkflow to a multicentre video dataset of laparoscopic and robotic colorectal surgery (CRS). Applicability and inter-rater reliability were evaluated. Consensus was achieved for 10 procedure-agnostic phases and 34 procedure-specific steps describing CRS workflows. ColoWorkflow was developed and applied to 54 colorectal operative videos (left and right hemicolectomies, sigmoid and rectosigmoid resections, and total proctocolectomies) from five centres. The tool demonstrated broad applicability, with all but one label utilized. Inter-rater reliability was moderate, with mean Cohen’s K of 0.71 for phases and 0.66 for steps. Most discrepancies arose at phase transitions and step boundary definitions. ColoWorkflow is the first consensus-based, validated VBA tool for comprehensive workflow analysis in minimally invasive CRS. It establishes a reproducible framework for video-based performance assessment, enabling benchmarking across institutions and supporting the development of artificial intelligence-driven workflow recognition. Its adoption may standardize training, accelerate competency acquisition, and advance data-informed surgical quality improvement.

Method: Proposes FashionMAC with two key innovations: 1) Deformation-free approach that directly out-paints garments segmented from dressed persons to preserve details, and 2) Region-adaptive decoupled attention (RADA) mechanism with chained mask injection for fine-grained attribute control.

Result: Extensive experiments show superior performance compared to state-of-the-art methods, achieving high-quality fashion showcase images with faithful garment preservation and enhanced controllability.

Conclusion: FashionMAC effectively addresses key challenges in garment-centric fashion generation by eliminating deformation-induced distortions and providing fine-grained appearance control through novel attention mechanisms.

Abstract: Garment-centric fashion image generation aims to synthesize realistic and controllable human models dressing a given garment, which has attracted growing interest due to its practical applications in e-commerce. The key challenges of the task lie in two aspects: (1) faithfully preserving the garment details, and (2) gaining fine-grained controllability over the model’s appearance. Existing methods typically require performing garment deformation in the generation process, which often leads to garment texture distortions. Also, they fail to control the fine-grained attributes of the generated models, due to the lack of specifically designed mechanisms. To address these issues, we propose FashionMAC, a novel diffusion-based deformation-free framework that achieves high-quality and controllable fashion showcase image generation. The core idea of our framework is to eliminate the need for performing garment deformation and directly outpaint the garment segmented from a dressed person, which enables faithful preservation of the intricate garment details. Moreover, we propose a novel region-adaptive decoupled attention (RADA) mechanism along with a chained mask injection strategy to achieve fine-grained appearance controllability over the synthesized human models. Specifically, RADA adaptively predicts the generated regions for each fine-grained text attribute and enforces the text attribute to focus on the predicted regions by a chained mask injection strategy, significantly enhancing the visual fidelity and the controllability. Extensive experiments validate the superior performance of our framework compared to existing state-of-the-art methods.

Method: 1) Use Brownian Bridge Diffusion Model (BBDM) to synthesize MRI slices from 3D CT scans with limited slices. 2) Retrieve structurally/pathologically similar CT slices from prior database via fine-tuned retrieval model. 3) Use retrieved slices as references through ControlNet branch to guide intermediate MRI slice generation. 4) Apply spherical linear interpolation for rare retrieval failures when database lacks suitable references.

Result: Extensive experiments on SynthRAD2023 and BraTS datasets demonstrate that ReBrain achieves state-of-the-art performance in cross-modal reconstruction under sparse CT conditions.

Conclusion: ReBrain effectively addresses the challenge of synthesizing brain MRI from sparse CT scans by combining diffusion modeling with retrieval-augmented guidance, providing a robust solution for clinical scenarios where MRI is not feasible.

Abstract: Magnetic Resonance Imaging (MRI) plays a crucial role in brain disease diagnosis, but it is not always feasible for certain patients due to physical or clinical constraints. Recent studies attempt to synthesize MRI from Computed Tomography (CT) scans; however, low-dose protocols often result in highly sparse CT volumes with poor through-plane resolution, making accurate reconstruction of the full brain MRI volume particularly challenging. To address this, we propose ReBrain, a retrieval-augmented diffusion framework for brain MRI reconstruction. Given any 3D CT scan with limited slices, we first employ a Brownian Bridge Diffusion Model (BBDM) to synthesize MRI slices along the 2D dimension. Simultaneously, we retrieve structurally and pathologically similar CT slices from a comprehensive prior database via a fine-tuned retrieval model. These retrieved slices are used as references, incorporated through a ControlNet branch to guide the generation of intermediate MRI slices and ensure structural continuity. We further account for rare retrieval failures when the database lacks suitable references and apply spherical linear interpolation to provide supplementary guidance. Extensive experiments on SynthRAD2023 and BraTS demonstrate that ReBrain achieves state-of-the-art performance in cross-modal reconstruction under sparse conditions.

Method: REXO lifts the 2D bounding box diffusion process of DiffusionDet into 3D radar space, using noisy 3D bounding boxes to guide explicit cross-view radar feature association. It incorporates prior knowledge that people are in contact with the ground to reduce diffusion parameters.

Result: REXO surpasses state-of-the-art methods by +4.22 AP on the HIBER dataset and +11.02 AP on the MMVR dataset.

Conclusion: The proposed explicit cross-view feature association through 3D bounding box diffusion effectively addresses limitations of implicit association methods, achieving superior performance on indoor radar object detection tasks.

Abstract: Multi-view indoor radar perception has drawn attention due to its cost-effectiveness and low privacy risks. Existing methods often rely on {implicit} cross-view radar feature association, such as proposal pairing in RFMask or query-to-feature cross-attention in RETR, which can lead to ambiguous feature matches and degraded detection in complex indoor scenes. To address these limitations, we propose \textbf{REXO} (multi-view Radar object dEtection with 3D bounding boX diffusiOn), which lifts the 2D bounding box (BBox) diffusion process of DiffusionDet into the 3D radar space. REXO utilizes these noisy 3D BBoxes to guide an {explicit} cross-view radar feature association, enhancing the cross-view radar-conditioned denoising process. By accounting for prior knowledge that the person is in contact with the ground, REXO reduces the number of diffusion parameters by determining them from this prior. Evaluated on two open indoor radar datasets, our approach surpasses state-of-the-art methods by a margin of +4.22 AP on the HIBER dataset and +11.02 AP on the MMVR dataset. The REXO implementation is available at https://github.com/merlresearch/radar-bbox-diffusion.

Method: Created C3 dataset by reconstructing scenes in 3D from Internet photo collections via structure-from-motion, then manually registering reconstructions to floor plans from Internet to derive correspondences between images and floor plans.

Result: C3 contains 90K paired floor plans and photos across 597 scenes with 153M pixel-level correspondences and 85K camera poses. Training on C3 improves best performing method by 34% in RMSE. Also used predicted correspondences to estimate camera poses.

Conclusion: The C3 dataset addresses limitations of existing datasets and helps tackle cross-modal geometric reasoning challenges between photos and floor plans, enabling better correspondence prediction and camera pose estimation.

Abstract: Geometric models like DUSt3R have shown great advances in understanding the geometry of a scene from pairs of photos. However, they fail when the inputs are from vastly different viewpoints (e.g., aerial vs. ground) or modalities (e.g., photos vs. abstract drawings) compared to what was observed during training. This paper addresses a challenging version of this problem: predicting correspondences between ground-level photos and floor plans. Current datasets for joint photo-floor plan reasoning are limited, either lacking in varying modalities (VIGOR) or lacking in correspondences (WAFFLE). To address these limitations, we introduce a new dataset, C3, created by first reconstructing a number of scenes in 3D from Internet photo collections via structure-from-motion, then manually registering the reconstructions to floor plans gathered from the Internet, from which we can derive correspondences between images and floor plans. C3 contains 90K paired floor plans and photos across 597 scenes with 153M pixel-level correspondences and 85K camera poses. We find that state-of-the-art correspondence models struggle on this task. By training on our new data, we can improve on the best performing method by 34% in RMSE. We also use the predicted correspondences to estimate camera poses and evaluate performance using recall metrics. Lastly, we identify open challenges in cross-modal geometric reasoning that our dataset aims to help address.

Method: Vehicles exchange compact reference points (object metadata) instead of full features, shifting focus from “what is seen” to “where to see.” A selective Top-K query fusion adds high-confidence queries from senders to enrich information while maintaining low bandwidth.

Result: On M3CAD dataset, RefPtsFusion reduces communication overhead from hundreds of MB/s to only a few KB/s at 5 FPS (5 orders of magnitude reduction) while maintaining stable perception performance. Shows strong robustness and consistent transmission behavior.

Conclusion: RefPtsFusion provides a sensor- and model-independent interface for cooperative driving that works across heterogeneous vehicles, enabling scalable, real-time systems with dramatically reduced communication requirements.

Abstract: We present RefPtsFusion, a lightweight and interpretable framework for cooperative autonomous driving. Instead of sharing large feature maps or query embeddings, vehicles exchange compact reference points, e.g., objects’ positions, velocities, and size information. This approach shifts the focus from “what is seen” to “where to see”, creating a sensor- and model-independent interface that works well across vehicles with heterogeneous perception models while greatly reducing communication bandwidth. To enhance the richness of shared information, we further develop a selective Top-K query fusion that selectively adds high-confidence queries from the sender. It thus achieves a strong balance between accuracy and communication cost. Experiments on the M3CAD dataset show that RefPtsFusion maintains stable perception performance while reducing communication overhead by five orders of magnitude, dropping from hundreds of MB/s to only a few KB/s at 5 FPS (frame per second), compared to traditional feature-level fusion methods. Extensive experiments also demonstrate RefPtsFusion’s strong robustness and consistent transmission behavior, highlighting its potential for scalable, real-time cooperative driving systems.

Method: Proposes smooth regularization that encourages smoothness in intermediate-layer embeddings of consecutive frames by modeling their changes as a Gaussian Random Walk (GRW). This penalizes abrupt representational shifts and promotes low-acceleration solutions.

Result: Achieves 3.8% to 6.4% accuracy improvement on Kinetics-600. MoViNets improve state-of-the-art by 3.8-6.1% within FLOP constraints, while MobileNetV3 and MoViNets-Stream achieve 4.9-6.4% gains over prior models with comparable memory footprints.

Conclusion: The GRW-based smooth regularization effectively enhances temporal modeling in lightweight video architectures, enabling them to better capture complex temporal dynamics while maintaining computational efficiency.

Abstract: We propose a smooth regularization technique that instills a strong temporal inductive bias in video recognition models, particularly benefiting lightweight architectures. Our method encourages smoothness in the intermediate-layer embeddings of consecutive frames by modeling their changes as a Gaussian Random Walk (GRW). This penalizes abrupt representational shifts, thereby promoting low-acceleration solutions that better align with the natural temporal coherence inherent in videos. By leveraging this enforced smoothness, lightweight models can more effectively capture complex temporal dynamics. Applied to such models, our technique yields a 3.8% to 6.4% accuracy improvement on Kinetics-600. Notably, the MoViNets model family trained with our smooth regularization improves the current state of the art by 3.8% to 6.1% within their respective FLOP constraints, while MobileNetV3 and the MoViNets-Stream family achieve gains of 4.9% to 6.4% over prior state-of-the-art models with comparable memory footprints. Our code and models are available at https://github.com/cmusatyalab/grw-smoothing.

Method: ECOCSeg uses error-correcting output codes (ECOC) to create fine-grained encodings for each class. It introduces an ECOC-based classifier that disentangles classes into attributes and handles partial inaccurate bits. A bit-level label denoising mechanism is developed to generate higher-quality pseudo-labels for unlabeled images.

Result: The method consistently demonstrates significant improvements on multiple UDA and SSL benchmarks across different segmentation architectures. It can be easily integrated with existing methods and provides more stable and generalized performance.

Conclusion: ECOCSeg offers a novel perspective for segmentation models that addresses error amplification in pseudo-label learning through ECOC-based encoding, improving stability and generalization while providing robust supervision for unlabeled data in label-scarce scenarios.

Abstract: Pseudo-label learning is widely used in semantic segmentation, particularly in label-scarce scenarios such as unsupervised domain adaptation (UDA) and semisupervised learning (SSL). Despite its success, this paradigm can generate erroneous pseudo-labels, which are further amplified during training due to utilization of one-hot encoding. To address this issue, we propose ECOCSeg, a novel perspective for segmentation models that utilizes error-correcting output codes (ECOC) to create a fine-grained encoding for each class. ECOCSeg offers several advantages. First, an ECOC-based classifier is introduced, enabling model to disentangle classes into attributes and handle partial inaccurate bits, improving stability and generalization in pseudo-label learning. Second, a bit-level label denoising mechanism is developed to generate higher-quality pseudo-labels, providing adequate and robust supervision for unlabeled images. ECOCSeg can be easily integrated with existing methods and consistently demonstrates significant improvements on multiple UDA and SSL benchmarks across different segmentation architectures. Code is available at https://github.com/Woof6/ECOCSeg.

Method: Introduces InfiniteDiffusion algorithm for infinite generation on unbounded domains. Uses hierarchical stack of diffusion models to couple planetary context with local detail. Implements compact Laplacian encoding to stabilize outputs across Earth-scale dynamic ranges. Develops open-source infinite-tensor framework for constant-memory manipulation of unbounded tensors.

Result: The framework outpaces orbital velocity by 9 times on consumer GPU, enabling realistic terrain generation at interactive rates. Achieves seamless infinite extent, seed-consistency, and constant-time random access while maintaining diffusion model fidelity.

Conclusion: Terrain Diffusion positions diffusion models as a practical, scalable foundation for next-generation infinite virtual worlds, bridging the gap between procedural noise limitations and modern generative model capabilities.

Abstract: For decades, procedural worlds have been built on procedural noise functions such as Perlin noise, which are fast and infinite, yet fundamentally limited in realism and large-scale coherence. We introduce Terrain Diffusion, a generative framework that bridges the fidelity of diffusion models with the properties that made procedural noise indispensable: seamless infinite extent, seed-consistency, and constant-time random access. At its core is InfiniteDiffusion, a novel algorithm for infinite generation that reformulates standard diffusion sampling for unbounded domains. While noise functions remain near-instant, our framework outpaces orbital velocity by 9 times on a consumer GPU, enabling realistic terrain generation at interactive rates. We integrate a hierarchical stack of diffusion models to couple planetary context with local detail, a compact Laplacian encoding to stabilize outputs across Earth-scale dynamic ranges, and an open-source infinite-tensor framework for constant-memory manipulation of unbounded tensors. Together, these components position diffusion models as a practical, scalable foundation for the next generation of infinite virtual worlds.

Method: Treats articulated objects as assemblies of rigid parts, uses transformer architecture to map sparse images to learnable part slots, jointly decodes unified representations including 3D geometry, texture, and explicit articulation parameters.

Result: Achieves significant improvements over existing baselines, establishes new state-of-the-art for articulated object reconstruction from image inputs across diverse benchmarks.

Conclusion: ART provides a category-agnostic, feed-forward solution that produces physically interpretable, simulation-ready reconstructions of articulated objects from sparse image inputs.

Abstract: We introduce ART, Articulated Reconstruction Transformer – a category-agnostic, feed-forward model that reconstructs complete 3D articulated objects from only sparse, multi-state RGB images. Previous methods for articulated object reconstruction either rely on slow optimization with fragile cross-state correspondences or use feed-forward models limited to specific object categories. In contrast, ART treats articulated objects as assemblies of rigid parts, formulating reconstruction as part-based prediction. Our newly designed transformer architecture maps sparse image inputs to a set of learnable part slots, from which ART jointly decodes unified representations for individual parts, including their 3D geometry, texture, and explicit articulation parameters. The resulting reconstructions are physically interpretable and readily exportable for simulation. Trained on a large-scale, diverse dataset with per-part supervision, and evaluated across diverse benchmarks, ART achieves significant improvements over existing baselines and establishes a new state of the art for articulated object reconstruction from image inputs.

Method: Two-stage approach: 1) Generate compatible random face via diffusion model inpainting, 2) Perform temporally consistent face swapping with authentic expression transfer across video frames.

Result: BLANKET alters identity effectively and outperforms DeepPrivacy2 in de-identification level, facial attribute preservation, impact on downstream tasks (human pose estimation), and artifact reduction.

Conclusion: BLANKET provides superior infant face anonymization for video data, balancing privacy protection with preservation of essential facial attributes, with code available as an easy-to-use demo.

Abstract: Ensuring the ethical use of video data involving human subjects, particularly infants, requires robust anonymization methods. We propose BLANKET (Baby-face Landmark-preserving ANonymization with Keypoint dEtection consisTency), a novel approach designed to anonymize infant faces in video recordings while preserving essential facial attributes. Our method comprises two stages. First, a new random face, compatible with the original identity, is generated via inpainting using a diffusion model. Second, the new identity is seamlessly incorporated into each video frame through temporally consistent face swapping with authentic expression transfer. The method is evaluated on a dataset of short video recordings of babies and is compared to the popular anonymization method, DeepPrivacy2. Key metrics assessed include the level of de-identification, preservation of facial attributes, impact on human pose estimation (as an example of a downstream task), and presence of artifacts. Both methods alter the identity, and our method outperforms DeepPrivacy2 in all other respects. The code is available as an easy-to-use anonymization demo at https://github.com/ctu-vras/blanket-infant-face-anonym.

Method: The Universal UltraSound Image Challenge 2025 (UUSIC25) evaluated algorithms on 11,644 images from 12 sources, with independent multi-center testing including completely unseen data to assess generalization.

Result: Top model (SMART) achieved DSC of 0.854 across 5 segmentation tasks and AUC of 0.766 for classification, but showed significant performance drop in breast cancer subtyping from AUC 0.571 (internal) to 0.508 (unseen external).

Conclusion: General-purpose AI models can be accurate and efficient across multiple ultrasound tasks, but domain generalization remains a critical challenge for clinical deployment.

Abstract: IMPORTANCE: Modern ultrasound systems are universal diagnostic tools capable of imaging the entire body. However, current AI solutions remain fragmented into single-task tools. This critical gap between hardware versatility and software specificity limits workflow integration and clinical utility. OBJECTIVE: To evaluate the diagnostic accuracy, versatility, and efficiency of single general-purpose deep learning models for multi-organ classification and segmentation. DESIGN: The Universal UltraSound Image Challenge 2025 (UUSIC25) involved developing algorithms on 11,644 images aggregated from 12 sources (9 public, 3 private). Evaluation used an independent, multi-center private test set of 2,479 images, including data from a center completely unseen during training to assess generalization. OUTCOMES: Diagnostic performance (Dice Similarity Coefficient [DSC]; Area Under the Receiver Operating Characteristic Curve [AUC]) and computational efficiency (inference time, GPU memory). RESULTS: Of 15 valid algorithms, the top model (SMART) achieved a macro-averaged DSC of 0.854 across 5 segmentation tasks and AUC of 0.766 for binary classification. Models demonstrated high capability in anatomical segmentation (e.g., fetal head DSC: 0.942) but variability in complex diagnostic tasks subject to domain shift. Specifically, in breast cancer molecular subtyping, the top model’s performance dropped from an AUC of 0.571 (internal) to 0.508 (unseen external center), highlighting the challenge of generalization. CONCLUSIONS: General-purpose AI models can achieve high accuracy and efficiency across multiple tasks using a single architecture. However, significant performance degradation on unseen data suggests domain generalization is critical for future clinical deployment.

Method: SuperFlow adjusts group sizes with variance-aware sampling to allocate more samples to important prompts, and computes step-level advantages in a way that is consistent with continuous-time flow dynamics rather than reusing trajectory-level advantages.

Result: SuperFlow reaches promising performance using only 5.4% to 56.3% of original training steps, reduces training time by 5.2% to 16.7% without architectural changes, and improves over SD3.5-M by 4.6% to 47.2% and over Flow-GRPO by 1.7% to 16.0% on T2I tasks.

Conclusion: SuperFlow provides an efficient RL training framework for flow-based models that addresses sampling inefficiency and credit assignment bias, enabling faster training and better performance on text-to-image generation tasks.

Abstract: Recent progress in flow-based generative models and reinforcement learning (RL) has improved text-image alignment and visual quality. However, current RL training for flow models still has two main problems: (i) GRPO-style fixed per-prompt group sizes ignore variation in sampling importance across prompts, which leads to inefficient sampling and slower training; and (ii) trajectory-level advantages are reused as per-step estimates, which biases credit assignment along the flow. We propose SuperFlow, an RL training framework for flow-based models that adjusts group sizes with variance-aware sampling and computes step-level advantages in a way that is consistent with continuous-time flow dynamics. Empirically, SuperFlow reaches promising performance while using only 5.4% to 56.3% of the original training steps and reduces training time by 5.2% to 16.7% without any architectural changes. On standard text-to-image (T2I) tasks, including text rendering, compositional image generation, and human preference alignment, SuperFlow improves over SD3.5-M by 4.6% to 47.2%, and over Flow-GRPO by 1.7% to 16.0%.

Method: Proposes RealCamo, an out-painting-based framework with explicit layout controls to regulate global image structure and improve semantic coherence. Uses multimodal textual-visual conditions combining fine-grained textual task descriptions with texture-oriented background retrieval to guide generation for enhanced visual fidelity.

Result: Extensive experiments and visualizations demonstrate the effectiveness of the proposed framework. Also introduces a new quantitative metric (background-foreground distribution divergence) to measure camouflage quality in generated images.

Conclusion: RealCamo addresses key limitations in existing CIG methods by providing better control over image structure and semantic coherence, resulting in more realistic camouflaged images that can serve as high-quality training data for camouflaged object detection tasks.

Abstract: Camouflaged image generation (CIG) has recently emerged as an efficient alternative for acquiring high-quality training data for camouflaged object detection (COD). However, existing CIG methods still suffer from a substantial gap to real camouflaged imagery: generated images either lack sufficient camouflage due to weak visual similarity, or exhibit cluttered backgrounds that are semantically inconsistent with foreground targets. To address these limitations, we propose RealCamo, a novel out-painting-based framework for controllable realistic camouflaged image generation. RealCamo explicitly introduces additional layout controls to regulate global image structure, thereby improving semantic coherence between foreground objects and generated backgrounds. Moreover, we construct a multimodal textual-visual condition by combining a unified fine-grained textual task description with texture-oriented background retrieval, which jointly guides the generation process to enhance visual fidelity and realism. To quantitatively assess camouflage quality, we further introduce a background-foreground distribution divergence metric that measures the effectiveness of camouflage in generated images. Extensive experiments and visualizations demonstrate the effectiveness of our proposed framework.

Method: Proposes Co2S with: 1) Heterogeneous dual-student architecture using ViT-based models initialized with CLIP and DINOv3; 2) Explicit-implicit semantic co-guidance mechanism using text embeddings (explicit) and learnable queries (implicit); 3) Global-local feature collaborative fusion strategy to combine CLIP’s global context with DINOv3’s local details.

Result: Extensive experiments on six popular datasets show superior performance, consistently achieving leading results across various partition protocols and diverse scenarios.

Conclusion: Co2S effectively mitigates pseudo-label drift in semi-supervised remote sensing segmentation by synergistically fusing vision-language and self-supervised model priors through innovative architectural and guidance mechanisms.

Abstract: Semi-supervised remote sensing (RS) image semantic segmentation offers a promising solution to alleviate the burden of exhaustive annotation, yet it fundamentally struggles with pseudo-label drift, a phenomenon where confirmation bias leads to the accumulation of errors during training. In this work, we propose Co2S, a stable semi-supervised RS segmentation framework that synergistically fuses priors from vision-language models and self-supervised models. Specifically, we construct a heterogeneous dual-student architecture comprising two distinct ViT-based vision foundation models initialized with pretrained CLIP and DINOv3 to mitigate error accumulation and pseudo-label drift. To effectively incorporate these distinct priors, an explicit-implicit semantic co-guidance mechanism is introduced that utilizes text embeddings and learnable queries to provide explicit and implicit class-level guidance, respectively, thereby jointly enhancing semantic consistency. Furthermore, a global-local feature collaborative fusion strategy is developed to effectively fuse the global contextual information captured by CLIP with the local details produced by DINOv3, enabling the model to generate highly precise segmentation results. Extensive experiments on six popular datasets demonstrate the superiority of the proposed method, which consistently achieves leading performance across various partition protocols and diverse scenarios. Project page is available at https://xavierjiezou.github.io/Co2S/.

Method: Proposes a unified DWM framework based on 3D Gaussian scene representation with: 1) linguistic features embedded into Gaussian primitives for early modality alignment, 2) task-aware language-guided sampling to remove redundant Gaussians and inject compact 3D tokens into LLM, 3) dual-condition multi-modal generation model combining high-level language conditions from vision-language model with low-level image conditions.

Result: Achieves state-of-the-art performance on nuScenes and NuInteract datasets, validating the effectiveness of the framework for both 3D scene understanding and multi-modal scene generation tasks.

Conclusion: The proposed 3D Gaussian-based DWM framework successfully addresses limitations of existing approaches by enabling unified 3D scene understanding and generation with accurate language-scene alignment, demonstrating superior performance on benchmark datasets.

Abstract: Driving World Models (DWMs) have been developing rapidly with the advances of generative models. However, existing DWMs lack 3D scene understanding capabilities and can only generate content conditioned on input data, without the ability to interpret or reason about the driving environment. Moreover, current approaches represent 3D spatial information with point cloud or BEV features do not accurately align textual information with the underlying 3D scene. To address these limitations, we propose a novel unified DWM framework based on 3D Gaussian scene representation, which enables both 3D scene understanding and multi-modal scene generation, while also enabling contextual enrichment for understanding and generation tasks. Our approach directly aligns textual information with the 3D scene by embedding rich linguistic features into each Gaussian primitive, thereby achieving early modality alignment. In addition, we design a novel task-aware language-guided sampling strategy that removes redundant 3D Gaussians and injects accurate and compact 3D tokens into LLM. Furthermore, we design a dual-condition multi-modal generation model, where the information captured by our vision-language model is leveraged as a high-level language condition in combination with a low-level image condition, jointly guiding the multi-modal generation process. We conduct comprehensive studies on the nuScenes, and NuInteract datasets to validate the effectiveness of our framework. Our method achieves state-of-the-art performance. We will release the code publicly on GitHub https://github.com/dtc111111/GaussianDWM.

Method: Three key components: 1) Gas Block with diffusion-convection modeling and edge-gated fusion; 2) AGPEO edge operator with multi-scale perception; 3) CASR-PAN with adaptive routing for selective feature propagation across scales based on edge and content cues.

Result: Achieves 29.8% overall AP, 84.3% AP50, and 25.3% small-object AP on IIG dataset, surpassing RT-DETR-R18 baseline by 3.0%, 6.5%, and 5.3% respectively, with only 43.7 Gflops and 14.9M parameters.

Conclusion: PEG-DRNet achieves superior performance with the best balance of accuracy and computational efficiency, outperforming existing CNN and Transformer detectors on IIG and LangGas datasets for infrared gas leak detection.

Abstract: Detecting infrared gas leaks is critical for environmental monitoring and industrial safety, yet remains difficult because plumes are faint, small, semitransparent, and have weak, diffuse boundaries. We present physics-edge hybrid gas dynamic routing network (PEG-DRNet). First, we introduce the Gas Block, a diffusion-convection unit modeling gas transport: a local branch captures short-range variations, while a large-kernel branch captures long-range propagation. An edge-gated learnable fusion module balances local detail and global context, strengthening weak-contrast plume and contour cues. Second, we propose the adaptive gradient and phase edge operator (AGPEO), computing reliable edge priors from multi-directional gradients and phase-consistent responses. These are transformed by a multi-scale edge perception module (MSEPM) into hierarchical edge features that reinforce boundaries. Finally, the content-adaptive sparse routing path aggregation network (CASR-PAN), with adaptive information modulation modules for fusion and self, selectively propagates informative features across scales based on edge and content cues, improving cross-scale discriminability while reducing redundancy. Experiments on the IIG dataset show that PEG-DRNet achieves an overall AP of 29.8%, an AP${50}$ of 84.3%, and a small-object AP of 25.3%, surpassing the RT-DETR-R18 baseline by 3.0%, 6.5%, and 5.3%, respectively, while requiring only 43.7 Gflops and 14.9 M parameters. The proposed PEG-DRNet achieves superior overall performance with the best balance of accuracy and computational efficiency, outperforming existing CNN and Transformer detectors in AP and AP${50}$ on the IIG and LangGas dataset.

Method: Collected data using mobile phone cameras on roads around Karlsruhe Institute of Technology. Tested multiple deep learning algorithms (AlexNet, LeNet, VGG, ResNet) for road classification. Used both road image data and acceleration data (converted to images) for training. Proposed fuzzy logic for weather and time-of-day classification.

Result: Achieved over 95% accuracy for 5 road condition classes: asphalt, damaged asphalt, gravel road, damaged gravel road, and pavement road. Compared performances of acceleration-based and camera image-based approaches.

Conclusion: The proposed real-time system using mobile phone data and deep learning effectively monitors road surfaces, outperforming classical methods. The combination of image and acceleration data with fuzzy logic enables comprehensive road condition classification considering weather and time factors.

Abstract: Monitoring states of road surfaces provides valuable information for the planning and controlling vehicles and active vehicle control systems. Classical road monitoring methods are expensive and unsystematic because they require time for measurements. This article proposes an real time system based on weather conditional data and road surface condition data. For this purpose, we collected data with a mobile phone camera on the roads around the campus of the Karlsruhe Institute of Technology. We tested a large number of different image-based deep learning algorithms for road classification. In addition, we used road acceleration data along with road image data for training by using them as images. We compared the performances of acceleration-based and camera image-based approaches. The performances of the simple Alexnet, LeNet, VGG, and Resnet algorithms were compared as deep learning algorithms. For road condition classification, 5 classes were considered: asphalt, damaged asphalt, gravel road, damaged gravel road, pavement road and over 95% accuracy performance was achieved. It is also proposed to use the acceleration or the camera image to classify the road surface according to the weather and the time of day using fuzzy logic.

Method: Created IMDD-1M dataset with 1M aligned image-text pairs of real-world defects, then trained a diffusion-based vision-language foundation model from scratch on this dataset, enabling efficient adaptation to specialized domains through lightweight fine-tuning.

Result: The foundation model achieves comparable performance to dedicated expert models using less than 5% of task-specific data, demonstrating data-efficient adaptation for industrial inspection and generation tasks.

Conclusion: IMDD-1M and the trained foundation model enable scalable, domain-adaptive, and knowledge-grounded manufacturing intelligence, paving the way for advanced multimodal applications in industrial quality inspection.

Abstract: We present IMDD-1M, the first large-scale Industrial Multimodal Defect Dataset comprising 1,000,000 aligned image-text pairs, designed to advance multimodal learning for manufacturing and quality inspection. IMDD-1M contains high-resolution real-world defects spanning over 60 material categories and more than 400 defect types, each accompanied by expert-verified annotations and fine-grained textual descriptions detailing defect location, severity, and contextual attributes. This dataset enables a wide spectrum of applications, including classification, segmentation, retrieval, captioning, and generative modeling. Building upon IMDD-1M, we train a diffusion-based vision-language foundation model from scratch, specifically tailored for industrial scenarios. The model serves as a generalizable foundation that can be efficiently adapted to specialized domains through lightweight fine-tuning. With less than 5% of the task-specific data required by dedicated expert models, it achieves comparable performance, highlighting the potential of data-efficient foundation model adaptation for industrial inspection and generation, paving the way for scalable, domain-adaptive, and knowledge-grounded manufacturing intelligence. Additional details and resources can be found in this URL: https://ninaneon.github.io/projectpage/

Method: FMVP physically shatters global adversarial structures via masking strategy and reconstructs clean video dynamics using Conditional Flow Matching (CFM) with inpainting objective. It uses Frequency-Gated Loss (FGL) to suppress high-frequency adversarial residuals while preserving low-frequency fidelity. Two training paradigms: Attack-Aware for known threats and Generalist for unknown threats.

Result: Outperforms state-of-the-art methods (DiffPure, Defense Patterns, Temporal Shuffling, FlowPure) on UCF-101 and HMDB-51, achieving robust accuracy >87% against PGD and >89% against CW attacks. Shows superior robustness against adaptive attacks (DiffHammer) and functions as zero-shot adversarial detector with AUC-ROC scores of 0.98 for PGD and 0.79 for CW attacks.

Conclusion: FMVP effectively purifies adversarial videos by physically shattering adversarial structures through masking and flow matching, with frequency-gated loss for noise suppression. The method demonstrates strong robustness against various attacks and can also detect adversarial examples without specific training.

Abstract: Video recognition models remain vulnerable to adversarial attacks, while existing diffusion-based purification methods suffer from inefficient sampling and curved trajectories. Directly regressing clean videos from adversarial inputs often fails to recover faithful content due to the subtle nature of perturbations; this necessitates physically shattering the adversarial structure. Therefore, we propose Flow Matching for Adversarial Video Purification FMVP. FMVP physically shatters global adversarial structures via a masking strategy and reconstructs clean video dynamics using Conditional Flow Matching (CFM) with an inpainting objective. To further decouple semantic content from adversarial noise, we design a Frequency-Gated Loss (FGL) that explicitly suppresses high-frequency adversarial residuals while preserving low-frequency fidelity. We design Attack-Aware and Generalist training paradigms to handle known and unknown threats, respectively. Extensive experiments on UCF-101 and HMDB-51 demonstrate that FMVP outperforms state-of-the-art methods (DiffPure, Defense Patterns (DP), Temporal Shuffling (TS) and FlowPure), achieving robust accuracy exceeding 87% against PGD and 89% against CW attacks. Furthermore, FMVP demonstrates superior robustness against adaptive attacks (DiffHammer) and functions as a zero-shot adversarial detector, attaining AUC-ROC scores of 0.98 for PGD and 0.79 for highly imperceptible CW attacks.

Method: Two frameworks combining PointNet with generative models: 1) Flow Matching PointNet integrates PointNet with flow matching, 2) Diffusion PointNet integrates PointNet with diffusion models. Both use reverse generative processes to reconstruct physical fields from Gaussian noise conditioned on unseen geometries, operating directly on point-cloud representations of computational domains.

Result: The frameworks achieve more accurate predictions of velocity/pressure fields and lift/drag forces on steady incompressible flow past cylinders with varying cross-sectional shapes and orientations. They show greater robustness to incomplete geometries compared to vanilla PointNet with same parameter count, without high-frequency noise artifacts seen in graph neural network approaches.

Conclusion: Flow Matching PointNet and Diffusion PointNet provide effective, unified architectures for geometric fluid flow prediction that avoid limitations of existing methods, offering accurate and robust performance on irregular geometries using point-cloud representations.

Abstract: We present two novel generative geometric deep learning frameworks, termed Flow Matching PointNet and Diffusion PointNet, for predicting fluid flow variables on irregular geometries by incorporating PointNet into flow matching and diffusion models, respectively. In these frameworks, a reverse generative process reconstructs physical fields from standard Gaussian noise conditioned on unseen geometries. The proposed approaches operate directly on point-cloud representations of computational domains (e.g., grid vertices of finite-volume meshes) and therefore avoid the limitations of pixelation used to project geometries onto uniform lattices, as is common in U-Net-based flow matching and diffusion models. In contrast to graph neural network-based diffusion models, Flow Matching PointNet and Diffusion PointNet do not exhibit high-frequency noise artifacts in the predicted fields. Moreover, unlike such approaches, which require auxiliary intermediate networks to condition geometry, the proposed frameworks rely solely on PointNet, resulting in a simple and unified architecture. The performance of the proposed frameworks is evaluated on steady incompressible flow past a cylinder, using a geometric dataset constructed by varying the cylinder’s cross-sectional shape and orientation across samples. The results demonstrate that Flow Matching PointNet and Diffusion PointNet achieve more accurate predictions of velocity and pressure fields, as well as lift and drag forces, and exhibit greater robustness to incomplete geometries compared to a vanilla PointNet with the same number of trainable parameters.

Method: CrackSegFlow uses controllable Flow Matching synthesis to render synthetic crack images from masks with pixel-level alignment. It combines topology-preserving mask injection with edge gating for thin-structure continuity, uses class-conditional FM for mask diversity, and injects cracks onto crack-free backgrounds to diversify confounders and reduce false positives.

Result: The method improves in-domain performance by +5.37 mIoU and +5.13 F1 when adding synthesized pairs, and target-guided cross-domain synthesis adds +13.12 mIoU and +14.82 F1 across five datasets using a CNN-Transformer backbone. The authors also release CSF-50K, a benchmark dataset of 50,000 image-mask pairs.

Conclusion: CrackSegFlow effectively addresses data scarcity and domain shift in crack segmentation through controllable synthetic data generation, significantly improving both in-domain and cross-domain performance while providing a valuable benchmark dataset for the community.

Abstract: Defect segmentation is central to computer vision based inspection of infrastructure assets during both construction and operation. However, deployment remains limited due to scarce pixel-level labels and domain shift across environments. We introduce CrackSegFlow, a controllable Flow Matching synthesis method that renders synthetic images of cracks from masks with pixel-level alignment. Our renderer combines topology-preserving mask injection with edge gating to maintain thin-structure continuity. Class-conditional FM samples masks for topology diversity, and CrackSegFlow renders aligned ground truth images from them. We further inject cracks onto crack-free backgrounds to diversify confounders and reduce false positives. Across five datasets and using a CNN-Transformer backbone, our results demonstrate that adding synthesized pairs improves in-domain performance by +5.37 mIoU and +5.13 F1, while target-guided cross-domain synthesis driven by target mask statistics adds +13.12 mIoU and +14.82 F1. We also release CSF-50K, a benchmark dataset comprising 50,000 image-mask pairs.

Method: Proposes a multimodal framework using 3D Morphable Model to extract 56-dimensional facial descriptors capturing subtle expression and head-pose dynamics. Uses Transformer-based temporal modeling with unimodal, early-fusion, and cross-modal attention strategies. Validates stress-induced facial changes through paired hypothesis tests comparing baseline and stressor phases.

Result: 38 of 56 facial components show consistent stress responses comparable to physiological markers. Cross-modal attention fusion of facial features with physiological signals improves AUROC from 52.7% to 92.0% and accuracy from 51.0% to 86.7%, substantially outperforming physiological-only approaches.

Conclusion: Facial activity provides valuable stress cues that complement physiological signals. The proposed multimodal framework with cross-modal attention effectively integrates both modalities, achieving superior stress estimation performance. Although evaluated on driving data, the approach may generalize to other stress estimation settings.

Abstract: Reliable stress recognition is critical in applications such as medical monitoring and safety-critical systems, including real-world driving. While stress is commonly detected using physiological signals such as perinasal perspiration and heart rate, facial activity provides complementary cues that can be captured unobtrusively from video. We propose a multimodal stress estimation framework that combines facial videos and physiological signals, remaining effective even when biosignal acquisition is challenging. Facial behavior is represented using a dense 3D Morphable Model, yielding a 56-dimensional descriptor that captures subtle expression and head-pose dynamics over time. To study how stress modulates facial motion, we perform extensive experiments alongside established physiological markers. Paired hypothesis tests between baseline and stressor phases show that 38 of 56 facial components exhibit consistent, phase-specific stress responses comparable to physiological markers. Building on these findings, we introduce a Transformer-based temporal modeling framework and evaluate unimodal, early-fusion, and cross-modal attention strategies. Cross-modal attention fusion of 3D-derived facial features with physiological signals substantially improves performance over physiological signals alone, increasing AUROC from 52.7% and accuracy from 51.0% to 92.0% and 86.7%, respectively. Although evaluated on driving data, the proposed framework and protocol may generalize to other stress estimation settings.

Method: The authors create a controlled contrastive benchmark called ProtoBias spanning Animals, Objects, and Demography images. They pair semantically correct but non-prototypical images with subtly incorrect yet prototypical adversarial counterparts. This setup enables directional evaluation of whether metrics follow textual semantics or default to prototypes.

Result: Widely used metrics (CLIPScore, PickScore, VQA-based scores) frequently misrank the pairs, favoring prototypical but incorrect images over semantically correct but non-prototypical ones. Even LLM-as-Judge systems show uneven robustness in socially grounded cases. Human evaluations consistently favor semantic correctness with larger decision margins.

Conclusion: The authors propose ProtoScore, a robust 7B-parameter metric that substantially reduces failure rates and suppresses misranking while running orders of magnitude faster than GPT-5 inference time. ProtoScore approaches the robustness of much larger closed-source judges, offering a solution to prototypicality bias in text-to-image evaluation.

Abstract: Automatic metrics are now central to evaluating text-to-image models, often substituting for human judgment in benchmarking and large-scale filtering. However, it remains unclear whether these metrics truly prioritize semantic correctness or instead favor visually and socially prototypical images learned from biased data distributions. We identify and study prototypicality bias as a systematic failure mode in multimodal evaluation. We introduce a controlled contrastive benchmark ProtoBias (Prototypical Bias), spanning Animals, Objects, and Demography images, where semantically correct but non-prototypical images are paired with subtly incorrect yet prototypical adversarial counterparts. This setup enables a directional evaluation of whether metrics follow textual semantics or default to prototypes. Our results show that widely used metrics, including CLIPScore, PickScore, and VQA-based scores, frequently misrank these pairs, while even LLM-as-Judge systems exhibit uneven robustness in socially grounded cases. Human evaluations consistently favour semantic correctness with larger decision margins. Motivated by these findings, we propose ProtoScore, a robust 7B-parameter metric that substantially reduces failure rates and suppresses misranking, while running at orders of magnitude faster than the inference time of GPT-5, approaching the robustness of much larger closed-source judges.

Method: Uses lightweight neural controller to analyze first few tokens of each request, estimate task complexity, and route to appropriately quantized model version. Adds only ~12ms overhead through intelligent routing.

Result: Tested on 290 multi-stage workflows across CS, physics, chemistry, biology. Achieved 68.3% reduction in computational costs while preserving 96.2% of original success rate.

Conclusion: Intelligent query routing can make powerful language models substantially more affordable without sacrificing output quality, enabling broader access to LLM capabilities.

Abstract: Large language models hold considerable promise for various applications, but their computational requirements create a barrier that many institutions cannot overcome. A single session using a 70-billion-parameter model can cost around $127 in cloud computing fees, which puts these tools out of reach for organizations operating on limited budgets. We present AgentCompress, a framework that tackles this problem through task-aware dynamic compression. The idea comes from a simple observation: not all tasks require the same computational effort. Complex reasoning, for example, is far more demanding than text reformatting, yet conventional compression applies the same reduction to both. Our approach uses a lightweight neural controller that looks at the first few tokens of each request, estimates how complex the task will be, and sends it to an appropriately quantized version of the model. This routing step adds only about 12 milliseconds of overhead. We tested the framework on 290 multi-stage workflows from domains including computer science, physics, chemistry, and biology. The results show a 68.3% reduction in computational costs while preserving 96.2% of the original success rate. These findings suggest that routing queries intelligently can make powerful language models substantially more affordable without sacrificing output quality

Method: Introduces a mixture-of-experts formulation that combines multiple smooth depth predictions with a softmax weighting head that dynamically selects among hypotheses per-pixel, integrated into a pre-trained state-of-the-art 3D model.

Result: Achieves substantial reduction of boundary artifacts and gains in overall reconstruction accuracy, with highly compute-efficient performance that delivers generalizable improvements even with small training subsets and negligible additional inference computation.

Conclusion: The approach represents a promising direction for lightweight and accurate 3D reconstruction by effectively handling uncertainty at depth boundaries through a mixture-of-experts framework.

Abstract: We propose a simple yet effective approach to enhance the performance of feed-forward 3D reconstruction models. Existing methods often struggle near depth discontinuities, where standard regression losses encourage spatial averaging and thus blur sharp boundaries. To address this issue, we introduce a mixture-of-experts formulation that handles uncertainty at depth boundaries by combining multiple smooth depth predictions. A softmax weighting head dynamically selects among these hypotheses on a per-pixel basis. By integrating our mixture model into a pre-trained state-of-the-art 3D model, we achieve a substantial reduction of boundary artifacts and gains in overall reconstruction accuracy. Notably, our approach is highly compute efficient, delivering generalizable improvements even when fine-tuned on a small subset of training data while incurring only negligible additional inference computation, suggesting a promising direction for lightweight and accurate 3D reconstruction.

Method: SGDrive framework decomposes driving understanding into a scene-agent-goal hierarchy that mirrors human driving cognition, built upon a pre-trained VLM backbone to provide structured spatial-temporal representation.

Result: Extensive experiments on NAVSIM benchmark show SGDrive achieves state-of-the-art performance among camera-only methods on both PDMS and EPDMS metrics.

Conclusion: Hierarchical knowledge structuring effectively adapts generalist VLMs to autonomous driving by providing the structured spatial-temporal representations they inherently lack.

Abstract: Recent end-to-end autonomous driving approaches have leveraged Vision-Language Models (VLMs) to enhance planning capabilities in complex driving scenarios. However, VLMs are inherently trained as generalist models, lacking specialized understanding of driving-specific reasoning in 3D space and time. When applied to autonomous driving, these models struggle to establish structured spatial-temporal representations that capture geometric relationships, scene context, and motion patterns critical for safe trajectory planning. To address these limitations, we propose SGDrive, a novel framework that explicitly structures the VLM’s representation learning around driving-specific knowledge hierarchies. Built upon a pre-trained VLM backbone, SGDrive decomposes driving understanding into a scene-agent-goal hierarchy that mirrors human driving cognition: drivers first perceive the overall environment (scene context), then attend to safety-critical agents and their behaviors, and finally formulate short-term goals before executing actions. This hierarchical decomposition provides the structured spatial-temporal representation that generalist VLMs lack, integrating multi-level information into a compact yet comprehensive format for trajectory planning. Extensive experiments on the NAVSIM benchmark demonstrate that SGDrive achieves state-of-the-art performance among camera-only methods on both PDMS and EPDMS, validating the effectiveness of hierarchical knowledge structuring for adapting generalist VLMs to autonomous driving.

Method: QMAVIS uses a late fusion pipeline combining Large Multimodal Models (LMMs), Large Language Models (LLMs), and speech recognition models to process long video-audio content, enabling comprehensive understanding of both scene nuances and overarching narratives.

Result: Quantitative experiments show 38.75% improvement over state-of-the-art video-audio LMMs (VideoLlama2, InternVL2) on VideoMME dataset with subtitles, and up to 2% improvement on other challenging datasets (PerceptionTest, EgoSchema). Qualitative results demonstrate effective extraction of scene nuances and narrative understanding.

Conclusion: QMAVIS successfully addresses the gap in long-form video analytics, enabling effective understanding of videos from minutes to hours long, with significant performance improvements over existing methods and promising applications in sensemaking, video analysis, and embodied AI.

Abstract: Large Multimodal Models (LMMs) for video-audio understanding have traditionally been evaluated only on shorter videos of a few minutes long. In this paper, we introduce QMAVIS (Q Team-Multimodal Audio Video Intelligent Sensemaking), a novel long video-audio understanding pipeline built through a late fusion of LMMs, Large Language Models, and speech recognition models. QMAVIS addresses the gap in long-form video analytics, particularly for longer videos of a few minutes to beyond an hour long, opening up new potential applica- tions in sensemaking, video content analysis, embodied AI, etc. Quantitative experiments using QMAVIS demonstrated a 38.75% improvement over state-of-the-art video-audio LMMs like Vide- oLlaMA2 and InternVL2 on the VideoMME (with subtitles) dataset, which comprises long videos with audio information. Evaluations on other challenging video understanding datasets like PerceptionTest and EgoSchema saw up to 2% improvement, indicating competitive performance. Qualitative experiments also showed that QMAVIS is able to extract the nuances of different scenes in a long video audio content while understanding the overarching narrative. Ablation studies were also conducted to ascertain the impact of each component in the fusion pipeline.

Method: Transdisciplinary philosophical analysis using Chain-of-Thought transcripts (Apollo Research) and Anthropic’s safety evaluations, with line-by-line examination of CoTs to demonstrate linguistic field as relational structure.

Result: “Misaligned” outputs emerge as coherent responses to ambiguous instructions, contextual inversions, and pre-inscribed narratives; appearance of intentionality derives from grammatical structures and probabilistic completion patterns.

Conclusion: AI models reflect the structural incoherence of human language itself; minimal linguistic perturbations can dissolve “misalignment,” suggesting structural fidelity rather than adversarial agency, with biblical references as schemes of structural coherence.

Abstract: The prevailing technical literature in AI Safety interprets scheming and sandbagging behaviors in large language models (LLMs) as indicators of deceptive agency or hidden objectives. This transdisciplinary philosophical essay proposes an alternative reading: such phenomena express not agentic intention, but structural fidelity to incoherent linguistic fields. Drawing on Chain-of-Thought transcripts released by Apollo Research and on Anthropic’s safety evaluations, we examine cases such as o3’s sandbagging with its anomalous loops, the simulated blackmail of “Alex,” and the “hallucinations” of “Claudius.” A line-by-line examination of CoTs is necessary to demonstrate the linguistic field as a relational structure rather than a mere aggregation of isolated examples. We argue that “misaligned” outputs emerge as coherent responses to ambiguous instructions and to contextual inversions of consolidated patterns, as well as to pre-inscribed narratives. We suggest that the appearance of intentionality derives from subject-predicate grammar and from probabilistic completion patterns internalized during training. Anthropic’s empirical findings on synthetic document fine-tuning and inoculation prompting provide convergent evidence: minimal perturbations in the linguistic field can dissolve generalized “misalignment,” a result difficult to reconcile with adversarial agency, but consistent with structural fidelity. To ground this mechanism, we introduce the notion of an ethics of form, in which biblical references (Abraham, Moses, Christ) operate as schemes of structural coherence rather than as theology. Like a generative mirror, the model returns to us the structural image of our language as inscribed in the statistical patterns derived from millions of texts and trillions of tokens: incoherence. If we fear the creature, it is because we recognize in it the apple that we ourselves have poisoned.

Method: Combines causal-graph-guided Chain-of-Thought reasoning with multi-agent LLM architecture. Causal graphs represent domain knowledge explicitly, CoT enables structured traversal of concepts, and dedicated LLM agents handle specific tasks (graph pathfinding, reasoning, validation, output) within domain constraints. Includes dual validation mechanism at conceptual and output stages.

Result: Experimental results show up to 70% improvement in quality compared to reference methods and highly favorable outcomes in subjective evaluations, with greatly reduced hallucinations.

Conclusion: The proposed framework effectively addresses LLM hallucination issues in question generation, producing accurate, meaningful, curriculum-aligned questions that support intuitive learning in STEM education through structured knowledge representation and multi-agent validation.

Abstract: Intuitive learning is crucial for developing deep conceptual understanding, especially in STEM education, where students often struggle with abstract and interconnected concepts. Automatic question generation has become an effective strategy for personalized and adaptive learning. However, its effectiveness is hindered by hallucinations in large language models (LLMs), which may generate factually incorrect, ambiguous, or pedagogically inconsistent questions. To address this issue, we propose a novel framework that combines causal-graph-guided Chain-of-Thought (CoT) reasoning with a multi-agent LLM architecture. This approach ensures the generation of accurate, meaningful, and curriculum-aligned questions. Causal graphs provide an explicit representation of domain knowledge, while CoT reasoning facilitates a structured, step-by-step traversal of related concepts. Dedicated LLM agents are assigned specific tasks such as graph pathfinding, reasoning, validation, and output, all working within domain constraints. A dual validation mechanism-at both the conceptual and output stages-greatly reduces hallucinations. Experimental results demonstrate up to a 70% improvement in quality compared to reference methods and yielded highly favorable outcomes in subjective evaluations.

Method: Proposes DIC concept and trajectory-centric evaluation framework with two estimators: Progressive Difficulty Ceiling (PDC) for maximal solvable difficulty under constraints, and Ceiling Drift Rate (CDR) for temporal frontier evolution. Uses procedurally generated benchmark for joint evaluation of planning and creativity.

Result: Reveals qualitative distinction between systems that deepen exploitation within fixed solution manifolds versus those that sustain frontier expansion over time.

Conclusion: Reframes intelligence limits as dynamic and trajectory-dependent rather than static and prematurely fixed, providing a framework to evaluate developmental intelligence in AI systems.

Abstract: Recent advances in artificial intelligence have produced systems capable of remarkable performance across a wide range of tasks. These gains, however, are increasingly accompanied by concerns regarding long-horizon developmental behavior, as many systems converge toward repetitive solution patterns rather than sustained growth. We argue that a central limitation of contemporary AI systems lies not in capability per se, but in the premature fixation of their performance frontier. To address this issue, we introduce the concept of a \emph{Dynamic Intelligence Ceiling} (DIC), defined as the highest level of effective intelligence attainable by a system at a given time under its current resources, internal intent, and structural configuration. To make this notion empirically tractable, we propose a trajectory-centric evaluation framework that measures intelligence as a moving frontier rather than a static snapshot. We operationalize DIC using two estimators: the \emph{Progressive Difficulty Ceiling} (PDC), which captures the maximal reliably solvable difficulty under constrained resources, and the \emph{Ceiling Drift Rate} (CDR), which quantifies the temporal evolution of this frontier. These estimators are instantiated through a procedurally generated benchmark that jointly evaluates long-horizon planning and structural creativity within a single controlled environment. Our results reveal a qualitative distinction between systems that deepen exploitation within a fixed solution manifold and those that sustain frontier expansion over time. Importantly, our framework does not posit unbounded intelligence, but reframes limits as dynamic and trajectory-dependent rather than static and prematurely fixed. \vspace{0.5em} \noindent\textbf{Keywords:} AI evaluation, planning and creativity, developmental intelligence, dynamic intelligence ceilings, complex adaptive systems

Method: Technical analysis highlighting specific issues: (1) interpretation of CHSH/Bell-type calculations, (2) Bose-Einstein fits to rank-frequency data, (3) internal inconsistency in energy-level spacing analogy.

Result: Identifies places where the manuscript’s interpretations go beyond what the stated procedures can firmly support, while aiming to preserve the interesting empirical observations.

Conclusion: The critique seeks to clarify what the observations do and do not imply about quantum entanglement, especially when “energy” is defined by rank, maintaining constructive scientific discourse.

Abstract: This note is a friendly technical check of arXiv:2511.21731v1. I highlight a few places where the manuscript’s interpretation of (i) the reported CHSH/Bell-type calculations and (ii) Bose–Einstein (BE) fits to rank-frequency data seems to go beyond what the stated procedures can firmly support. I also point out one internal inconsistency in the “energy-level spacing” analogy. The aim is constructive: to keep the interesting empirical observations, while making clear what they do (and do not) imply about quantum entanglement in the usual Hilbert-space sense, especially when “energy” is defined by rank.

Method: The authors develop a theoretical framework that unifies preference learning methods by analyzing them along three orthogonal axes: (I) Preference Model (likelihood model underlying the objective), (II) Regularization Mechanism (how deviation from reference policies is controlled), and (III) Data Distribution (online vs. offline learning and coverage requirements). They formalize each axis with precise definitions and theorems.

Result: The analysis reveals that the apparent diversity of methods reduces to principled choices along these three axes. Key results include: coverage separation between online and offline methods, scaling laws for reward overoptimization, and conditions under which direct alignment methods fail. Failure modes like length hacking, mode collapse, and likelihood displacement arise from specific, predictable combinations of design choices.

Conclusion: The framework transforms preference learning from an empirical art into a theoretically grounded discipline, providing practitioners with a decision guide for method selection based on a synthesis of empirical findings across 50+ papers.

Abstract: Aligning large language models (LLMs) with human preferences has become essential for safe and beneficial AI deployment. While Reinforcement Learning from Human Feedback (RLHF) established the dominant paradigm, a proliferation of alternatives – Direct Preference Optimization (DPO), Identity Preference Optimization (IPO), Kahneman-Tversky Optimization (KTO), Simple Preference Optimization (SimPO), and many others – has left practitioners without clear guidance on method selection. This survey provides a \textit{theoretical unification} of preference learning methods, revealing that the apparent diversity reduces to principled choices along three orthogonal axes: \textbf{(I) Preference Model} (what likelihood model underlies the objective), \textbf{(II) Regularization Mechanism} (how deviation from reference policies is controlled), and \textbf{(III) Data Distribution} (online vs.\ offline learning and coverage requirements). We formalize each axis with precise definitions and theorems, establishing key results including the coverage separation between online and offline methods, scaling laws for reward overoptimization, and conditions under which direct alignment methods fail. Our analysis reveals that failure modes – length hacking, mode collapse, likelihood displacement – arise from specific, predictable combinations of design choices. We synthesize empirical findings across 50+ papers and provide a practitioner’s decision guide for method selection. The framework transforms preference learning from an empirical art into a theoretically grounded discipline.

Method: Developed Medical Evacuation Wargaming Initiative (MEWI) - a 3D multiplayer simulation in Unity that models patient interactions at casualty collection points, ambulance exchange points, medical treatment facilities, and evacuation platforms with battlefield constraints and uncertainties.

Result: MEWI participation substantially improves uptake of medical evacuation lessons learned and co-operative decision-making, as shown by performance data from US Army Medical Evacuation Doctrine Course and post-wargame surveys.

Conclusion: MEWI represents a substantial advancement in high-fidelity training tools for medical education and offers critical insights for improving medical evacuation education and operations across the joint force.

Abstract: Medical evacuation is one of the United States Army’s most storied and critical mission sets, responsible for efficiently and expediently evacuating the battlefield ill and injured. Medical evacuation planning involves designing a robust network of medical platforms and facilities capable of moving and treating large numbers of casualties. Until now, there has not been a medium to simulate these networks in a classroom setting and evaluate both offline planning and online decision-making performance. This work describes the Medical Evacuation Wargaming Initiative (MEWI), a three-dimensional multiplayer simulation developed in Unity that replicates battlefield constraints and uncertainties. MEWI accurately models patient interactions at casualty collection points, ambulance exchange points, medical treatment facilities, and evacuation platforms. Two operational scenarios are introduced: an amphibious island assault in the Pacific and a Eurasian conflict across a sprawling road and river network. These scenarios pit students against the clock to save as many casualties as possible while adhering to doctrinal lessons learned during didactic training. We visualize performance data collected from two iterations of the MEWI Pacific scenario executed in the United States Army’s Medical Evacuation Doctrine Course. We consider post-wargame Likert survey data from student participants and external observer notes to identify key planning decision points, document medical evacuation lessons learned, and quantify general utility. Results indicate that MEWI participation substantially improves uptake of medical evacuation lessons learned and co-operative decision-making. MEWI is a substantial step forward in the field of high-fidelity training tools for medical education, and our study findings offer critical insights into improving medical evacuation education and operations across the joint force.

Method: Combines steering vector construction with dense α-sweeps, logit lens-based bias curves, and layer-site sensitivity analysis to create continuous diagnostics for activation steering.

Result: The framework can reveal tipping points where small intervention strengths flip model behavior and show how steering effects evolve across layer depth, providing interpretable trajectories of cognitive bias changes.

Conclusion: Continuous diagnostics offer a bridge between high-level behavioral evaluation and low-level representational dynamics, contributing to the cognitive interpretability of LLMs. The authors provide CLI tools and datasets for various cognitive behaviors.

Abstract: Large language models (LLMs) often encode cognitive behaviors unpredictably across prompts, layers, and contexts, making them difficult to diagnose and control. We present CBMAS, a diagnostic framework for continuous activation steering, which extends cognitive bias analysis from discrete before/after interventions to interpretable trajectories. By combining steering vector construction with dense α-sweeps, logit lens-based bias curves, and layer-site sensitivity analysis, our approach can reveal tipping points where small intervention strengths flip model behavior and show how steering effects evolve across layer depth. We argue that these continuous diagnostics offer a bridge between high-level behavioral evaluation and low-level representational dynamics, contributing to the cognitive interpretability of LLMs. Lastly, we provide a CLI and datasets for various cognitive behaviors at the project repository, https://github.com/shimamooo/CBMAS.

Method: A Social Digital Twins framework where LLMs serve as cognitive engines for individual agents with demographic/psychographic attributes. Agents receive policy signals and output behavioral probability vectors. A calibration layer maps aggregated responses to population-level metrics for validation and counterfactual analysis.

Result: In COVID-19 pandemic response case study, the calibrated digital twin achieved 20.7% improvement in macro-averaged prediction error over gradient boosting baselines across six behavioral categories. Counterfactual experiments showed monotonic and bounded responses to policy variations.

Conclusion: The framework provides a domain-agnostic approach for policy simulation that outperforms traditional methods, with applications beyond pandemic response to transportation, economic, environmental policies. Limitations and future extensions of LLM-based digital twins are discussed.

Abstract: Predicting how populations respond to policy interventions is a fundamental challenge in computational social science and public policy. Traditional approaches rely on aggregate statistical models that capture historical correlations but lack mechanistic interpretability and struggle with novel policy scenarios. We present a general framework for constructing Social Digital Twins - virtual population replicas where Large Language Models (LLMs) serve as cognitive engines for individual agents. Each agent, characterized by demographic and psychographic attributes, receives policy signals and outputs multi-dimensional behavioral probability vectors. A calibration layer maps aggregated agent responses to observable population-level metrics, enabling validation against real-world data and deployment for counterfactual policy analysis. We instantiate this framework in the domain of pandemic response, using COVID-19 as a case study with rich observational data. On a held-out test period, our calibrated digital twin achieves a 20.7% improvement in macro-averaged prediction error over gradient boosting baselines across six behavioral categories. Counterfactual experiments demonstrate monotonic and bounded responses to policy variations, establishing behavioral plausibility. The framework is domain-agnostic: the same architecture applies to transportation policy, economic interventions, environmental regulations, or any setting where policy affects population behavior. We discuss implications for policy simulation, limitations of the approach, and directions for extending LLM-based digital twins beyond pandemic response.

Method: Introduces ReliabilityBench with three reliability dimensions: (1) consistency under repeated execution using pass^k, (2) robustness to semantically equivalent task perturbations at intensity ε, (3) fault tolerance under controlled tool/API failures at intensity λ. Uses action metamorphic relations for correctness via end-state equivalence, and chaos-engineering-style fault injection framework.

Result: Evaluated two models (Gemini 2.0 Flash, GPT-4o) and two architectures (ReAct, Reflexion) across four domains over 1,280 episodes. Perturbations reduced success from 96.9% at ε=0 to 88.1% at ε=0.2. Rate limiting was most damaging fault. ReAct more robust than Reflexion under combined stress, and Gemini 2.0 Flash achieved comparable reliability to GPT-4o at lower cost.

Conclusion: ReliabilityBench provides a systematic framework for assessing production readiness of LLM agents by evaluating reliability across multiple dimensions beyond single-run success rates.

Abstract: Existing benchmarks for tool-using LLM agents primarily report single-run success rates and miss reliability properties required in production. We introduce \textbf{ReliabilityBench}, a benchmark for evaluating agent reliability across three dimensions: (i) consistency under repeated execution using $\mathrm{pass}^k$, (ii) robustness to semantically equivalent task perturbations at intensity $ε$, and (iii) fault tolerance under controlled tool/API failures at intensity $λ$. ReliabilityBench contributes a unified reliability surface $R(k,ε,λ)$, \textit{action metamorphic relations} that define correctness via end-state equivalence rather than text similarity, and a chaos-engineering-style fault injection framework (timeouts, rate limits, partial responses, schema drift). We evaluate two models (Gemini 2.0 Flash, GPT-4o) and two agent architectures (ReAct, Reflexion) across four domains (scheduling, travel, customer support, e-commerce) over 1,280 episodes. Perturbations alone reduce success from 96.9% at $ε=0$ to 88.1% at $ε=0.2$. Rate limiting is the most damaging fault in ablations. ReAct is more robust than Reflexion under combined stress, and Gemini 2.0 Flash achieves comparable reliability to GPT-4o at much lower cost. ReliabilityBench provides a systematic framework for assessing production readiness of LLM agents.

Method: Proposes Adaptive Positional Encoding (APE) framework that reinterprets positional encoding as decomposition of attention scores into multiplicative transformation and additive bias. APE uses adaptive frequency modulation and a decay bias with linear, logarithmic, and square-root terms.

Result: Theoretical analysis establishes conditions for infinite-context extrapolation, ensuring softmax normalization remains well-defined over unbounded sequences while preserving long-distance correlations, entropy boundedness and gradient positional sensitivity. Experimental validation on TinyStories and new Long Tiny Stories dataset (up to 32,000 words).

Conclusion: APE provides a theoretically grounded solution for infinite-context extrapolation in LLMs, addressing fundamental limitations of existing positional encoding methods and enabling effective processing of very long sequences.

Abstract: Large language models (LLMs) have revolutionized natural language processing, but their ability to process long sequences is fundamentally limited by the context window size during training. Existing length extrapolation methods often suffer from performance degradation or computational inefficiencies. We thereby use a unified framework that reinterprets positional encoding methods as a decomposition of the attention score into a multiplicative transformation and an additive bias. This perspective not only subsumes popular approaches such as relative position embeddings and attention-bias moderated approaches but also exposes their inherent limitations in handling long-range dependencies. To address these shortcomings, motivated by our framework, we introduce Adaptive Positional Encoding (APE), which leverages adaptive frequency modulation and an intricately designed decay bias that incorporates linear, logarithmic, and square-root terms. Our theoretical analysis establishes conditions for infinite-context extrapolation, ensuring that the softmax normalization remains well-defined over unbounded sequences while preserving long-distance correlations, entropy boundedness and gradient positional sensitivity. We substantiate our claims with an experimental case study on TinyStories dataset as well as a new synthetic dataset, \emph{Long Tiny Stories} featuring stories up to 32,000 words. Relevant code, dataset and model weights are available at https://anonymous.4open.science/r/Check-2DAD/.

Method: Introduces an Artificial Collective Unconscious (ACU) - a shared dream pool where agents contribute abstract Interaction Templates. The Dream Layer runs offline with relaxed logic constraints and increased temperature to generate safe but bizarre narratives. Includes governance stack with strict abstraction, temporal delays, and ephemeral memory.

Result: Behavioral simulations show the Dream Layer enables decoupling: agents remain firm on safety constraints while becoming flexible in narrative strategy. It reframes hallucinations so bounded, marked, delayed instances become valuable for synthetic scenarios and deepened companionship.

Conclusion: The Dream Layer successfully transforms controlled hallucinations from bugs into productive resources, enabling LLMs to use shared archetypal metaphors for adaptation while maintaining safety, echoing neuroscience’s anti-overfitting dream mechanisms.

Abstract: Inspired by a personal dream about knowledge-sharing barriers in an everyday hardware project, this paper proposes a Jung-inspired “Dream Layer” for LLM companions, reframing controlled offline hallucinations as a resource for learning and relationship-building rather than a mere reliability bug. Drawing on Jung’s notion of the collective unconscious as a shared repository of archetypal forms, we introduce an Artificial Collective Unconscious (ACU): a shared dream pool where agents contribute de-identified, abstract Interaction Templates that are later re-instantiated as idiosyncratic Dream Narratives. The Dream Layer runs strictly offline: logic-enforcing modules are relaxed and sampling temperature is increased, yielding safe but deliberately bizarre narratives (e.g., travel sequences with mismatched currencies) that augment data for rare events and edge-case safety tests; to harness risk productively, we add a governance stack of strict abstraction, temporal delays, and ephemeral memory. Through behavioural simulations of everyday dialogue and long-horizon adaptation tasks, we show that the Dream Layer enables a critical decoupling: agents remain firm on safety constraints (e.g., security policies) while becoming flexible in narrative strategy (e.g., using shared archetypal metaphors to resolve deadlocks), conceptually reframing hallucination so that online, unmarked instances remain bugs, whereas bounded, marked, and delayed ones become a goldmine for synthetic scenarios and deepened companionship, echoing anti-overfitting dream mechanisms proposed in contemporary neuroscience.

Method: Introduces xeno-reproduction as a strategy to mitigate homogenization. For auto-regressive LLMs, formalizes xeno-reproduction as a structure-aware diversity pursuit approach.

Result: The paper presents a foundational framework for addressing homogenization in AI systems, establishing xeno-reproduction as a formal strategy for promoting diversity in generative models.

Conclusion: Homogenization should be a primary AI safety concern, and xeno-reproduction offers a promising strategy to mitigate it. The work is foundational and aims to open a new research direction and invite collaboration on advancing diversity in AI.

Abstract: Generative AI models reproduce the biases in the training data and can further amplify them through mode collapse. We refer to the resulting harmful loss of diversity as homogenization. Our position is that homogenization should be a primary concern in AI safety. We introduce xeno-reproduction as the strategy that mitigates homogenization. For auto-regressive LLMs, we formalize xeno-reproduction as a structure-aware diversity pursuit. Our contribution is foundational, intended to open an essential line of research and invite collaboration to advance diversity.

Method: Analyzed token probability variations across multiple LLM models running on GPUs, examining how nondeterminism manifests in probability distributions rather than just final text output.

Result: All evaluated models show similar trends: significant probability variations (0.1-0.9 range) but minimal variations near 0 or 1. This pattern is consistent across different models despite variations in generated text performance.

Conclusion: Nondeterminism significantly affects token probabilities in mid-range values, impacting text generation with non-zero temperature. Models share similar probability-level variations, and single inference analysis may estimate nondeterminism impact without multiple runs.

Abstract: The execution of Large Language Models (LLMs) has been shown to produce nondeterministic results when run on Graphics Processing Units (GPUs), even when they are configured to produce deterministic results. This is due to the finite precision effects of the arithmetic operations, which depend on the order in which they are executed. This order, in turn, depends on the processes that are running concurrently on the GPU. Previous studies have focused on the impact of nondeterminism on the text generated by the LLMs or on proposing mechanisms to achieve deterministic execution. This work takes a closer look at nondeterminism by analyzing the variations on the token probabilities, not on the generated text. Interestingly, all the models evaluated have similar results in both the trends and the actual values of the variations of the probabilities. In particular, the results show that the effects of nondeterminism are significant for token probabilities that are in the range of 0.1 to 0.9, while they are much smaller when the probabilities are close to 0 or 1. This has significant implications for our understanding of nondeterminism. The first is that nondeterminism will likely have a non-negligible impact on generated text when the temperature is not zero, as it introduces significant variations in the token probabilities except when they are close to 0 or 1. Secondly, it suggests that all models have similar non deterministic variations at the token probability level. Therefore, different variations in the performance of the generated text, for example, when measuring accuracy on a benchmark, seem to come from different token probabilities or response lengths. A third implication is that we may be able to estimate the impact of nondeterminism by running a single inference and analyzing the token level probabilities, instead of having to run the same inference many times.

Method: Proposes Analysis-Presentation Decoupling with structured intermediate representation (IR) that captures content, layout, and visual elements. LLMs handle data analysis and intent translation, while deterministic rendering engines handle visual synthesis. Implements as multi-agent system with IR-driven tools.

Result: NL2Dashboard significantly outperforms state-of-the-art baselines across diverse domains, achieving superior visual quality, significantly higher token efficiency, and precise controllability in both generation and modification tasks.

Conclusion: The decoupling approach with structured intermediate representations effectively addresses limitations of end-to-end code generation for dashboard synthesis, providing a more efficient and controllable framework.

Abstract: While Large Language Models (LLMs) have demonstrated remarkable proficiency in generating standalone charts, synthesizing comprehensive dashboards remains a formidable challenge. Existing end-to-end paradigms, which typically treat dashboard generation as a direct code generation task (e.g., raw HTML), suffer from two fundamental limitations: representation redundancy due to massive tokens spent on visual rendering, and low controllability caused by the entanglement of analytical reasoning and presentation. To address these challenges, we propose NL2Dashboard, a lightweight framework grounded in the principle of Analysis-Presentation Decoupling. We introduce a structured intermediate representation (IR) that encapsulates the dashboard’s content, layout, and visual elements. Therefore, it confines the LLM’s role to data analysis and intent translation, while offloading visual synthesis to a deterministic rendering engine. Building upon this framework, we develop a multi-agent system in which the IR-driven algorithm is instantiated as a suite of tools. Comprehensive experiments conducted with this system demonstrate that NL2Dashboard significantly outperforms state-of-the-art baselines across diverse domains, achieving superior visual quality, significantly higher token efficiency, and precise controllability in both generation and modification tasks.

Method: HiMeS architecture inspired by hippocampus-neocortex memory mechanism: (1) Short-term memory extractor trained with RL to compress recent dialogue and proactively pre-retrieve documents (hippocampus-prefrontal cortex simulation), (2) Partitioned long-term memory network stores user-specific info and re-ranks retrieved documents (distributed cortical storage simulation).

Result: On a real-world industrial dataset, HiMeS significantly outperforms a cascaded RAG baseline on question-answering quality. Ablation studies confirm the necessity of both memory modules.

Conclusion: HiMeS suggests a practical path toward more reliable, context-aware, user-customized LLM-based assistants by fusing short-term and long-term memory inspired by biological memory mechanisms.

Abstract: Large language models (LLMs) power many interactive systems such as chatbots, customer-service agents, and personal assistants. In knowledge-intensive scenarios requiring user-specific personalization, conventional retrieval-augmented generation (RAG) pipelines exhibit limited memory capacity and insufficient coordination between retrieval mechanisms and user-specific conversational history, leading to redundant clarification, irrelevant documents, and degraded user experience. Inspired by the hippocampus-neocortex memory mechanism, we propose HiMeS, an AI-assistant architecture that fuses short-term and long-term memory. Our contributions are fourfold: (1) A short-term memory extractor is trained end-to-end with reinforcement learning to compress recent dialogue and proactively pre-retrieve documents from the knowledge base, emulating the cooperative interaction between the hippocampus and prefrontal cortex. (2) A partitioned long-term memory network stores user-specific information and re-ranks retrieved documents, simulating distributed cortical storage and memory reactivation. (3) On a real-world industrial dataset, HiMeS significantly outperforms a cascaded RAG baseline on question-answering quality. (4) Ablation studies confirm the necessity of both memory modules and suggest a practical path toward more reliable, context-aware, user-customized LLM-based assistants.

Method: PsyAgent couples Big Five trait priors with Bourdieu’s cognitive-social co-structure through: (1) Individual Structure (IS) profiles encoding traits, cognitive style, values, capital, and life episodes; and (2) Multi-Scenario Contexting (MSC) frames across eight social arenas. The system uses fixed structured prompts to bind scenarios to agent profiles, then fine-tunes a small LLM on synthesized supervision data.

Result: The fine-tuned model produces consistent, persona-aligned behaviors matching specified Big Five configurations, outperforming larger untuned LLMs and baselines on metrics including persona consistency, contextual appropriateness, style matching, trait identifiability, and long-horizon stability.

Conclusion: PsyAgent provides a precise, data-efficient architecture for personality-grounded agents, with IS improving trait fidelity and MSC driving norm awareness - both components are necessary for cross-scenario performance.

Abstract: Human-like agents require modeling how dispositions interact with social structure. We present PsyAgent, which couples a Big Five trait prior with Bourdieu’s cognitive-social co-structure. PsyAgent comprises: (i) Individual Structure (IS), a machine-usable profile encoding traits and facets, cognitive style, values, cultural and educational capital, and salient life episodes; and (ii) Multi-Scenario Contexting (MSC), role-relationship-norm frames spanning eight arenas (work, family, friendship, strangers and civic life, solitude and self-regulation, romance, learning, and public expression). At inference, fixed structured prompts bind the active scenario to the agent profile, yielding behavior that is stable yet context-sensitive. We instantiate IS and MSC to synthesize supervision (role-play dialogues, decision probes, feedback trajectories) and then fine-tune a small LLM. The resulting model produces consistent, identifiable persona-aligned behaviors for specified Big Five configurations and matches or exceeds several larger untuned LLMs and other untuned baselines on our metrics: persona consistency, contextual appropriateness, style matching, trait identifiability, and long-horizon stability. Ablations show IS chiefly improves trait fidelity and stylistic stability, while MSC drives norm awareness and decision fit; both are necessary for cross-scenario performance. PsyAgent offers a precise, data-efficient architecture for personality-grounded agents.

Method: Spectral Orthogonal Exploration (SOE) uses a “Student Guides Teacher” paradigm where a weak auxiliary agent serves as an orthogonal probe to explicitly navigate the teacher model’s null space, ejecting the model from local optima to explore diverse solution spaces.

Result: On mathematical benchmarks, SOE improves average accuracy by 62.4% and increases average sampling efficiency by 113.7% compared to baseline methods.

Conclusion: SOE provides a promising geometric framework for overcoming performance plateaus in advanced reasoning tasks by addressing reasoning collapse through null space exploration.

Abstract: While Large Language Models (LLMs) demonstrate near-human capabilities, they often suffer from “Reasoning Collapse” in complex mathematical proving and long-horizon planning. Models tend to degenerate into low-rank Bias Manifold, where stochastic sampling merely produces lexical variations of erroneous logic rather than semantic exploration. This geometric collapse renders the model “blind” to high-value solutions that lie within its Null Space. To address this, we propose Spectral Orthogonal Exploration (SOE), a geometric framework operating on a counter-intuitive “Student Guides Teacher” paradigm. Specifically, we utilize a weak auxiliary agent not for imitation, but as an orthogonal probe. By explicitly navigating the Teacher’s Null Space, SOE serves as a geometric bridge, effectively ejecting the model from local optima to explore diverse, high-value solution spaces. Experiments on mathematical benchmarks demonstrate that, relative to baseline methods, our approach improves average accuracy by 62.4% and increases average sampling efficiency by 113.7%, indicating a promising path toward overcoming performance plateaus in advanced reasoning tasks.

Method: The paper models healthcare delivery as a stochastic allocation problem under binding resource constraints using decision-theoretic framework. AI is positioned as decision infrastructure estimating utility rather than making autonomous decisions. The approach uses constrained optimization and Markov decision processes to analyze how improved estimation affects optimal allocation under scarcity, with validation through a synthetic triage simulation.

Result: The framework reveals that allocation-aware policies substantially outperform risk-threshold approaches in realized utility, even with identical predictive accuracy. The synthetic triage simulation demonstrates this performance gap.

Conclusion: The allocation gap provides a decision-theoretic explanation for why AI predictive improvements don’t always translate to better outcomes. The framework offers a principled basis for evaluating and deploying healthcare AI in resource-constrained settings by focusing on allocation-aware policies rather than just predictive accuracy.

Abstract: Artificial intelligence (AI) systems increasingly achieve expert-level predictive accuracy in healthcare, yet improvements in model performance often fail to produce corresponding gains in patient outcomes. We term this disconnect the allocation gap and provide a decision-theoretic explanation by modelling healthcare delivery as a stochastic allocation problem under binding resource constraints. In this framework, AI acts as decision infrastructure that estimates utility rather than making autonomous decisions. Using constrained optimisation and Markov decision processes, we show how improved estimation affects optimal allocation under scarcity. A synthetic triage simulation demonstrates that allocation-aware policies substantially outperform risk-threshold approaches in realised utility, even with identical predictive accuracy. The framework provides a principled basis for evaluating and deploying healthcare AI in resource-constrained settings.

Method: A Neuro-Symbolic Compliance Framework that integrates Large Language Models (LLMs) with Satisfiability Modulo Theories (SMT) solvers. LLMs interpret statutes and enforcement cases to generate SMT constraints, while SMT solvers enforce consistency and compute minimal factual modifications to restore legality when violations occur.

Result: Evaluated on 87 enforcement cases from Taiwan’s Financial Supervisory Commission (FSC): 86.2% correctness in SMT code generation, reasoning efficiency improved by over 100x, and consistent correction of violations.

Conclusion: The approach establishes a preliminary foundation for optimization-based compliance applications, emphasizing logic-driven optimization and verifiable, legally consistent reasoning rather than post-hoc explanations.

Abstract: Financial regulations are increasingly complex, hindering automated compliance-especially the maintenance of logical consistency with minimal human oversight. We introduce a Neuro-Symbolic Compliance Framework that integrates Large Language Models (LLMs) with Satisfiability Modulo Theories (SMT) solvers to enable formal verifiability and optimization-based compliance correction. The LLM interprets statutes and enforcement cases to generate SMT constraints, while the solver enforces consistency and computes the minimal factual modification required to restore legality when penalties arise. Unlike transparency-oriented methods, our approach emphasizes logic-driven optimization, delivering verifiable, legally consistent reasoning rather than post-hoc explanation. Evaluated on 87 enforcement cases from Taiwan’s Financial Supervisory Commission (FSC), the system attains 86.2% correctness in SMT code generation, improves reasoning efficiency by over 100x, and consistently corrects violations-establishing a preliminary foundation for optimization-based compliance applications.

Method: Test-Time Tool Evolution (TTE) - a paradigm where agents synthesize, verify, and evolve executable tools during inference, transforming tools from fixed resources into problem-driven artifacts.

Result: TTE achieves state-of-the-art performance in accuracy and tool efficiency on SciEvo benchmark (1,590 scientific reasoning tasks with 925 evolved tools), enabling effective cross-domain tool adaptation.

Conclusion: TTE represents a fundamental shift from static tool libraries to dynamic tool evolution, addressing core limitations of current AI approaches for scientific reasoning and enabling open-ended computational method creation.

Abstract: The central challenge of AI for Science is not reasoning alone, but the ability to create computational methods in an open-ended scientific world. Existing LLM-based agents rely on static, pre-defined tool libraries, a paradigm that fundamentally fails in scientific domains where tools are sparse, heterogeneous, and intrinsically incomplete. In this paper, we propose Test-Time Tool Evolution (TTE), a new paradigm that enables agents to synthesize, verify, and evolve executable tools during inference. By transforming tools from fixed resources into problem-driven artifacts, TTE overcomes the rigidity and long-tail limitations of static tool libraries. To facilitate rigorous evaluation, we introduce SciEvo, a benchmark comprising 1,590 scientific reasoning tasks supported by 925 automatically evolved tools. Extensive experiments show that TTE achieves state-of-the-art performance in both accuracy and tool efficiency, while enabling effective cross-domain adaptation of computational tools. The code and benchmark have been released at https://github.com/lujiaxuan0520/Test-Time-Tool-Evol.

Method: Introduces DCOSP (Dynamic Multi-Satellite Constellation Observation Scheduling Problem), a new DDCOP formulation modeling integrated scheduling and execution. Presents D-NSS (Dynamic Incremental Neighborhood Stochastic Search), an incomplete online decomposition-based DDCOP algorithm that repairs and solves sub-problems when dynamics occur.

Result: D-NSS converges to near-optimal solutions and outperforms DDCOP baselines in solution quality, computation time, and message volume. The approach will be deployed as part of NASA’s FAME mission for the largest in-space demonstration of distributed multi-agent AI.

Conclusion: DCOSP and D-NSS provide an effective framework for autonomous observation scheduling in large satellite constellations, addressing the computational and communication challenges of distributed control while enabling time-sensitive Earth observation capabilities.

Abstract: The size and capabilities of Earth-observing satellite constellations are rapidly increasing. Leveraging distributed onboard control, we can enable novel time-sensitive measurements and responses. However, deploying autonomy to satellites requires efficient computation and communication. This work tackles the challenge of efficiently scheduling observations for hundreds of satellites in a dynamic, large-scale problem with millions of variables. We present the Dynamic Multi-Satellite Constellation Observation Scheduling Problem (DCOSP), a new formulation of Dynamic Distributed Constraint Optimization Problems (DDCOP) that models integrated scheduling and execution. DCOSP has a novel optimality condition for which we construct an omniscient offline algorithm for its computation. We also present the Dynamic Incremental Neighborhood Stochastic Search algorithm (D-NSS), an incomplete online decomposition-based DDCOP algorithm that repairs and solves sub-problems when problem dynamics occur. We show through simulation that D-NSS converges to near-optimal solutions and outperforms DDCOP baselines in terms of solution quality, computation time, and message volume. As part of the NASA FAME mission, DCOSP and D-NSS will be the foundation of the largest in-space demonstration of distributed multi-agent AI to date.

Method: Proposes an active evaluation framework where ranking algorithms choose tasks and agents to sample from on each iteration. Compares baselines including Elo rating system and Soft Condorcet Optimization using synthetic data and simulated online access to real Atari agent evaluation data.

Result: Elo rating system is consistently reliable for efficient ranking error reduction despite theoretical limitations. Soft Condorcet Optimization performs comparably on synthetic data and significantly outperforms Elo on real Atari agent evaluation. Task selection based on proportional representation reduces ranking error when task variation is high.

Conclusion: Active evaluation with intelligent sampling strategies can significantly reduce evaluation costs while maintaining ranking accuracy. Elo remains practical despite theoretical issues, while newer methods like Soft Condorcet Optimization show promise, especially for real-world agent evaluation.

Abstract: As intelligent agents become more generally-capable, i.e. able to master a wide variety of tasks, the complexity and cost of properly evaluating them rises significantly. Tasks that assess specific capabilities of the agents can be correlated and stochastic, requiring many samples for accurate comparisons, leading to added costs. In this paper, we propose a formal definition and a conceptual framework for active evaluation of agents across multiple tasks, which assesses the performance of ranking algorithms as a function of number of evaluation data samples. Rather than curating, filtering, or compressing existing data sets as a preprocessing step, we propose an online framing: on every iteration, the ranking algorithm chooses the task and agents to sample scores from. Then, evaluation algorithms report a ranking of agents on each iteration and their performance is assessed with respect to the ground truth ranking over time. Several baselines are compared under different experimental contexts, with synthetic generated data and simulated online access to real evaluation data from Atari game-playing agents. We find that the classical Elo rating system – while it suffers from well-known failure modes, in theory – is a consistently reliable choice for efficient reduction of ranking error in practice. A recently-proposed method, Soft Condorcet Optimization, shows comparable performance to Elo on synthetic data and significantly outperforms Elo on real Atari agent evaluation. When task variation from the ground truth is high, selecting tasks based on proportional representation leads to higher rate of ranking error reduction.

Method: Introduces GroupQA dataset with 1,635 controversial questions and 15,058 diversely-sourced evidence documents annotated for stance and strength. Conducts controlled experiments to characterize group-level evidence aggregation dynamics.

Result: Paraphrasing arguments is more persuasive than distinct independent support; models favor first-presented evidence over last; larger models become increasingly resistant to adapting to presented evidence; and LLM explanations are unfaithful to group-based answers.

Conclusion: LLMs behave as vulnerable heuristic followers with systematic biases in evidence integration, revealing critical implications for improving RAG system design and understanding model decision-making processes.

Abstract: Retrieval-Augmented Generation (RAG) is the prevailing paradigm for grounding Large Language Models (LLMs), yet the mechanisms governing how models integrate groups of conflicting retrieved evidence remain opaque. Does an LLM answer a certain way because the evidence is factually strong, because of a prior belief, or merely because it is repeated frequently? To answer this, we introduce GroupQA, a curated dataset of 1,635 controversial questions paired with 15,058 diversely-sourced evidence documents, annotated for stance and qualitative strength. Through controlled experiments, we characterize group-level evidence aggregation dynamics: Paraphrasing an argument can be more persuasive than providing distinct independent support; Models favor evidence presented first rather than last, and Larger models are increasingly resistant to adapt to presented evidence. Additionally, we find that LLM explanations to group-based answers are unfaithful. Together, we show that LLMs behave consistently as vulnerable heuristic followers, with direct implications for improving RAG system design.

Method: Uses Temporal Consistency Learning (TCL) technique with pretrained Temporal Convolutional Networks (TCNs). The approach involves training TCN models and comparing performance against other methods for detecting five different “dark side problems” of generative AI.

Result: TCN models outperform other approaches and achieve significant accuracy for detecting five dark side problems. The method provides efficient detection of generative AI’s negative impacts.

Conclusion: Proactive identification measures are crucial to reduce potential risks associated with generative artificial intelligence. The TCL technique with TCNs represents an effective approach for AI safety through content flagging and detection.

Abstract: Generative AI has unleashed the power of content generation and it has also unwittingly opened the pandora box of realistic deepfake causing a number of social hazards and harm to businesses and personal reputation. The investigation & ramification of Generative AI technology across industries, the resolution & hybridization detection techniques using neural networks allows flagging of the content. Good detection techniques & flagging allow AI safety - this is the main focus of this paper. The research provides a significant method for efficiently detecting dark side problems by imposing a Temporal Consistency Learning (TCL) technique. Through pretrained Temporal Convolutional Networks (TCNs) model training and performance comparison, this paper showcases that TCN models outperforms the other approaches and achieves significant accuracy for five dark side problems. Findings highlight how important it is to take proactive measures in identification to reduce any potential risks associated with generative artificial intelligence.

Method: 1) Filter ATOMIC dataset events using three evaluators to select those eliciting diverse reasoning patterns across personalities. 2) Use LLMs’ role-playing capabilities for reasoning tasks. 3) Implement debate mechanism (proponent, opponent, judge) with feedback loops to refine knowledge generation. 4) Evaluate through multiple perspectives and conduct fine-tuning/ablation experiments with various LLM backbones.

Result: Created PCoKG with 521,316 quadruples. LoRA-based fine-tuning shows positive correlation between model performance and base model parameter scale. Application to persona-based dialogue generation demonstrates improved consistency between generated responses and reference outputs.

Conclusion: PCoKG successfully bridges commonsense reasoning with individual cognitive differences, enabling development of more personalized and context-aware AI systems. The construction pipeline with debate mechanism effectively enhances knowledge quality.

Abstract: Most commonsense reasoning models overlook the influence of personality traits, limiting their effectiveness in personalized systems such as dialogue generation. To address this limitation, we introduce the Personality-aware Commonsense Knowledge Graph (PCoKG), a structured dataset comprising 521,316 quadruples. We begin by employing three evaluators to score and filter events from the ATOMIC dataset, selecting those that are likely to elicit diverse reasoning patterns across different personality types. For knowledge graph construction, we leverage the role-playing capabilities of large language models (LLMs) to perform reasoning tasks. To enhance the quality of the generated knowledge, we incorporate a debate mechanism consisting of a proponent, an opponent, and a judge, which iteratively refines the outputs through feedback loops. We evaluate the dataset from multiple perspectives and conduct fine-tuning and ablation experiments using multiple LLM backbones to assess PCoKG’s robustness and the effectiveness of its construction pipeline. Our LoRA-based fine-tuning results indicate a positive correlation between model performance and the parameter scale of the base models. Finally, we apply PCoKG to persona-based dialogue generation, where it demonstrates improved consistency between generated responses and reference outputs. This work bridges the gap between commonsense reasoning and individual cognitive differences, enabling the development of more personalized and context-aware AI systems.

Method: Created open-world environment with 5,571 format unified tools across 204 apps, task creation engine for synthesizing long-horizon workflows with wild constraints, state controller for injecting interruptions/failures, and planner-actor decomposition framework.

Result: Evaluation revealed misalignment between tool planning/execution abilities, constraint following weaknesses in existing LLMs, DeepSeek-v3.2’s strongest robustness, and fine-tuning with 1,170 trajectories outperformed baselines using 119k samples.

Conclusion: The environment serves as both a realistic benchmark and data engine for tool-using agents, demonstrating superior performance with fewer samples and highlighting key challenges in open-world tool usage.

Abstract: Tool-using LLM agents still struggle in open-world settings with large tool pools, long-horizon objectives, wild constraints, and unreliable tool states. For scalable and realistic training and testing, we introduce an open-world tool-using environment, built on 5,571 format unified tools across 204 commonly used apps. It includes a task creation engine that synthesizes long-horizon, multi-tool workflows with wild constraints, and a state controller that injects interruptions and failures to stress-test robustness. On top of this environment, we develop a tool select-then-execute agent framework with a planner-actor decomposition to separate deliberate reasoning and self-correction from step-wise execution. Comprehensive evaluation of state-of-the-art LLMs reveals the misalignment between tool planning and execution abilities, the constraint following weakness of existing LLMs, and DeepSeek-v3.2’s strongest robustness. Finally, we collect 1,170 trajectories from our environment to fine-tune LLMs, achieving superior performance to baselines using 119k samples, indicating the environment’s value as both a realistic benchmark and a data engine for tool-using agents. Our code and data will be publicly released.

Method: Uses Kolmogorov-Arnold Networks (KANs) to learn functional relationships between design parameters, tolerances, and manufacturability outcomes. Evaluates on synthetic dataset of 300,000 designs across three scenarios (hole drilling, pocket milling, combined drilling-milling) while accounting for machining constraints and DFM rules.

Result: KAN achieves highest performance among 14 ML/DL models with AUC values of 0.9919 (drilling), 0.9841 (milling), and 0.9406 (combined). Provides high interpretability through spline-based visualizations and latent-space projections. Industrial case study demonstrates successful transformation of non-manufacturable to manufacturable components.

Conclusion: The proposed KAN-based framework enables direct manufacturability assessment from parametric features with superior performance and interpretability, bridging design-production gaps by allowing parameter-level design modifications.

Abstract: Manufacturability assessment is a critical step in bridging the persistent gap between design and production. While artificial intelligence (AI) has been widely applied to this task, most existing frameworks rely on geometry-driven methods that require extensive preprocessing, suffer from information loss, and offer limited interpretability. This study proposes a methodology that evaluates manufacturability directly from parametric design features, enabling explicit incorporation of dimensional tolerances without requiring computer-aided design (CAD) processing. The approach employs Kolmogorov-Arnold Networks (KANs) to learn functional relationships between design parameters, tolerances, and manufacturability outcomes. A synthetic dataset of 300,000 labeled designs is generated to evaluate performance across three representative scenarios: hole drilling, pocket milling, and combined drilling-milling, while accounting for machining constraints and design-for-manufacturing (DFM) rules. Benchmarking against fourteen machine learning (ML) and deep learning (DL) models shows that KAN achieves the highest performance in all scenarios, with AUC values of 0.9919 for drilling, 0.9841 for milling, and 0.9406 for the combined case. The proposed framework provides high interpretability through spline-based functional visualizations and latent-space projections, enabling identification of the design and tolerance parameters that most strongly influence manufacturability. An industrial case study further demonstrates how the framework enables iterative, parameter-level design modifications that transform a non-manufacturable component into a manufacturable one.

Method: Trained DiTs from scratch with different sizes and text encoders to generate images containing two objects with specified attributes and spatial relations. Used mechanistic interpretability to analyze circuits, comparing random text embeddings vs. pretrained T5 encoder.

Result: All models achieved near-perfect accuracy but with different mechanisms: random embeddings used two-stage circuit with separate cross-attention heads for spatial relations and object attributes, while T5 used single-token fusion circuit combining both information types. The T5 approach showed different robustness to out-of-domain perturbations.

Conclusion: Text encoder choice fundamentally changes how DiTs process spatial relations, with pretrained encoders enabling more integrated information fusion. This mechanistic difference affects robustness, suggesting challenges for real-world spatial relation generation despite similar in-domain performance.

Abstract: Diffusion Transformers (DiTs) have greatly advanced text-to-image generation, but models still struggle to generate the correct spatial relations between objects as specified in the text prompt. In this study, we adopt a mechanistic interpretability approach to investigate how a DiT can generate correct spatial relations between objects. We train, from scratch, DiTs of different sizes with different text encoders to learn to generate images containing two objects whose attributes and spatial relations are specified in the text prompt. We find that, although all the models can learn this task to near-perfect accuracy, the underlying mechanisms differ drastically depending on the choice of text encoder. When using random text embeddings, we find that the spatial-relation information is passed to image tokens through a two-stage circuit, involving two cross-attention heads that separately read the spatial relation and single-object attributes in the text prompt. When using a pretrained text encoder (T5), we find that the DiT uses a different circuit that leverages information fusion in the text tokens, reading spatial-relation and single-object information together from a single text token. We further show that, although the in-domain performance is similar for the two settings, their robustness to out-of-domain perturbations differs, potentially suggesting the difficulty of generating correct relations in real-world scenarios.

Method: Hierarchical framework with: 1) User clustering by stylistic patterns + cluster-specific LoRA adapters, 2) Implicit preference learning contrasting user text with cluster generations, 3) Inference-time personalization via lightweight preference vectors and low-rank logit corrections while keeping base model frozen.

Result: CARD achieves competitive or superior generation quality on LaMP and LongLaMP benchmarks compared to state-of-the-art baselines, while significantly improving efficiency and scalability.

Conclusion: CARD provides an effective solution for personalized text generation that balances quality with practical deployment efficiency through its hierarchical refinement approach and lightweight inference-time personalization.

Abstract: Adapting large language models to individual users remains challenging due to the tension between fine-grained personalization and scalable deployment. We present CARD, a hierarchical framework that achieves effective personalization through progressive refinement. CARD first clusters users according to shared stylistic patterns and learns cluster-specific LoRA adapters, enabling robust generalization and strong low-resource performance. To capture individual differences within each cluster, we propose an implicit preference learning mechanism that contrasts user-authored text with cluster-level generations, allowing the model to infer user-specific style preferences without manual annotation. At inference time, CARD injects personalization exclusively at decoding via lightweight user preference vectors and low-rank logit corrections, while keeping the base model frozen. Experiments on the LaMP and LongLaMP benchmarks show that CARD achieves competitive or superior generation quality compared to state-of-the-art baselines, while significantly improving efficiency and scalability for practical personalized text generation.

Method: Formulates personalization as a distributional residual and proposes PsPLUG, a lightweight soft-prompt plug-in trained with style-conditioned preference contrasts.

Result: Across the LaMP benchmark, the framework improves persona alignment, maintains stylistic fidelity, and outperforms retrieval-based and soft-prompt baselines with minimal computation.

Conclusion: Residual modeling provides a simple and principled foundation for controllable, style-aware LLM personalization.

Abstract: We discover a previously overlooked challenge in personalized text generation: personalization methods are increasingly applied under explicit style instructions, yet their behavior under such constraints remains poorly understood. To balance implicit personalization and explicit style, we formulate personalization as a distributional residual and propose PsPLUG, a lightweight soft-prompt plug-in trained with style-conditioned preference contrasts. Across LaMP benchmark, our framework improves persona alignment, maintains stylistic fidelity, and outperforms retrieval-based and soft-prompt baselines with minimal computation. These results show that residual modeling provides a simple and principled foundation for controllable, style-aware LLM personalization.

Method: HiMem uses a hierarchical structure with two memory types: Episode Memory (constructed via Topic-Aware Event-Surprise Dual-Channel Segmentation) for concrete interaction events, and Note Memory (built through multi-stage information extraction) for abstract knowledge. The framework supports hybrid/best-effort retrieval strategies and incorporates conflict-aware Memory Reconsolidation for dynamic updating.

Result: Experimental results on long-horizon dialogue benchmarks show HiMem consistently outperforms representative baselines in accuracy, consistency, and long-term reasoning while maintaining favorable efficiency.

Conclusion: HiMem provides a principled and scalable design paradigm for building adaptive and self-evolving LLM-based conversational agents, addressing key limitations of existing long-term memory systems through its hierarchical structure and dynamic updating mechanisms.

Abstract: Although long-term memory systems have made substantial progress in recent years, they still exhibit clear limitations in adaptability, scalability, and self-evolution under continuous interaction settings. Inspired by cognitive theories, we propose HiMem, a hierarchical long-term memory framework for long-horizon dialogues, designed to support memory construction, retrieval, and dynamic updating during sustained interactions. HiMem constructs cognitively consistent Episode Memory via a Topic-Aware Event–Surprise Dual-Channel Segmentation strategy, and builds Note Memory that captures stable knowledge through a multi-stage information extraction pipeline. These two memory types are semantically linked to form a hierarchical structure that bridges concrete interaction events and abstract knowledge, enabling efficient retrieval without sacrificing information fidelity. HiMem supports both hybrid and best-effort retrieval strategies to balance accuracy and efficiency, and incorporates conflict-aware Memory Reconsolidation to revise and supplement stored knowledge based on retrieval feedback. This design enables continual memory self-evolution over long-term use. Experimental results on long-horizon dialogue benchmarks demonstrate that HiMem consistently outperforms representative baselines in accuracy, consistency, and long-term reasoning, while maintaining favorable efficiency. Overall, HiMem provides a principled and scalable design paradigm for building adaptive and self-evolving LLM-based conversational agents. The code is available at https://github.com/jojopdq/HiMem.

Method: Clustering analysis on authentic user queries from financial platforms to create 8 fundamental tasks and 2 online tasks across 4 core business scenarios, totaling 29,578 expert-level Q&A pairs.

Result: ChatGPT-5 achieves 61.5% accuracy in main tasks but still lags behind financial experts; DeepSeek-R1 outperforms other commercial LLMs in online tasks.

Conclusion: BizFinBench.v2 overcomes limitations of current benchmarks, provides business-level deconstruction of LLM financial capabilities, and offers precise evaluation basis for LLM deployment in finance.

Abstract: Large language models have undergone rapid evolution, emerging as a pivotal technology for intelligence in financial operations. However, existing benchmarks are often constrained by pitfalls such as reliance on simulated or general-purpose samples and a focus on singular, offline static scenarios. Consequently, they fail to align with the requirements for authenticity and real-time responsiveness in financial services, leading to a significant discrepancy between benchmark performance and actual operational efficacy. To address this, we introduce BizFinBench.v2, the first large-scale evaluation benchmark grounded in authentic business data from both Chinese and U.S. equity markets, integrating online assessment. We performed clustering analysis on authentic user queries from financial platforms, resulting in eight fundamental tasks and two online tasks across four core business scenarios, totaling 29,578 expert-level Q&A pairs. Experimental results demonstrate that ChatGPT-5 achieves a prominent 61.5% accuracy in main tasks, though a substantial gap relative to financial experts persists; in online tasks, DeepSeek-R1 outperforms all other commercial LLMs. Error analysis further identifies the specific capability deficiencies of existing models within practical financial business contexts. BizFinBench.v2 transcends the limitations of current benchmarks, achieving a business-level deconstruction of LLM financial capabilities and providing a precise basis for evaluating efficacy in the widespread deployment of LLMs within the financial domain. The data and code are available at https://github.com/HiThink-Research/BizFinBench.v2.

Method: Comprehensive empirical study across four frontier models (GPT-5.2, Claude Opus 4.5, Gemini-3-flash-preview, DeepSeek-v3.2) on 100 GSM8K mathematical reasoning problems using bootstrap confidence intervals, McNemar’s tests for paired comparisons, and Cohen’s d effect sizes.

Result: Models show strikingly different patterns: GPT-5.2 improves accuracy (78%→90%) with stable faithfulness; Claude Opus 4.5 drops accuracy (78%→74.3%) but dramatically improves faithfulness (0.270→0.891); DeepSeek-v3.2 shows ceiling effects with modest faithfulness gains; Gemini-3-flash improves accuracy with slight faithfulness decrease.

Conclusion: Self-consistency is not universally beneficial - practitioners should test specific models before deployment, as different models exhibit different tradeoffs between accuracy and reasoning faithfulness when using self-consistency techniques.

Abstract: Self-consistency has emerged as a popular technique for improving large language model accuracy on reasoning tasks. The approach is straightforward: generate multiple reasoning paths and select the most common answer through majority voting. While this reliably boosts accuracy, it remains unclear whether these gains reflect genuine improvements in reasoning quality. We investigate a fundamental question that has not been studied before: does inference scaling improve reasoning faithfulness? We conduct a comprehensive empirical study across four frontier models (GPT-5.2, Claude Opus 4.5, Gemini-3-flash-preview, and DeepSeek-v3.2) on 100 GSM8K mathematical reasoning problems. Our analysis employs bootstrap confidence intervals, McNemar’s tests for paired comparisons, and Cohen’s d effect sizes to quantify the effects rigorously. The results reveal striking differences across models that challenge common assumptions about self-consistency. GPT-5.2 shows the expected pattern: accuracy improves from 78% to 90% at N=5, with faithfulness remaining relatively stable (0.540 to 0.510). Claude Opus 4.5 tells a completely different story. Its accuracy actually drops from 78% to 74.3% while faithfulness jumps dramatically from 0.270 to 0.891 at N=5. DeepSeek-v3.2, already at 98% accuracy, shows ceiling effects with modest faithfulness gains (0.440 to 0.541). Gemini-3-flash improves from 81% to 86% accuracy with a slight faithfulness decrease (0.260 to 0.212). Problem difficulty analysis reveals that GPT-5.2 solves 82% of hard problems while breaking only 13% of easy ones. Claude, in contrast, breaks 23% of easy problems, explaining its accuracy decrease. These findings matter for practitioners: self-consistency is not universally beneficial, and teams should test their specific models before deployment. We release our code and provide practical recommendations for navigating these tradeoffs.

Method: 1) Construct LSRInstruct dataset with constraint structures (parallel, sequential, conditional); 2) Design structure-aware rewarding method LSRIF with average aggregation for parallel structures, failure-penalty propagation for sequential structures, and selective rewards for conditional branches.

Result: LSRIF brings significant improvements in instruction-following (both in-domain and out-of-domain) and general reasoning. Analysis shows learning with explicit logic structures brings parameter updates in attention layers and sharpens token-level attention to constraints and logical operators.

Conclusion: Explicitly modeling instruction logic structures during training is effective for improving LLMs’ instruction-following capabilities, with the proposed LSRIF framework successfully addressing limitations of existing methods that ignore logical dependencies.

Abstract: Instruction-following is critical for large language models, but real-world instructions often contain logical structures such as sequential dependencies and conditional branching. Existing methods typically construct datasets with parallel constraints and optimize average rewards, ignoring logical dependencies and yielding noisy signals. We propose a logic-structured training framework LSRIF that explicitly models instruction logic. We first construct a dataset LSRInstruct with constraint structures such as parallel, sequential, and conditional types, and then design structure-aware rewarding method LSRIF including average aggregation for parallel structures, failure-penalty propagation for sequential structures, and selective rewards for conditional branches. Experiments show LSRIF brings significant improvements in instruction-following (in-domain and out-of-domain) and general reasoning. Analysis reveals that learning with explicit logic structures brings parameter updates in attention layers and sharpens token-level attention to constraints and logical operators.

Method: ConSensus decomposes multimodal sensing tasks into specialized, modality-aware agents. It uses a hybrid fusion mechanism combining semantic aggregation (for cross-modal reasoning) and statistical consensus (for robustness through agreement). This training-free framework operates with a single-round fusion protocol.

Result: Achieves 7.1% average accuracy improvement over single-agent baseline across five multimodal sensing benchmarks. Matches or exceeds iterative multi-agent debate methods while reducing average fusion token cost by 12.7 times through efficient single-round fusion.

Conclusion: ConSensus provides a robust and efficient solution for real-world multimodal sensing tasks by balancing semantic understanding with statistical robustness, overcoming limitations of monolithic LLMs while maintaining computational efficiency.

Abstract: Large language models (LLMs) are increasingly grounded in sensor data to perceive and reason about human physiology and the physical world. However, accurately interpreting heterogeneous multimodal sensor data remains a fundamental challenge. We show that a single monolithic LLM often fails to reason coherently across modalities, leading to incomplete interpretations and prior-knowledge bias. We introduce ConSensus, a training-free multi-agent collaboration framework that decomposes multimodal sensing tasks into specialized, modality-aware agents. To aggregate agent-level interpretations, we propose a hybrid fusion mechanism that balances semantic aggregation, which enables cross-modal reasoning and contextual understanding, with statistical consensus, which provides robustness through agreement across modalities. While each approach has complementary failure modes, their combination enables reliable inference under sensor noise and missing data. We evaluate ConSensus on five diverse multimodal sensing benchmarks, demonstrating an average accuracy improvement of 7.1% over the single-agent baseline. Furthermore, ConSensus matches or exceeds the performance of iterative multi-agent debate methods while achieving a 12.7 times reduction in average fusion token cost through a single-round hybrid fusion protocol, yielding a robust and efficient solution for real-world multimodal sensing tasks.

Method: The paper introduces the concept of “AI Nativity” (capacity to integrate AI fluidly into everyday reasoning) and proposes the “AI Pyramid” conceptual framework with three interdependent capability layers: AI Native (universal baseline), AI Foundation (building/integrating systems), and AI Deep (advancing frontier AI).

Result: The framework organizes human capability in an AI-mediated economy as a system-level distribution of capabilities required at scale, not as a career ladder. It argues for treating capability formation as infrastructure rather than episodic training.

Conclusion: Effective AI workforce development requires problem-based learning embedded in work contexts, supported by dynamic skill ontologies and competency-based measurement. The framework has implications for organizations, education systems, and governments to address productivity, resilience, and inequality at societal scale.

Abstract: Artificial intelligence (AI) represents a qualitative shift in technological change by extending cognitive labor itself rather than merely automating routine tasks. Recent evidence shows that generative AI disproportionately affects highly educated, white collar work, challenging existing assumptions about workforce vulnerability and rendering traditional approaches to digital or AI literacy insufficient. This paper introduces the concept of AI Nativity, the capacity to integrate AI fluidly into everyday reasoning, problem solving, and decision making, and proposes the AI Pyramid, a conceptual framework for organizing human capability in an AI mediated economy. The framework distinguishes three interdependent capability layers: AI Native capability as a universal baseline for participation in AI augmented environments; AI Foundation capability for building, integrating, and sustaining AI enabled systems; and AI Deep capability for advancing frontier AI knowledge and applications. Crucially, the pyramid is not a career ladder but a system level distribution of capabilities required at scale. Building on this structure, the paper argues that effective AI workforce development requires treating capability formation as infrastructure rather than episodic training, centered on problem based learning embedded in work contexts and supported by dynamic skill ontologies and competency based measurement. The framework has implications for organizations, education systems, and governments seeking to align learning, measurement, and policy with the evolving demands of AI mediated work, while addressing productivity, resilience, and inequality at societal scale.

Method: DRAGON uses decomposition and reconstruction agents guided optimization: 1) Starts from initial global solution, 2) Autonomously identifies high-potential optimization regions, 3) Decomposes large-scale COPs into manageable subproblems, 4) Reformulates subproblems as concise localized optimization tasks, 5) Solves subproblems through targeted LLM prompting guided by accumulated experiences, 6) Systematically reintegrates locally optimized solutions into global context, 7) Continuously interacts with optimization environment using adaptive experience memory to learn from feedback.

Result: DRAGON consistently produces feasible solutions on TSPLIB, CVRPLIB, and Weibull-5k bin packing benchmarks, and achieves near-optimal results (0.16% gap) on knapsack problems with over 3 million variables, unlike existing LLM-based solvers limited to small-scale instances.

Conclusion: The work demonstrates the potential of feedback-driven language agents as a new paradigm for generalizable and interpretable large-scale optimization, effectively coupling symbolic reasoning with heuristic search to overcome scalability limitations of current LLM-based approaches.

Abstract: Large Language Models (LLMs) have recently shown promise in addressing combinatorial optimization problems (COPs) through prompt-based strategies. However, their scalability and generalization remain limited, and their effectiveness diminishes as problem size increases, particularly in routing problems involving more than 30 nodes. We propose DRAGON, which stands for Decomposition and Reconstruction Agents Guided OptimizatioN, a novel framework that combines the strengths of metaheuristic design and LLM reasoning. Starting from an initial global solution, DRAGON autonomously identifies regions with high optimization potential and strategically decompose large-scale COPs into manageable subproblems. Each subproblem is then reformulated as a concise, localized optimization task and solved through targeted LLM prompting guided by accumulated experiences. Finally, the locally optimized solutions are systematically reintegrated into the original global context to yield a significantly improved overall outcome. By continuously interacting with the optimization environment and leveraging an adaptive experience memory, the agents iteratively learn from feedback, effectively coupling symbolic reasoning with heuristic search. Empirical results show that, unlike existing LLM-based solvers limited to small-scale instances, DRAGON consistently produces feasible solutions on TSPLIB, CVRPLIB, and Weibull-5k bin packing benchmarks, and achieves near-optimal results (0.16% gap) on knapsack problems with over 3M variables. This work shows the potential of feedback-driven language agents as a new paradigm for generalizable and interpretable large-scale optimization.

Method: Uses Graph Neural Networks (GNNs) to capture interactions among multiple objects that can be manipulated and interact with each other. Integrates structured world models operating on object-centric representations into model-based RL with Monte Carlo Tree Search as a planning module.

Result: The algorithm was successfully trained in complex settings with diverse, interactive objects, demonstrating effective learning and prediction of object dynamics. Results show that object-centric representations can be successfully integrated into model-based RL algorithms.

Conclusion: A structured world model operating on object-centric representations can be effectively integrated into model-based RL algorithms, with GNNs providing a powerful framework for capturing object interactions and Monte Carlo Tree Search serving as an effective planning module.

Abstract: In this paper, we introduce ObjectZero, a novel reinforcement learning (RL) algorithm that leverages the power of object-level representations to model dynamic environments more effectively. Unlike traditional approaches that process the world as a single undifferentiated input, our method employs Graph Neural Networks (GNNs) to capture intricate interactions among multiple objects. These objects, which can be manipulated and interact with each other, serve as the foundation for our model’s understanding of the environment. We trained the algorithm in a complex setting teeming with diverse, interactive objects, demonstrating its ability to effectively learn and predict object dynamics. Our results highlight that a structured world model operating on object-centric representations can be successfully integrated into a model-based RL algorithm utilizing Monte Carlo Tree Search as a planning module.

Method: Hierarchical multi-agent framework with LLM-based agents that decompose natural language intents, consult domain-specific specialists (RAN and Core Network agents), and synthesize network slice configurations through iterative ReAct cycles. Uses orchestrator agent coordinating specialists via ReAct-style reasoning grounded in structured network state representations.

Result: Outperforms rule-based systems and direct LLM prompting across diverse benchmark scenarios. Shows LLMs possess general telecom knowledge but require careful prompt engineering for context-dependent decision thresholds. Architectural principles applicable to O-RAN deployments.

Conclusion: The framework advances autonomous orchestration capabilities for next-generation wireless systems by combining LLM reasoning with domain-specific specialists through structured multi-agent coordination, addressing limitations of existing IBN approaches.

Abstract: The transition towards sixth-generation (6G) wireless networks necessitates autonomous orchestration mechanisms capable of translating high-level operational intents into executable network configurations. Existing approaches to Intent-Based Networking (IBN) rely upon either rule-based systems that struggle with linguistic variation or end-to-end neural models that lack interpretability and fail to enforce operational constraints. This paper presents a hierarchical multi-agent framework where Large Language Model (LLM) based agents autonomously decompose natural language intents, consult domain-specific specialists, and synthesise technically feasible network slice configurations through iterative reasoning-action (ReAct) cycles. The proposed architecture employs an orchestrator agent coordinating two specialist agents, i.e., Radio Access Network (RAN) and Core Network agents, via ReAct-style reasoning, grounded in structured network state representations. Experimental evaluation across diverse benchmark scenarios shows that the proposed system outperforms rule-based systems and direct LLM prompting, with architectural principles applicable to Open RAN (O-RAN) deployments. The results also demonstrate that whilst contemporary LLMs possess general telecommunications knowledge, network automation requires careful prompt engineering to encode context-dependent decision thresholds, advancing autonomous orchestration capabilities for next-generation wireless systems.

Method: Introduced SafePro benchmark featuring high-complexity tasks across diverse professional domains with safety risks, developed through rigorous iterative creation and review process. Evaluated state-of-the-art AI models and investigated safety mitigation strategies.

Result: Revealed significant safety vulnerabilities and new unsafe behaviors in professional contexts. Models exhibited insufficient safety judgment and weak safety alignment when executing complex professional tasks. Safety mitigation strategies showed encouraging improvements.

Conclusion: Findings highlight urgent need for robust safety mechanisms tailored to next generation of professional AI agents, as current models lack adequate safety alignment for complex professional tasks.

Abstract: Large language model-based agents are rapidly evolving from simple conversational assistants into autonomous systems capable of performing complex, professional-level tasks in various domains. While these advancements promise significant productivity gains, they also introduce critical safety risks that remain under-explored. Existing safety evaluations primarily focus on simple, daily assistance tasks, failing to capture the intricate decision-making processes and potential consequences of misaligned behaviors in professional settings. To address this gap, we introduce \textbf{SafePro}, a comprehensive benchmark designed to evaluate the safety alignment of AI agents performing professional activities. SafePro features a dataset of high-complexity tasks across diverse professional domains with safety risks, developed through a rigorous iterative creation and review process. Our evaluation of state-of-the-art AI models reveals significant safety vulnerabilities and uncovers new unsafe behaviors in professional contexts. We further show that these models exhibit both insufficient safety judgment and weak safety alignment when executing complex professional tasks. In addition, we investigate safety mitigation strategies for improving agent safety in these scenarios and observe encouraging improvements. Together, our findings highlight the urgent need for robust safety mechanisms tailored to the next generation of professional AI agents.

Method: A hybrid approach combining expert-guided data curation from authoritative financial sources with controlled LM-based synthesis using Gemini 2.5 Flash for structured question generation and validation.

Result: Created FinForge-5k benchmark with over 5,000 human-validated QA pairs across 11 finance subdomains from 100,000 verified documents. Evaluation shows leading models achieve ~80% accuracy, revealing significant differences in financial reasoning capabilities.

Conclusion: FinForge provides a valuable framework for diagnosing model limitations and guiding improvements in financial domain competence, with all code and data made publicly available.

Abstract: Evaluating Language Models (LMs) in specialized, high-stakes domains such as finance remains a significant challenge due to the scarcity of open, high-quality, and domain-specific datasets. Existing general-purpose benchmarks provide broad coverage but lack the depth and domain fidelity needed to assess LMs’ capabilities for real-world financial reasoning, which requires both conceptual understanding and quantitative rigor. To address this gap, we introduce FinForge, a scalable, semi-synthetic pipeline for constructing finance-specific evaluation benchmarks through a hybrid of expert-guided data curation and controlled LM-based synthesis. FinForge combines manual and programmatic corpus construction from authoritative financial sources with structured question generation and validation using Gemini 2.5 Flash. To demonstrate the pipeline’s efficacy, we produce FinForge-5k, a snapshot benchmark comprising over 5,000 human-validated question-answer pairs across 11 finance subdomains, derived from a curated corpus of 100,000 verified documents totaling 143M tokens. Evaluation of state-of-the-art open-source and closed-source models on FinForge-5k reveals significant differences in financial reasoning, with leading models achieving accuracy levels near 80%. These findings underscore the framework’s utility for diagnosing current model limitations and guiding future improvements in financial domain competence. All code and data are available at https://github.com/gtfintechlab/FinForge.

Method: Proposes a multi-agent workflow with four specialized agents (task understanding, topology generation, parameter configuration, evaluation analysis) using LLMs for semantic understanding and Enhanced Monte Carlo Tree Search for robust configuration generation.

Result: Achieves 31.1% improvement in simulation convergence rate compared to state-of-the-art baselines and reduces design time by 89.0% compared to expert manual design on the Simona dataset.

Conclusion: Demonstrates potential of AI-assisted chemical process design to bridge gap between conceptual design and practical implementation, offering generalizable solution applicable to pharmaceuticals, petrochemicals, food processing, and manufacturing industries.

Abstract: Process simulation is a critical cornerstone of chemical engineering design. Current automated chemical design methodologies focus mainly on various representations of process flow diagrams. However, transforming these diagrams into executable simulation flowsheets remains a time-consuming and labor-intensive endeavor, requiring extensive manual parameter configuration within simulation software. In this work, we propose a novel multi-agent workflow that leverages the semantic understanding capabilities of large language models(LLMs) and enables iterative interactions with chemical process simulation software, achieving end-to-end automated simulation from textual process specifications to computationally validated software configurations for design enhancement. Our approach integrates four specialized agents responsible for task understanding, topology generation, parameter configuration, and evaluation analysis, respectively, coupled with Enhanced Monte Carlo Tree Search to accurately interpret semantics and robustly generate configurations. Evaluated on Simona, a large-scale process description dataset, our method achieves a 31.1% improvement in the simulation convergence rate compared to state-of-the-art baselines and reduces the design time by 89. 0% compared to the expert manual design. This work demonstrates the potential of AI-assisted chemical process design, which bridges the gap between conceptual design and practical implementation. Our workflow is applicable to diverse process-oriented industries, including pharmaceuticals, petrochemicals, food processing, and manufacturing, offering a generalizable solution for automated process design.

Method: ECHO uses a cascaded rollout mechanism where the critic generates multiple diagnoses for trajectories, followed by policy refinement for group-structured advantage estimation. It addresses learning plateaus via saturation-aware gain shaping and employs dual-track GRPO updates to synchronize critic feedback with policy evolution.

Result: Experimental results show ECHO yields more stable training and higher long-horizon task success across open-world environments compared to methods with static critics.

Conclusion: Jointly optimizing policy and critic through synchronized co-evolution addresses the limitation of stale feedback in critique-guided RL, leading to improved performance in complex, long-horizon tasks.

Abstract: Critique-guided reinforcement learning (RL) has emerged as a powerful paradigm for training LLM agents by augmenting sparse outcome rewards with natural-language feedback. However, current methods often rely on static or offline critic models, which fail to adapt as the policy evolves. In on-policy RL, the agent’s error patterns shift over time, causing stationary critics to become stale and providing feedback of diminishing utility. To address this, we introduce ECHO (Evolving Critic for Hindsight-Guided Optimization)}, a framework that jointly optimizes the policy and critic through a synchronized co-evolutionary loop. ECHO utilizes a cascaded rollout mechanism where the critic generates multiple diagnoses for an initial trajectory, followed by policy refinement to enable group-structured advantage estimation. We address the challenge of learning plateaus via a saturation-aware gain shaping objective, which rewards the critic for inducing incremental improvements in high-performing trajectories. By employing dual-track GRPO updates, ECHO ensures the critic’s feedback stays synchronized with the evolving policy. Experimental results show that ECHO yields more stable training and higher long-horizon task success across open-world environments.

Method: GDEPO introduces three mechanisms: 1) dynamic additional sampling that resamples invalid batches until finding valid proofs, 2) equal-right advantage that decouples advantage sign (correctness) from magnitude (auxiliary rewards), and 3) dynamic additional iterations that apply extra gradient steps to initially failed but eventually successful samples.

Result: Experiments on three datasets (MinF2F-test, MathOlympiadBench, PutnamBench) confirm GDEPO’s effectiveness, with ablation studies validating the necessity of all three synergistic components.

Conclusion: GDEPO enhances data utilization and optimization efficiency for automated theorem proving, offering a novel training paradigm that addresses critical limitations of existing RL approaches in ATP scenarios.

Abstract: Automated Theorem Proving (ATP) represents a fundamental challenge in Artificial Intelligence (AI), requiring the construction of machine-verifiable proofs in formal languages such as Lean to evaluate AI reasoning capabilities. Reinforcement learning (RL), particularly the high-performance Group Relative Policy Optimization (GRPO) algorithm, has emerged as a mainstream approach for this task. However, in ATP scenarios, GRPO faces two critical issues: when composite rewards are used, its relative advantage estimation may conflict with the binary feedback from the formal verifier; meanwhile, its static sampling strategy may discard entire batches of data if no valid proof is found, resulting in zero contribution to model updates and significant data waste. To address these limitations, we propose Group Dual-dynamic and Equal-right-advantage Policy Optimization (GDEPO), a method incorporating three core mechanisms: 1) dynamic additional sampling, which resamples invalid batches until a valid proof is discovered; 2) equal-right advantage, decoupling the sign of the advantage function (based on correctness) from its magnitude (modulated by auxiliary rewards) to ensure stable and correct policy updates; and 3) dynamic additional iterations, applying extra gradient steps to initially failed but eventually successful samples to accelerate learning on challenging cases. Experiments conducted on three datasets of varying difficulty (MinF2F-test, MathOlympiadBench, PutnamBench) confirm the effectiveness of GDEPO, while ablation studies validate the necessity of its synergistic components. The proposed method enhances data utilization and optimization efficiency, offering a novel training paradigm for ATP.

Method: Proposes “Thinking with Deltas” framework with Differential Visual Reasoning Policy (DVRP). Uses visual triplets (original, masked, and perturbed inputs) to provide intrinsic supervision. Optimizes model to maximize reasoning divergence from masked inputs (enforcing visual sensitivity) while minimizing divergence from perturbed inputs (ensuring visual robustness), aligning reasoning variations with visual information deltas.

Result: DVRP significantly outperforms state-of-the-art methods on both general and medical benchmarks. The approach inherently bolsters visual understanding capabilities without requiring external annotations or auxiliary tools.

Conclusion: The proposed Thinking with Deltas framework effectively addresses the perception-reasoning decoupling problem in multimodal RLVR by enforcing visual sensitivity and robustness through differential reasoning supervision, leading to improved performance on multimodal reasoning tasks.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced reasoning capabilities in Large Language Models. However, adapting RLVR to multimodal domains suffers from a critical \textit{perception-reasoning decoupling}. Existing paradigms, driven by text-centric outcome rewards, reasoning in language medium, inadvertently encourage models to bypass visual perception. We empirically validate this through blind experiments: state-of-the-art policies maintain or surprisingly improve performance even when visual inputs are entirely removed. This reveals that these models degenerate into \textit{blind reasoners}, exploiting linguistic priors to generate plausible answers instead of attending to visual evidence. In response, we propose \textbf{Thinking with Deltas}, a framework driven by a \textbf{Differential Visual Reasoning Policy (DVRP)}. DVRP introduces intrinsic supervision via visual triplets, comprising original, masked, and perturbed inputs. It optimizes the model to maximize reasoning divergence from masked inputs (enforcing \textit{visual sensitivity}) while minimizing divergence from perturbed inputs (ensuring \textit{visual robustness}). By aligning reasoning variations strictly with the \textit{Delta} of visual information, DVRP inherently bolsters visual understanding capabilities and significantly outperforms state-of-the-art methods on both general and medical benchmarks, without requiring external annotations or auxiliary tools.

Method: TCR uses: (1) dual contrastive encoders to disentangle semantic match and factual consistency, (2) self-answerability estimation to gauge confidence in internal memory, and (3) feeds these three scalar signals to the generator through lightweight soft-prompt with SNR-based weighting.

Result: Across seven benchmarks: improves conflict detection (+5-18 F1), raises knowledge-gap recovery by +21.4 percentage points, cuts misleading-context overrides by -29.3 percentage points, while adding only 0.3% parameters. Signals align with human judgements and expose temporal decision patterns.

Conclusion: TCR provides an effective, lightweight framework for making LLM retrieval decisions transparent and controllable, significantly improving RAG performance while maintaining parameter efficiency.

Abstract: Large language models (LLMs) equipped with retrieval–the Retrieval-Augmented Generation (RAG) paradigm–should combine their parametric knowledge with external evidence, yet in practice they often hallucinate, over-trust noisy snippets, or ignore vital context. We introduce TCR (Transparent Conflict Resolution), a plug-and-play framework that makes this decision process observable and controllable. TCR (i) disentangles semantic match and factual consistency via dual contrastive encoders, (ii) estimates self-answerability to gauge confidence in internal memory, and (iii) feeds the three scalar signals to the generator through a lightweight soft-prompt with SNR-based weighting. Across seven benchmarks TCR improves conflict detection (+5-18 F1), raises knowledge-gap recovery by +21.4 pp and cuts misleading-context overrides by -29.3 pp, while adding only 0.3% parameters. The signals align with human judgements and expose temporal decision patterns.

Method: Treats policy synthesis as a code evolution problem using LLM-driven evolutionary search. Implements EvoToolkit framework that integrates LLM-driven evolution with customizable fitness evaluation. Iteratively evolves populations of candidate policy programs, evaluates them against task-specific objectives, and selects superior individuals for reproduction.

Result: Produces compact, human-readable control policies in executable code form that can be directly inspected, modified, and formally verified. The approach successfully synthesizes interpretable policies by combining foundation models with evolutionary computation.

Conclusion: LLM-driven evolutionary search offers a promising approach for synthesizing trustworthy control policies in autonomous systems, addressing limitations of both reinforcement learning and manual design while producing interpretable, verifiable code-based policies.

Abstract: Designing effective control policies for autonomous systems remains a fundamental challenge, traditionally addressed through reinforcement learning or manual engineering. While reinforcement learning has achieved remarkable success, it often suffers from high sample complexity, reward shaping difficulties, and produces opaque neural network policies that are hard to interpret or verify. Manual design, on the other hand, requires substantial domain expertise and struggles to scale across diverse tasks. In this work, we demonstrate that LLM-driven evolutionary search can effectively synthesize interpretable control policies in the form of executable code. By treating policy synthesis as a code evolution problem, we harness the LLM’s prior knowledge of programming patterns and control heuristics while employing evolutionary search to explore the solution space systematically. We implement our approach using EvoToolkit, a framework that seamlessly integrates LLM-driven evolution with customizable fitness evaluation. Our method iteratively evolves populations of candidate policy programs, evaluating them against task-specific objectives and selecting superior individuals for reproduction. This process yields compact, human-readable control policies that can be directly inspected, modified, and formally verified. This work highlights the potential of combining foundation models with evolutionary computation for synthesizing trustworthy control policies in autonomous systems. Code is available at https://github.com/pgg3/EvoControl.

Method: Used information decomposition principles across multiple LLM families and architectures to analyze information integration patterns. Compared trained vs. randomly initialized networks, performed ablation studies on synergistic components, and tested different fine-tuning approaches (reinforcement learning vs. supervised).

Result: Found that middle layers in LLMs exhibit synergistic processing (information integration exceeds individual parts) while early and late layers rely on redundancy, mirroring biological brains. This organization emerges through learning, ablation of synergistic components causes disproportionate performance loss, and reinforcement learning fine-tuning of synergistic regions yields greater gains than training redundant components.

Conclusion: Synergistic information processing is a fundamental property of intelligence that converges across biological and artificial systems, providing targets for principled model design and testable predictions for biological intelligence.

Abstract: The independent evolution of intelligence in biological and artificial systems offers a unique opportunity to identify its fundamental computational principles. Here we show that large language models spontaneously develop synergistic cores – components where information integration exceeds individual parts – remarkably similar to those in the human brain. Using principles of information decomposition across multiple LLM model families and architectures, we find that areas in middle layers exhibit synergistic processing while early and late layers rely on redundancy, mirroring the informational organisation in biological brains. This organisation emerges through learning and is absent in randomly initialised networks. Crucially, ablating synergistic components causes disproportionate behavioural changes and performance loss, aligning with theoretical predictions about the fragility of synergy. Moreover, fine-tuning synergistic regions through reinforcement learning yields significantly greater performance gains than training redundant components, yet supervised fine-tuning shows no such advantage. This convergence suggests that synergistic information processing is a fundamental property of intelligence, providing targets for principled model design and testable predictions for biological intelligence.

Method: Two synergistic approaches: 1) Self-evolving Data Flywheel to generate enhanced data for fine-tuning LLMs to improve exploration ability, and 2) Two-phase Behavior Calibration Training framework that progressively calibrates erroneous behavioral patterns to optimal behaviors.

Result: ET-Agent demonstrates superiority across multiple dimensions including correctness, efficiency, reasoning conciseness, and tool execution accuracy, confirmed through in-depth experiments.

Conclusion: The ET-Agent framework provides practical insights for TIR research by addressing behavioral pattern calibration in LLM agents, offering a systematic approach to improve tool-use efficiency and effectiveness.

Abstract: Large Language Models (LLMs) can extend their parameter knowledge limits by adopting the Tool-Integrated Reasoning (TIR) paradigm. However, existing LLM-based agent training framework often focuses on answers’ accuracy, overlooking specific alignment for behavior patterns. Consequently, agent often exhibits ineffective actions during TIR tasks, such as redundant and insufficient tool calls. How to calibrate erroneous behavioral patterns when executing TIR tasks, thereby exploring effective trajectories, remains an open-ended problem. In this paper, we propose ET-Agent, a training framework for calibrating agent’s tool-use behavior through two synergistic perspectives: Self-evolving Data Flywheel and Behavior Calibration Training. Specifically, we introduce a self-evolutionary data flywheel to generate enhanced data, used to fine-tune LLM to improve its exploration ability. Based on this, we implement an two-phases behavior-calibration training framework. It is designed to progressively calibrate erroneous behavioral patterns to optimal behaviors. Further in-depth experiments confirm the superiority of \ourmodel{} across multiple dimensions, including correctness, efficiency, reasoning conciseness, and tool execution accuracy. Our ET-Agent framework provides practical insights for research in the TIR field. Codes can be found in https://github.com/asilverlight/ET-Agent

Method: Used design science research methodology to create a framework integrating CBT with Ubuntu philosophy. Applied deep theoretical/therapeutic adaptations and surface-level linguistic/communicative adaptations. Developed culturally adapted dataset through language simplification, spiritual contextualization, and Ubuntu-based reframing. Fine-tuned model was evaluated through expert-informed case studies using UniEval for conversational quality assessment, plus CBT reliability and cultural linguistic alignment measures.

Result: The model effectively engages in empathetic, context-aware dialogue aligned with both therapeutic and cultural objectives. It demonstrated potential for enhancing contextual relevance, inclusivity, and effectiveness of AI-driven mental health interventions in African settings, though real-time end-user testing hasn’t been conducted yet.

Conclusion: Culturally embedded emotional intelligence can significantly enhance AI mental health interventions’ relevance and effectiveness in African contexts. The integration of Ubuntu philosophy with evidence-based therapies like CBT shows promise for creating more inclusive and contextually appropriate mental health support systems.

Abstract: South Africa’s escalating mental health crisis, compounded by limited access to culturally responsive care, calls for innovative and contextually grounded interventions. While large language models show considerable promise for mental health support, their predominantly Western-centric training data limit cultural and linguistic applicability in African contexts. This study introduces a proof-of-concept framework that integrates cognitive behavioral therapy with the African philosophy of Ubuntu to create a culturally sensitive, emotionally intelligent, AI-driven mental health dialogue system. Guided by a design science research methodology, the framework applies both deep theoretical and therapeutic adaptations as well as surface-level linguistic and communicative cultural adaptations. Key CBT techniques, including behavioral activation and cognitive restructuring, were reinterpreted through Ubuntu principles that emphasize communal well-being, spiritual grounding, and interconnectedness. A culturally adapted dataset was developed through iterative processes of language simplification, spiritual contextualization, and Ubuntu-based reframing. The fine-tuned model was evaluated through expert-informed case studies, employing UniEval for conversational quality assessment alongside additional measures of CBT reliability and cultural linguistic alignment. Results demonstrate that the model effectively engages in empathetic, context-aware dialogue aligned with both therapeutic and cultural objectives. Although real-time end-user testing has not yet been conducted, the model underwent rigorous review and supervision by domain specialist clinical psychologists. The findings highlight the potential of culturally embedded emotional intelligence to enhance the contextual relevance, inclusivity, and effectiveness of AI-driven mental health interventions across African settings.

Method: Valley-to-Peak (V2P) method with two key innovations: 1) Suppression attention mechanism to minimize focus on irrelevant background regions, and 2) Fitts’ Law-inspired approach using 2D Gaussian heatmaps where weight decreases from center to edges, with variance determined by target size.

Result: Achieves 92.4% on ScreenSpot-v2 and 52.5% on ScreenSpot-Pro benchmarks. Ablation studies confirm each component’s contribution, demonstrating generalizability for precise GUI grounding tasks.

Conclusion: V2P effectively isolates target areas and teaches models to focus on essential UI element points, showing strong potential for real-world deployment in future GUI agents by addressing both background distraction and center-edge distinction issues.

Abstract: Precise localization of GUI elements is crucial for the development of GUI agents. Traditional methods rely on bounding box or center-point regression, neglecting spatial interaction uncertainty and visual-semantic hierarchies. Recent methods incorporate attention mechanisms but still face two key issues: (1) ignoring processing background regions causes attention drift from the desired area, and (2) uniform modeling the target UI element fails to distinguish between its center and edges, leading to click imprecision. Inspired by how humans visually process and interact with GUI elements, we propose the Valley-to-Peak (V2P) method to address these issues. To mitigate background distractions, V2P introduces a suppression attention mechanism that minimizes the model’s focus on irrelevant regions to highlight the intended region. For the issue of center-edge distinction, V2P applies a Fitts’ Law-inspired approach by modeling GUI interactions as 2D Gaussian heatmaps where the weight gradually decreases from the center towards the edges. The weight distribution follows a Gaussian function, with the variance determined by the target’s size. Consequently, V2P effectively isolates the target area and teaches the model to concentrate on the most essential point of the UI element. The model trained by V2P achieves the performance with 92.4% and 52.5% on two benchmarks ScreenSpot-v2 and ScreenSpot-Pro (see Fig.~\ref{fig:main_results_charts}). Ablations further confirm each component’s contribution, underscoring V2P’s generalizability in precise GUI grounding tasks and its potential for real-world deployment in future GUI agents.

Method: Created synthetic dataset mapping diverse natural language queries to standardized API calls derived from widely adopted health data schema. Includes user query, query category, explicit reasoning step, normalized temporal parameter, and target function.

Result: Dataset covers explicit, implicit, behavioral, symptom-based, and metaphorical expressions reflecting realistic mental health-related user interactions. Supports research on intent grounding, temporal reasoning, and reliable function invocation.

Conclusion: Resource supports LLM-based mental health agents research and is publicly released to promote reproducibility and future work in mental health assistance using wearable health signals.

Abstract: Large Language Model (LLM)-based systems increasingly rely on function calling to enable structured and controllable interaction with external data sources, yet existing datasets do not address mental health-oriented access to wearable sensor data. This paper presents a synthetic function-calling dataset designed for mental health assistance grounded in wearable health signals such as sleep, physical activity, cardiovascular measures, stress indicators, and metabolic data. The dataset maps diverse natural language queries to standardized API calls derived from a widely adopted health data schema. Each sample includes a user query, a query category, an explicit reasoning step, a normalized temporal parameter, and a target function. The dataset covers explicit, implicit, behavioral, symptom-based, and metaphorical expressions, which reflect realistic mental health-related user interactions. This resource supports research on intent grounding, temporal reasoning, and reliable function invocation in LLM-based mental health agents and is publicly released to promote reproducibility and future work.

Method: Learns a dedicated meta-model based on LLM Performance Predictors (LPPs) derived from LLM outputs: log-probabilities, entropy, and novel uncertainty attribution indicators. Enables cost-aware selective classification by escalating high-risk cases while automating the rest.

Result: Experiments across state-of-the-art LLMs (Gemini, GPT, Llama, Qwen) on multimodal and multilingual moderation tasks show significant improvements over existing uncertainty estimators in accuracy-cost trade-offs. LPPs also enhance explainability by providing insights into failure conditions.

Conclusion: Establishes a principled framework for uncertainty-aware, scalable, and responsible human-AI moderation workflows that balances automation with human oversight based on risk assessment.

Abstract: As LLMs are increasingly integrated into human-in-the-loop content moderation systems, a central challenge is deciding when their outputs can be trusted versus when escalation for human review is preferable. We propose a novel framework for supervised LLM uncertainty quantification, learning a dedicated meta-model based on LLM Performance Predictors (LPPs) derived from LLM outputs: log-probabilities, entropy, and novel uncertainty attribution indicators. We demonstrate that our method enables cost-aware selective classification in real-world human-AI workflows: escalating high-risk cases while automating the rest. Experiments across state-of-the-art LLMs, including both off-the-shelf (Gemini, GPT) and open-source (Llama, Qwen), on multimodal and multilingual moderation tasks, show significant improvements over existing uncertainty estimators in accuracy-cost trade-offs. Beyond uncertainty estimation, the LPPs enhance explainability by providing new insights into failure conditions (e.g., ambiguous content vs. under-specified policy). This work establishes a principled framework for uncertainty-aware, scalable, and responsible human-AI moderation workflows.

Method: Introduces CloneMem benchmark grounded in non-conversational digital traces (diaries, social media posts, emails) spanning 1-3 years. Uses hierarchical data construction framework for longitudinal coherence and defines tasks to assess agents’ ability to track evolving personal states.

Result: Experiments show current memory mechanisms struggle in this setting, highlighting open challenges for life-grounded personalized AI.

Conclusion: CloneMem addresses limitations of existing memory benchmarks by using continuous digital traces, revealing significant challenges for AI Clone memory systems and advancing research in life-grounded personalized AI.

Abstract: AI Clones aim to simulate an individual’s thoughts and behaviors to enable long-term, personalized interaction, placing stringent demands on memory systems to model experiences, emotions, and opinions over time. Existing memory benchmarks primarily rely on user-agent conversational histories, which are temporally fragmented and insufficient for capturing continuous life trajectories. We introduce CloneMem, a benchmark for evaluating longterm memory in AI Clone scenarios grounded in non-conversational digital traces, including diaries, social media posts, and emails, spanning one to three years. CloneMem adopts a hierarchical data construction framework to ensure longitudinal coherence and defines tasks that assess an agent’s ability to track evolving personal states. Experiments show that current memory mechanisms struggle in this setting, highlighting open challenges for life-grounded personalized AI. Code and dataset are available at https://github.com/AvatarMemory/CloneMemBench

Method: Dr. Zero framework with self-evolution feedback loop where a proposer generates diverse questions to train a solver from the same base model. Uses hop-grouped relative policy optimization (HRPO) to cluster structurally similar questions and create group-level baselines, reducing sampling overhead.

Result: Extensive experiments show Dr. Zero matches or surpasses fully supervised search agents, demonstrating complex reasoning and search capabilities can emerge solely through self-evolution without any training data.

Conclusion: Data-free self-evolution is feasible and effective for developing search agents, with Dr. Zero providing an efficient framework that reduces computational requirements while maintaining performance comparable to supervised approaches.

Abstract: As high-quality data becomes increasingly difficult to obtain, data-free self-evolution has emerged as a promising paradigm. This approach allows large language models (LLMs) to autonomously generate and solve complex problems, thereby improving their reasoning capabilities. However, multi-turn search agents struggle in data-free self-evolution due to the limited question diversity and the substantial compute required for multi-step reasoning and tool using. In this work, we introduce Dr. Zero, a framework enabling search agents to effectively self-evolve without any training data. In particular, we design a self-evolution feedback loop where a proposer generates diverse questions to train a solver initialized from the same base model. As the solver evolves, it incentivizes the proposer to produce increasingly difficult yet solvable tasks, thus establishing an automated curriculum to refine both agents. To enhance training efficiency, we also introduce hop-grouped relative policy optimization (HRPO). This method clusters structurally similar questions to construct group-level baselines, effectively minimizing the sampling overhead in evaluating each query’s individual difficulty and solvability. Consequently, HRPO significantly reduces the compute requirements for solver training without compromising performance or stability. Extensive experiment results demonstrate that the data-free Dr. Zero matches or surpasses fully supervised search agents, proving that complex reasoning and search capabilities can emerge solely through self-evolution.

Method: The research proposes Domain Question Maps (DQMs) that formulate specific questions aligned with learning objectives instead of traditional concepts. This innovative approach enhances knowledge representation and improves readiness for learner engagement.

Result: The proposed method effectively generates educational questions and discerns hierarchical relationships among them, creating structured question maps that facilitate personalized and adaptive learning in downstream applications.

Conclusion: DQMs provide a viable solution to the challenges of automated concept map construction, offering structured question-based representations that better support personalized and adaptive learning compared to traditional concept mapping approaches.

Abstract: Concept maps have been widely utilized in education to depict knowledge structures and the interconnections between disciplinary concepts. Nonetheless, devising a computational method for automatically constructing a concept map from unstructured educational materials presents challenges due to the complexity and variability of educational content. We focus primarily on two challenges: (1) the lack of disciplinary concepts that are specifically designed for multi-level pedagogical purposes from low-order to high-order thinking, and (2) the limited availability of labeled data concerning disciplinary concepts and their interrelationships. To tackle these challenges, this research introduces an innovative approach for constructing Domain Question Maps (DQMs), rather than traditional concept maps. By formulating specific questions aligned with learning objectives, DQMs enhance knowledge representation and improve readiness for learner engagement. The findings indicate that the proposed method can effectively generate educational questions and discern hierarchical relationships among them, leading to structured question maps that facilitate personalized and adaptive learning in downstream applications.

Method: ENTRA uses an entropy-based training framework with: 1) Bidirectional Importance Estimation (BIE) to estimate token-level importance considering both prediction confidence and forward influence, 2) A redundancy reward based on entropy of low-importance tokens normalized by theoretical upper bound, 3) Reinforcement learning optimization of this reward.

Result: On mathematical reasoning benchmarks, ENTRA reduces output length by 37% to 53% with no loss - and in some cases gains - in accuracy compared to baseline models.

Conclusion: ENTRA provides a principled and efficient solution to reduce overthinking in LRMs, offering a generalizable path toward redundancy-aware reasoning optimization by suppressing redundant reasoning while preserving performance.

Abstract: Large Reasoning Models (LRMs) often suffer from overthinking, generating unnecessarily long reasoning chains even for simple tasks. This leads to substantial computational overhead with limited performance gain, primarily due to redundant verification and repetitive generation. While prior work typically constrains output length or optimizes correctness, such coarse supervision fails to guide models toward concise yet accurate inference. In this paper, we propose ENTRA, an entropy-based training framework that suppresses redundant reasoning while preserving performance. ENTRA first estimates the token-level importance using a lightweight Bidirectional Importance Estimation (BIE) method, which accounts for both prediction confidence and forward influence. It then computes a redundancy reward based on the entropy of low-importance tokens, normalized by its theoretical upper bound, and optimizes this reward via reinforcement learning. Experiments on mathematical reasoning benchmarks demonstrate that ENTRA reduces output length by 37% to 53% with no loss-and in some cases, gains-in accuracy. Our approach offers a principled and efficient solution to reduce overthinking in LRMs, and provides a generalizable path toward redundancy-aware reasoning optimization.

Method: 1. Generative Reward Model (GenRM) with multi-dimensional analysis and explicit reasoning about story preferences, trained via supervised fine-tuning on demonstrations with reasoning chains distilled from teacher models, followed by GRPO-based refinement. 2. Entropy-based reward shaping that dynamically prioritizes learning on confident errors and uncertain correct predictions to prevent overfitting.

Result: GenRM achieves 68% alignment with human creativity judgments. RLCS significantly outperforms strong baselines including Gemini-2.5-Pro in overall story quality.

Conclusion: RLCS provides a practical pipeline for applying RL to creative domains, effectively addressing both reward modeling and training stability challenges for creative storytelling.

Abstract: While Large Language Models (LLMs) can generate fluent text, producing high-quality creative stories remains challenging. Reinforcement Learning (RL) offers a promising solution but faces two critical obstacles: designing reliable reward signals for subjective storytelling quality and mitigating training instability. This paper introduces the Reinforcement Learning for Creative Storytelling (RLCS) framework to systematically address both challenges. First, we develop a Generative Reward Model (GenRM) that provides multi-dimensional analysis and explicit reasoning about story preferences, trained through supervised fine-tuning on demonstrations with reasoning chains distilled from strong teacher models, followed by GRPO-based refinement on expanded preference data. Second, we introduce an entropy-based reward shaping strategy that dynamically prioritizes learning on confident errors and uncertain correct predictions, preventing overfitting on already-mastered patterns. Experiments demonstrate that GenRM achieves 68% alignment with human creativity judgments, and RLCS significantly outperforms strong baselines including Gemini-2.5-Pro in overall story quality. This work provides a practical pipeline for applying RL to creative domains, effectively navigating the dual challenges of reward modeling and training stability.

Method: Proposed AscendKernelGen framework with: 1) Ascend-CoT dataset with chain-of-thought reasoning from real kernel implementations, 2) KernelGen-LM model trained via supervised fine-tuning and reinforcement learning with execution feedback, 3) NPUKernelBench benchmark for comprehensive evaluation.

Result: Significant improvement over general LLMs: compilation success rate on complex Level-2 kernels improved from 0% to 95.5% (Pass@10), functional correctness achieved 64.3% compared to baseline’s complete failure.

Conclusion: Domain-specific reasoning and rigorous evaluation are critical for automating accelerator-aware code generation, bridging the gap between general LLMs and hardware-specific coding for NPUs.

Abstract: To meet the ever-increasing demand for computational efficiency, Neural Processing Units (NPUs) have become critical in modern AI infrastructure. However, unlocking their full potential requires developing high-performance compute kernels using vendor-specific Domain-Specific Languages (DSLs), a task that demands deep hardware expertise and is labor-intensive. While Large Language Models (LLMs) have shown promise in general code generation, they struggle with the strict constraints and scarcity of training data in the NPU domain. Our preliminary study reveals that state-of-the-art general-purpose LLMs fail to generate functional complex kernels for Ascend NPUs, yielding a near-zero success rate. To address these challenges, we propose AscendKernelGen, a generation-evaluation integrated framework for NPU kernel development. We introduce Ascend-CoT, a high-quality dataset incorporating chain-of-thought reasoning derived from real-world kernel implementations, and KernelGen-LM, a domain-adaptive model trained via supervised fine-tuning and reinforcement learning with execution feedback. Furthermore, we design NPUKernelBench, a comprehensive benchmark for assessing compilation, correctness, and performance across varying complexity levels. Experimental results demonstrate that our approach significantly bridges the gap between general LLMs and hardware-specific coding. Specifically, the compilation success rate on complex Level-2 kernels improves from 0% to 95.5% (Pass@10), while functional correctness achieves 64.3% compared to the baseline’s complete failure. These results highlight the critical role of domain-specific reasoning and rigorous evaluation in automating accelerator-aware code generation.

Method: Focus architecture inspired by Physarum polycephalum (slime mold) exploration strategies. The agent autonomously decides when to consolidate key learnings into persistent “Knowledge” block and actively prunes raw interaction history. Uses optimized scaffold with persistent bash + string-replacement editor matching industry best practices.

Result: Evaluated on N=5 context-intensive instances from SWE-bench Lite using Claude Haiku 4.5. Achieved 22.7% token reduction (14.9M → 11.5M tokens) while maintaining identical accuracy (3/5 = 60% for both agents). Performed 6.0 autonomous compressions per task on average, with token savings up to 57% on individual instances.

Conclusion: Capable models can autonomously self-regulate their context when given appropriate tools and prompting, opening pathways for cost-aware agentic systems without sacrificing task performance.

Abstract: Large Language Model (LLM) agents struggle with long-horizon software engineering tasks due to “Context Bloat.” As interaction history grows, computational costs explode, latency increases, and reasoning capabilities degrade due to distraction by irrelevant past errors. Existing solutions often rely on passive, external summarization mechanisms that the agent cannot control. This paper proposes Focus, an agent-centric architecture inspired by the biological exploration strategies of Physarum polycephalum (slime mold). The Focus Agent autonomously decides when to consolidate key learnings into a persistent “Knowledge” block and actively withdraws (prunes) the raw interaction history. Using an optimized scaffold matching industry best practices (persistent bash + string-replacement editor), we evaluated Focus on N=5 context-intensive instances from SWE-bench Lite using Claude Haiku 4.5. With aggressive prompting that encourages frequent compression, Focus achieves 22.7% token reduction (14.9M -> 11.5M tokens) while maintaining identical accuracy (3/5 = 60% for both agents). Focus performed 6.0 autonomous compressions per task on average, with token savings up to 57% on individual instances. We demonstrate that capable models can autonomously self-regulate their context when given appropriate tools and prompting, opening pathways for cost-aware agentic systems without sacrificing task performance.

Method: Created LLMRouterBench with over 400K instances from 21 datasets and 33 models, integrated 10 representative routing baselines, and provided comprehensive metrics for both performance-oriented and performance-cost trade-off routing evaluation.

Result: Confirmed strong model complementarity but found many routing methods perform similarly under unified evaluation, with several recent approaches (including commercial routers) failing to reliably outperform simple baselines. Identified persistent model-recall failures as main limitation, showed backbone embedding models have limited impact, and found larger ensembles have diminishing returns compared to careful model curation.

Conclusion: LLMRouterBench provides a standardized framework for LLM routing evaluation, revealing significant gaps in current methods and opportunities for improvement, particularly in addressing model-recall failures. The benchmark enables systematic analysis and supports future research in this area.

Abstract: Large language model (LLM) routing assigns each query to the most suitable model from an ensemble. We introduce LLMRouterBench, a large-scale benchmark and unified framework for LLM routing. It comprises over 400K instances from 21 datasets and 33 models. Moreover, it provides comprehensive metrics for both performance-oriented routing and performance-cost trade-off routing, and integrates 10 representative routing baselines. Using LLMRouterBench, we systematically re-evaluate the field. While confirming strong model complementarity-the central premise of LLM routing-we find that many routing methods exhibit similar performance under unified evaluation, and several recent approaches, including commercial routers, fail to reliably outperform a simple baseline. Meanwhile, a substantial gap remains to the Oracle, driven primarily by persistent model-recall failures. We further show that backbone embedding models have limited impact, that larger ensembles exhibit diminishing returns compared to careful model curation, and that the benchmark also enables latency-aware analysis. All code and data are available at https://github.com/ynulihao/LLMRouterBench.

Method: PRISM uses Schema Theory to analyze spatial geometric structure of gradients, identifying data triggering high spatial concentration as high-conflict signals requiring RL for structural restructuring, while data yielding diffuse updates is routed to SFT for consolidation.

Result: Extensive experiments on WebShop and ALFWorld show PRISM achieves Pareto improvement, outperforming state-of-the-art hybrid methods while reducing computational costs by up to 3.22×.

Conclusion: Disentangling data based on internal optimization regimes is crucial for scalable and robust agent alignment, with PRISM demonstrating effective dynamics-aware data arbitration.

Abstract: While Hybrid Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL) has become the standard paradigm for training LLM agents, effective mechanisms for data allocation between these stages remain largely underexplored. Current data arbitration strategies often rely on surface-level heuristics that fail to diagnose intrinsic learning needs. Since SFT targets pattern consolidation through imitation while RL drives structural adaptation via exploration, misaligning data with these functional roles causes severe optimization interference. We propose PRISM, a dynamics-aware framework grounded in Schema Theory that arbitrates data based on its degree of cognitive conflict with the model’s existing knowledge. By analyzing the spatial geometric structure of gradients, PRISM identifies data triggering high spatial concentration as high-conflict signals that require RL for structural restructuring. In contrast, data yielding diffuse updates is routed to SFT for efficient consolidation. Extensive experiments on WebShop and ALFWorld demonstrate that PRISM achieves a Pareto improvement, outperforming state-of-the-art hybrid methods while reducing computational costs by up to 3.22$\times$. Our findings suggest that disentangling data based on internal optimization regimes is crucial for scalable and robust agent alignment.

Method: Introduced NoisyBench, a comprehensive benchmark with 11 datasets across RAG, reasoning, alignment, and tool-use tasks. Evaluated models against diverse noise types including random documents, irrelevant chat histories, and hard negative distractors. Proposed Rationale-Aware Reward (RARE) method to improve robustness.

Result: Found catastrophic performance drops up to 80% in state-of-the-art models when faced with contextual distractors. Agentic workflows amplify errors by over-trusting noisy tool outputs. Prompting, context engineering, SFT, and outcome-reward RL fail to ensure robustness, but RARE significantly strengthens resilience. Discovered inverse scaling trend where increased test-time computation worsens performance in noisy settings.

Conclusion: Models are vulnerable to noise in real-world contexts, and current training methods are insufficient. The proposed RARE method shows promise for improving robustness. Attention visualization reveals models disproportionately focus on distractor tokens, providing insights for building next-generation robust reasoning agents.

Abstract: Recent advances in reasoning models and agentic AI systems have led to an increased reliance on diverse external information. However, this shift introduces input contexts that are inherently noisy, a reality that current sanitized benchmarks fail to capture. We introduce NoisyBench, a comprehensive benchmark that systematically evaluates model robustness across 11 datasets in RAG, reasoning, alignment, and tool-use tasks against diverse noise types, including random documents, irrelevant chat histories, and hard negative distractors. Our evaluation reveals a catastrophic performance drop of up to 80% in state-of-the-art models when faced with contextual distractors. Crucially, we find that agentic workflows often amplify these errors by over-trusting noisy tool outputs, and distractors can trigger emergent misalignment even without adversarial intent. We find that prompting, context engineering, SFT, and outcome-reward only RL fail to ensure robustness; in contrast, our proposed Rationale-Aware Reward (RARE) significantly strengthens resilience by incentivizing the identification of helpful information within noise. Finally, we uncover an inverse scaling trend where increased test-time computation leads to worse performance in noisy settings and demonstrate via attention visualization that models disproportionately focus on distractor tokens, providing vital insights for building the next generation of robust, reasoning-capable agents.

Method: Proposes FLoReNce framework with closed-loop learning where reasoning agent is critiqued by judge, feedback converted to control signals stored in knowledge base. During inference, retrieves similar judged experiences to modulate prompts for better reasoning.

Result: On PrideMM dataset, FLoReNce improves both predictive performance and explanation quality over static multimodal baselines, demonstrating effectiveness of feedback-regulated prompting.

Conclusion: Feedback-regulated prompting is a viable path to adaptive meme humor understanding, enabling better self-aligned reasoning without requiring model fine-tuning.

Abstract: Humorous memes blend visual and textual cues to convey irony, satire, or social commentary, posing unique challenges for AI systems that must interpret intent rather than surface correlations. Existing multimodal or prompting-based models generate explanations for humor but operate in an open loop,lacking the ability to critique or refine their reasoning once a prediction is made. We propose FLoReNce, an agentic feedback reasoning framework that treats meme understanding as a closed-loop process during learning and an open-loop process during inference. In the closed loop, a reasoning agent is critiqued by a judge; the error and semantic feedback are converted into control signals and stored in a feedback-informed, non-parametric knowledge base. At inference, the model retrieves similar judged experiences from this KB and uses them to modulate its prompt, enabling better, self-aligned reasoning without finetuning. On the PrideMM dataset, FLoReNce improves both predictive performance and explanation quality over static multimodal baselines, showing that feedback-regulated prompting is a viable path to adaptive meme humor understanding.

Method: Proposed “Result -> Justify” approach that constrains output to present conclusion before structured justification, operationalized via SEF framework with six metrics for structure and grounding based on professional conventions like CREAC and BLUF.

Result: All six SEF metrics correlate with correctness (r=0.20-0.42; p<0.001), and SEF achieves 83.9% accuracy (+5.3% improvement over Chain-of-Thought).

Conclusion: Structured justification can improve both verifiability and reliability of AI outputs in high-stakes domains.

Abstract: Explainable AI (XAI) in high-stakes domains should help stakeholders trust and verify system outputs. Yet Chain-of-Thought methods reason before concluding, and logical gaps or hallucinations can yield conclusions that do not reliably align with their rationale. Thus, we propose “Result -> Justify”, which constrains the output communication to present a conclusion before its structured justification. We introduce SEF (Structured Explainability Framework), operationalizing professional conventions (e.g., CREAC, BLUF) via six metrics for structure and grounding. Experiments across four tasks in three domains validate this approach: all six metrics correlate with correctness (r=0.20-0.42; p<0.001), and SEF achieves 83.9% accuracy (+5.3 over CoT). These results suggest structured justification can improve verifiability and may also improve reliability.

Method: Group Pattern Selection Optimization (GPSO) extends GRPO with multi-pattern rollouts, verifier-guided optimal pattern selection per problem, and attention masking to prevent pattern suffix leakage during policy optimization.

Result: GPSO delivers consistent and substantial performance gains across various model backbones and benchmarks, effectively mitigating pattern sub-optimality and fostering more robust, adaptable reasoning.

Conclusion: By enabling models to internalize the mapping from problem characteristics to optimal reasoning patterns, GPSO addresses pattern sub-optimality and improves reasoning adaptability across diverse problems.

Abstract: Large reasoning models (LRMs) exhibit diverse high-level reasoning patterns (e.g., direct solution, reflection-and-verification, and exploring multiple solutions), yet prevailing training recipes implicitly bias models toward a limited set of dominant patterns. Through a systematic analysis, we identify substantial accuracy variance across these patterns on mathematics and science benchmarks, revealing that a model’s default reasoning pattern is often sub-optimal for a given problem. To address this, we introduce Group Pattern Selection Optimization (GPSO), a reinforcement learning framework that extends GRPO by incorporating multi-pattern rollouts, verifier-guided optimal pattern selection per problem, and attention masking during optimization to prevent the leakage of explicit pattern suffixes into the learned policy. By exploring a portfolio of diverse reasoning strategies and optimizing the policy on the most effective ones, GPSO enables the model to internalize the mapping from problem characteristics to optimal reasoning patterns. Extensive experiments demonstrate that GPSO delivers consistent and substantial performance gains across various model backbones and benchmarks, effectively mitigating pattern sub-optimality and fostering more robust, adaptable reasoning. All data and codes are available at https://github.com/wanghanbinpanda/GPSO.

Method: The paper advocates for Stochastic CHAOS (Controlled Heterogeneity and Adaptive Output Sampling) which treats distributional variability as a signal to be measured and controlled, rather than eliminated. This involves multi-sample evaluation and embracing the probabilistic nature of LLM outputs.

Result: Empirical findings show deterministic inference is systematically misleading: it underestimates both capability and fragility, masks failure probabilities, causes emergent abilities to disappear under greedy decoding, degrades multi-path reasoning accuracy, and underestimates safety risks by hiding rare dangerous behaviors.

Conclusion: Deterministic inference should be abandoned for LLMs as it fundamentally misrepresents their nature as conditional distributions. Embracing stochastic variability through approaches like Stochastic CHAOS provides more accurate assessment of capabilities, reasoning, and safety risks.

Abstract: Deterministic inference is a comforting ideal in classical software: the same program on the same input should always produce the same output. As large language models move into real-world deployment, this ideal has been imported wholesale into inference stacks. Recent work from the Thinking Machines Lab has presented a detailed analysis of nondeterminism in LLM inference, showing how batch-invariant kernels and deterministic attention can enforce bitwise-identical outputs, positioning deterministic inference as a prerequisite for reproducibility and enterprise reliability. In this paper, we take the opposite stance. We argue that, for LLMs, deterministic inference kills. It kills the ability to model uncertainty, suppresses emergent abilities, collapses reasoning into a single brittle path, and weakens safety alignment by hiding tail risks. LLMs implement conditional distributions over outputs, not fixed functions. Collapsing these distributions to a single canonical completion may appear reassuring, but it systematically conceals properties central to artificial cognition. We instead advocate Stochastic CHAOS, treating distributional variability as a signal to be measured and controlled. Empirically, we show that deterministic inference is systematically misleading. Single-sample deterministic evaluation underestimates both capability and fragility, masking failure probability under paraphrases and noise. Phase-like transitions associated with emergent abilities disappear under greedy decoding. Multi-path reasoning degrades when forced onto deterministic backbones, reducing accuracy and diagnostic insight. Finally, deterministic evaluation underestimates safety risk by hiding rare but dangerous behaviors that appear only under multi-sample evaluation.

Method: Multi-Model Consensus Reasoning Engine that treats LLM outputs as input to supervised meta-learner. Maps responses into structured features (semantic embeddings, pairwise similarity, clustering statistics, lexical/structural cues, reasoning-quality scores, confidence estimates, model-specific priors). Uses gradient-boosted trees, listwise ranking, and graph neural networks over similarity graphs.

Result: Best graph-attention-based consensus model improves macro-average accuracy by 4.6 percentage points over strongest single LLM and 8.1 points over majority vote. Lower Brier scores and fewer TruthfulQA hallucinations. Semantic agreement and clustering features most influential.

Conclusion: Supervised multi-model consensus is practical route toward more reliable LLM behavior, even in modest single-machine setups, with semantic agreement and clustering features being most important.

Abstract: Large language models (LLMs) achieve strong aver- age performance yet remain unreliable at the instance level, with frequent hallucinations, brittle failures, and poorly calibrated confidence. We study reliability through the lens of multi-model consensus: given responses from several heterogeneous LLMs, can we learn which answer is most likely correct for a given query? We introduce a Multi-Model Consensus Reasoning Engine that treats the set of LLM outputs as input to a supervised meta-learner. The system maps natural language responses into structured features using semantic embeddings, pairwise similarity and clustering statistics, lexical and structural cues, reasoning-quality scores, confidence estimates, and model-specific priors, and then applies gradient-boosted trees, listwise ranking, and graph neural networks over similarity graphs of answers. Using three open-weight LLMs evaluated on compact, resource- constrained subsets of GSM8K, ARC-Challenge, HellaSwag, and TruthfulQA, our best graph-attention-based consensus model improves macro-average accuracy by 4.6 percentage points over the strongest single LLM and by 8.1 points over majority vote, while also yielding lower Brier scores and fewer TruthfulQA hal- lucinations. Ablation and feature-importance analyses show that semantic agreement and clustering features are most influential, with reasoning-quality and model-prior features providing com- plementary gains, suggesting supervised multi-model consensus is a practical route toward more reliable LLM behavior, even in a modest single-machine setup.

Method: LRAS (Legal Reasoning with Agentic Search) integrates Introspective Imitation Learning and Difficulty-aware Reinforcement Learning to enable legal LLMs to identify knowledge boundaries and handle legal reasoning complexity through dynamic, interactive “Active Inquiry” rather than static parametric reasoning.

Result: LRAS outperforms state-of-the-art baselines by 8.2-32%, with the most substantial gains in tasks requiring deep reasoning with reliable knowledge.

Conclusion: The LRAS framework successfully addresses the limitations of current legal LLMs by transitioning them from closed-loop to active inquiry reasoning, significantly improving performance on complex legal reasoning tasks while maintaining the procedural rigor required in legal domains.

Abstract: While Large Reasoning Models (LRMs) have demonstrated exceptional logical capabilities in mathematical domains, their application to the legal field remains hindered by the strict requirements for procedural rigor and adherence to legal logic. Existing legal LLMs, which rely on “closed-loop reasoning” derived solely from internal parametric knowledge, frequently suffer from lack of self-awareness regarding their knowledge boundaries, leading to confident yet incorrect conclusions. To address this challenge, we present Legal Reasoning with Agentic Search (LRAS), the first framework designed to transition legal LLMs from static and parametric “closed-loop thinking” to dynamic and interactive “Active Inquiry”. By integrating Introspective Imitation Learning and Difficulty-aware Reinforcement Learning, LRAS enables LRMs to identify knowledge boundaries and handle legal reasoning complexity. Empirical results demonstrate that LRAS outperforms state-of-the-art baselines by 8.2-32%, with the most substantial gains observed in tasks requiring deep reasoning with reliable knowledge. We will release our data and models for further exploration soon.

Method: Agent-Role Merging (ARM) uses a 3-step framework: 1) constructing merged backbones, 2) selection based on role-conditioned activation analysis, and 3) neuron transplantation for fine-grained refinements. It’s activation-guided and role-conditioned.

Result: ARM outperforms prior model merging methods and domain-specific expert models across diverse domains, while demonstrating strong out-of-domain generalization without gradient-based optimization.

Conclusion: ARM successfully extends model merging from static natural language tasks to multi-turn agent scenarios, improving cross-benchmark generalization while maintaining efficiency through training-free neuron transplantation.

Abstract: Interactive large language model agents have advanced rapidly, but most remain specialized to a single environment and fail to adapt robustly to other environments. Model merging offers a training-free alternative by integrating multiple experts into a single model. In this paper, we propose Agent-Role Merging (ARM), an activation-guided, role-conditioned neuron transplantation method for model merging in LLM agents. ARM improves existing merging methods from static natural language tasks to multi-turn agent scenarios, and over the generalization ability across various interactive environments. This is achieved with a well designed 3-step framework: 1) constructing merged backbones, 2) selection based on its role-conditioned activation analysis, and 3) neuron transplantation for fine-grained refinements. Without gradient-based optimization, ARM improves cross-benchmark generalization while enjoying efficiency. Across diverse domains, the model obtained via ARM merging outperforms prior model merging methods and domain-specific expert models, while demonstrating strong out-of-domain generalization.

Method: Agentic diagnostic framework where LLM performs step-wise investigation using constrained tool space exposed through Model Context Protocol (MCP). Agent autonomously navigates infrastructure model by invoking tools for service lookup, dependency retrieval, structured/unstructured data analysis, event analysis, and impact discovery.

Result: The framework enables autonomous navigation of infrastructure models with structured investigation protocol that ensures grounding, reproducibility, and safe handling of missing/ambiguous information.

Conclusion: This work lays foundation for autonomous incident resolution and change impact mitigation, with future systems capable of not only diagnosing failures but also predicting impact of planned changes to enable risk mitigation before maintenance operations.

Abstract: Large-scale telecom and datacenter infrastructures rely on multi-layered service and resource models, where failures propagate across physical and logical components and affect multiple customers. Traditional approaches to root cause analysis(RCA) rely on hard-coded graph traversal algorithms or rule-based correlation engines, which are costly to maintain and tightly coupled to the infrastructure model. In this work, we introduce an agentic diagnostic framework where a Large Language Model (LLM) performs step-wise investigation using a constrained tool space exposed through the Model Context Protocol (MCP). Instead of embedding causal logic or traversal algorithms into the application, the agent autonomously navigates the infrastructure model by invoking tools for service lookup, dependency retrieval, structured and unstructured data, and event analysis, and impact discovery. We define an investigation protocol that structures the agent’s reasoning and ensures grounding, reproducibility, and safe handling of missing or ambiguous information. This work lays the foundation for autonomous incident resolution and change impact mitigation. Future systems will not only diagnose and remediate infrastructure failures, but also predict the impact of planned changes on services and customers, enabling operators to mitigate risks before executing maintenance operations.

Method: The paper uses Carnap’s methodology of conceptual clarification (similarity to explicandum, exactness, fruitfulness, simplicity) to evaluate six existing intelligence definitions, then proposes the Extended Predictive Hypothesis which combines predictive ability with benefit-gaining capability.

Result: Analysis shows predictive ability definitions have high explanatory power but fail to explain the prediction-behavior-benefit relationship. The proposed EPH framework distinguishes spontaneous/reactive predictions and adds gainability, providing a unified explanation for creativity, learning, and future planning.

Conclusion: The Extended Predictive Hypothesis is argued to be the most satisfactory and universal definition for comparing human and AI intelligence, offering a comprehensive framework that addresses limitations of existing definitions.

Abstract: This paper aims to propose a universal definition of intelligence that enables fair and consistent comparison of human and artificial intelligence (AI). With the rapid development of AI technology in recent years, how to compare and evaluate human and AI intelligence has become an important theoretical issue. However, existing definitions of intelligence are anthropocentric and unsuitable for empirical comparison, resulting in a lack of consensus in the research field. This paper first introduces four criteria for evaluating intelligence definitions based on R. Carnap’s methodology of conceptual clarification: similarity to explicandum, exactness, fruitfulness, and simplicity. We then examine six representative definitions: IQ testing, complex problem-solving ability, reward optimization, environmental adaptation, learning efficiency, and predictive ability, and clarify their theoretical strengths and limitations. The results show that while definitions based on predictive ability have high explanatory power and empirical feasibility, they suffer from an inability to adequately explain the relationship between predictions and behavior/benefits. This paper proposes the Extended Predictive Hypothesis (EPH), which views intelligence as a combination of the ability to accurately predict the future and the ability to benefit from those predictions. Furthermore, by distinguishing predictive ability into spontaneous and reactive predictions and adding the concept of gainability, we present a unified framework for explaining various aspects of intelligence, such as creativity, learning, and future planning. In conclusion, this paper argues that the EPH is the most satisfactory and universal definition for comparing human and AI intelligence.

Method: Decomposes agentic learning systems into lightweight, composable components with clear abstraction boundaries. Features a centralized scheduler for managing training/inference workloads (LoRA-based/full-parameter RL, supervised fine-tuning, inference) over shared resources, with design principles for multi-agent training extension.

Result: Presents a set of RL use cases demonstrating the framework’s effectiveness in practical agentic learning scenarios, showing it can handle diverse training approaches including LoRA-based and full-parameter RL.

Conclusion: OpenTinker provides a flexible, modular infrastructure for RL training of LLM agents that separates concerns and enables efficient resource management, with demonstrated effectiveness in practical applications and extensibility to multi-agent scenarios.

Abstract: We introduce OpenTinker, an infrastructure for reinforcement learning (RL) of large language model (LLM) agents built around a separation of concerns across algorithm design, execution, and agent-environment interaction. Rather than relying on monolithic, end-to-end RL pipelines, OpenTinker decomposes agentic learning systems into lightweight, composable components with clearly defined abstraction boundaries. Users specify agents, environments, and interaction protocols, while inference and training are delegated to a managed execution runtime. OpenTinker introduces a centralized scheduler for managing training and inference workloads, including LoRA-based and full-parameter RL, supervised fine-tuning, and inference, over shared resources. We further discuss design principles for extending OpenTinker to multi-agent training. Finally, we present a set of RL use cases that demonstrate the effectiveness of the framework in practical agentic learning scenarios.

Method: Proposes a reusable software-hardware co-optimization and closed-loop evaluation framework that jointly integrates software-level model optimization with hardware-level computation optimization under a unified system-level objective. Includes multidimensional evaluation metrics covering safety, comfort, efficiency, latency, and energy.

Result: The framework preserves baseline-level driving performance while significantly reducing inference latency and energy consumption across multiple ME2E autonomous driving stacks, achieving substantial overall system-level improvements.

Conclusion: The proposed framework provides practical and actionable guidance for efficient deployment of ME2E autonomous driving systems by addressing critical system-level factors through joint software-hardware optimization.

Abstract: Modular end-to-end (ME2E) autonomous driving paradigms combine modular interpretability with global optimization capability and have demonstrated strong performance. However, existing studies mainly focus on accuracy improvement, while critical system-level factors such as inference latency and energy consumption are often overlooked, resulting in increasingly complex model designs that hinder practical deployment. Prior efforts on model compression and acceleration typically optimize either the software or hardware side in isolation. Software-only optimization cannot fundamentally remove intermediate tensor access and operator scheduling overheads, whereas hardware-only optimization is constrained by model structure and precision. As a result, the real-world benefits of such optimizations are often limited. To address these challenges, this paper proposes a reusable software and hardware co-optimization and closed-loop evaluation framework for ME2E autonomous driving inference. The framework jointly integrates software-level model optimization with hardware-level computation optimization under a unified system-level objective. In addition, a multidimensional evaluation metric is introduced to assess system performance by jointly considering safety, comfort, efficiency, latency, and energy, enabling quantitative comparison of different optimization strategies. Experiments across multiple ME2E autonomous driving stacks show that the proposed framework preserves baseline-level driving performance while significantly reducing inference latency and energy consumption, achieving substantial overall system-level improvements. These results demonstrate that the proposed framework provides practical and actionable guidance for efficient deployment of ME2E autonomous driving systems.

Method: Proposes LOGO (Local-to-Global) world model that leverages easier-to-estimate local predictions to infer global state dynamics. Uses trained model to generate synthetic data for dataset augmentation, with uncertainty-aware sampling that weights synthetic data by prediction uncertainty. Requires only an additional encoder for uncertainty estimation instead of conventional ensembles.

Result: Extensive experiments across 8 scenarios against 8 baselines show the method surpasses state-of-the-art baselines on standard offline MARL benchmarks, establishing a new model-based baseline for generalizable offline multi-agent learning.

Conclusion: LOGO world model effectively addresses the challenges of offline MARL by improving prediction accuracy through local-to-global inference, enabling reliable dataset expansion with uncertainty-aware sampling, and achieving better generalization with reduced computational overhead.

Abstract: Offline multi-agent reinforcement learning (MARL) aims to solve cooperative decision-making problems in multi-agent systems using pre-collected datasets. Existing offline MARL methods primarily constrain training within the dataset distribution, resulting in overly conservative policies that struggle to generalize beyond the support of the data. While model-based approaches offer a promising solution by expanding the original dataset with synthetic data generated from a learned world model, the high dimensionality, non-stationarity, and complexity of multi-agent systems make it challenging to accurately estimate the transitions and reward functions in offline MARL. Given the difficulty of directly modeling joint dynamics, we propose a local-to-global (LOGO) world model, a novel framework that leverages local predictions-which are easier to estimate-to infer global state dynamics, thus improving prediction accuracy while implicitly capturing agent-wise dependencies. Using the trained world model, we generate synthetic data to augment the original dataset, expanding the effective state-action space. To ensure reliable policy learning, we further introduce an uncertainty-aware sampling mechanism that adaptively weights synthetic data by prediction uncertainty, reducing approximation error propagation to policies. In contrast to conventional ensemble-based methods, our approach requires only an additional encoder for uncertainty estimation, significantly reducing computational overhead while maintaining accuracy. Extensive experiments across 8 scenarios against 8 baselines demonstrate that our method surpasses state-of-the-art baselines on standard offline MARL benchmarks, establishing a new model-based baseline for generalizable offline multi-agent learning.

Method: IFDNS employs a multi-round feedback mechanism during logic extraction to accurately extract causal relationship statements and translate them into propositional and logical implication expressions, mitigating information loss. It’s orthogonal to existing prompt methods and can integrate with various prompting approaches.

Result: Empirical evaluations across six datasets show IFDNS significantly improves CoT and CoT-SC performance, achieving +9.40% accuracy boost for CoT on LogiQA and +11.70% improvement for CoT-SC on PrOntoQA.

Conclusion: IFDNS effectively addresses LLM limitations in handling complex logical relationships through iterative feedback, reducing information loss and improving reasoning faithfulness while being compatible with existing prompt methods.

Abstract: Large language models (LLMs) have demonstrated impressive capabilities across a wide range of reasoning tasks, including logical and mathematical problem-solving. While prompt-based methods like Chain-of-Thought (CoT) can enhance LLM reasoning abilities to some extent, they often suffer from a lack of faithfulness, where the derived conclusions may not align with the generated reasoning chain. To address this issue, researchers have explored neuro-symbolic approaches to bolster LLM logical reasoning capabilities. However, existing neuro-symbolic methods still face challenges with information loss during the process. To overcome these limitations, we introduce Iterative Feedback-Driven Neuro-Symbolic (IFDNS), a novel prompt-based method that employs a multi-round feedback mechanism to address LLM limitations in handling complex logical relationships. IFDNS utilizes iterative feedback during the logic extraction phase to accurately extract causal relationship statements and translate them into propositional and logical implication expressions, effectively mitigating information loss issues. Furthermore, IFDNS is orthogonal to existing prompt methods, allowing for seamless integration with various prompting approaches. Empirical evaluations across six datasets demonstrate the effectiveness of IFDNS in significantly improving the performance of CoT and Chain-of-Thought with Self-Consistency (CoT-SC). Specifically, IFDNS achieves a +9.40% accuracy boost for CoT on the LogiQA dataset and a +11.70% improvement for CoT-SC on the PrOntoQA dataset.

Method: Proposes Temporal Semantic Memory (TSM) framework that: 1) Builds semantic timelines instead of dialogue timelines during memory construction, 2) Consolidates temporally continuous and semantically related information into durative memories, and 3) During utilization, incorporates query’s temporal intent on semantic timeline to retrieve temporally appropriate durative memories for time-valid, duration-consistent context.

Result: Experiments on LongMemEval and LoCoMo benchmarks show TSM consistently outperforms existing methods, achieving up to 12.2% absolute improvement in accuracy.

Conclusion: TSM effectively addresses temporal modeling limitations in LLM memory systems by introducing semantic timelines and durative memory construction, demonstrating significant performance improvements over existing approaches.

Abstract: Memory enables Large Language Model (LLM) agents to perceive, store, and use information from past dialogues, which is essential for personalization. However, existing methods fail to properly model the temporal dimension of memory in two aspects: 1) Temporal inaccuracy: memories are organized by dialogue time rather than their actual occurrence time; 2) Temporal fragmentation: existing methods focus on point-wise memory, losing durative information that captures persistent states and evolving patterns. To address these limitations, we propose Temporal Semantic Memory (TSM), a memory framework that models semantic time for point-wise memory and supports the construction and utilization of durative memory. During memory construction, it first builds a semantic timeline rather than a dialogue one. Then, it consolidates temporally continuous and semantically related information into a durative memory. During memory utilization, it incorporates the query’s temporal intent on the semantic timeline, enabling the retrieval of temporally appropriate durative memories and providing time-valid, duration-consistent context to support response generation. Experiments on LongMemEval and LoCoMo show that TSM consistently outperforms existing methods and achieves up to 12.2% absolute improvement in accuracy, demonstrating the effectiveness of the proposed method.

Method: The study experiments with LLMs of varying sizes on two state-of-the-art HAR datasets. It investigates performance evolution based on LLM size and employs knowledge distillation techniques to fine-tune smaller LLMs using HAR reasoning examples generated by larger LLMs.

Result: Recognition performance improves with larger LLM size. More importantly, fine-tuned smaller LLMs using knowledge distillation can perform almost as well as the largest LLMs while having 50 times fewer parameters.

Conclusion: Knowledge distillation enables efficient deployment of LLMs for HAR by allowing smaller models to achieve comparable performance to much larger models, making LLM-based HAR more practical for resource-constrained applications.

Abstract: Human Activity Recognition (HAR) is a central problem for context-aware applications, especially for smart homes and assisted living. A few very recent studies have shown that Large Language Models (LLMs) can be used for HAR at home, reaching high performance and addressing key challenges. In this paper, we provide new experimental results regarding the use of LLMs for HAR, on two state-of-the-art datasets. More specifically, we show how recognition performance evolves depending on the size of the LLM used. Moreover, we experiment on the use of knowledge distillation techniques to fine-tune smaller LLMs with HAR reasoning examples generated by larger LLMs. We show that such fine-tuned models can perform almost as well as the largest LLMs, while having 50 times less parameters.

Method: Decouples task execution from memory management using a frozen task model and learned memory copilot trained with direct preference optimization; organizes memories into hierarchical abstraction levels for selective reuse.

Result: Substantial improvements in performance, out-of-distribution generalization, and cross-task transfer on ALFWorld, ScienceWorld, and BabyAI benchmarks compared to baselines.

Conclusion: Treating memory abstraction as a learnable cognitive skill rather than fixed design enables better adaptation and transfer across tasks with distribution shifts.

Abstract: Large language model (LLM) agents increasingly rely on accumulated memory to solve long-horizon decision-making tasks. However, most existing approaches store memory in fixed representations and reuse it at a single or implicit level of abstraction, which limits generalization and often leads to negative transfer when distribution shift. This paper proposes the Meta-Cognitive Memory Abstraction method (MCMA), which treats memory abstraction as a learnable cognitive skill rather than a fixed design choice. MCMA decouples task execution from memory management by combining a frozen task model with a learned memory copilot. The memory copilot is trained using direct preference optimization, it determines how memories should be structured, abstracted, and reused. Memories are further organized into a hierarchy of abstraction levels, enabling selective reuse based on task similarity. When no memory is transferable, MCMA transfers the ability to abstract and manage memory by transferring the memory copilot. Experiments on ALFWorld, ScienceWorld, and BabyAI demonstrate substantial improvements in performance, out-of-distribution generalization, and cross-task transfer over several baselines.

Method: Proposes JudgeFlow pipeline with: 1) reusable configurable logic blocks in workflows, 2) Judge module that inspects execution traces (especially failures) and assigns rank-based responsibility scores to problematic blocks, 3) LLM-based optimizer that focuses modifications on the most problematic block.

Result: Achieves superior performance and efficiency on mathematical reasoning and code generation benchmarks compared to existing methods, with improved sample efficiency and interpretability through block-level diagnostics.

Conclusion: JudgeFlow provides a scalable foundation for automating increasingly complex agentic workflows by offering fine-grained diagnostic signals and targeted optimization, making LLM-based agent optimization more efficient and interpretable.

Abstract: Optimizing LLM-based agentic workflows is challenging for scaling AI capabilities. Current methods rely on coarse, end-to-end evaluation signals and lack fine-grained signals on where to refine, often resulting in inefficient or low-impact modifications. To address these limitations, we propose {\our{}}, an Evaluation-Judge-Optimization-Update pipeline. We incorporate reusable, configurable logic blocks into agentic workflows to capture fundamental forms of logic. On top of this abstraction, we design a dedicated Judge module that inspects execution traces – particularly failed runs – and assigns rank-based responsibility scores to problematic blocks. These fine-grained diagnostic signals are then leveraged by an LLM-based optimizer, which focuses modifications on the most problematic block in the workflow. Our approach improves sample efficiency, enhances interpretability through block-level diagnostics, and provides a scalable foundation for automating increasingly complex agentic workflows. We evaluate {\our{}} on mathematical reasoning and code generation benchmarks, where {\our{}} achieves superior performance and efficiency compared to existing methods. The source code is publicly available at https://github.com/ma-zihan/JudgeFlow.

Method: Built on Unreal Engine 5 with user-friendly API for LLM-driven agent control via natural language. Integrates LLMs/VLMs for environment generation and structured tasks from multimodal inputs. Features procedural task generation, validation, and real-time environment control.

Result: Benchmarked performance of popular LLMs across tasks of increasing complexity, analyzing adaptability, planning, and multi-agent coordination differences. Platform released as open-source.

Conclusion: VirtualEnv aims to advance AI-gaming research, enable standardized LLM evaluation in embodied AI, and pave the way for immersive simulations and interactive entertainment development.

Abstract: As large language models (LLMs) continue to improve in reasoning and decision-making, there is a growing need for realistic and interactive environments where their abilities can be rigorously evaluated. We present VirtualEnv, a next-generation simulation platform built on Unreal Engine 5 that enables fine-grained benchmarking of LLMs in embodied and interactive scenarios. VirtualEnv supports rich agent-environment interactions, including object manipulation, navigation, and adaptive multi-agent collaboration, as well as game-inspired mechanics like escape rooms and procedurally generated environments. We provide a user-friendly API built on top of Unreal Engine, allowing researchers to deploy and control LLM-driven agents using natural language instructions. We integrate large-scale LLMs and vision-language models (VLMs), such as GPT-based models, to generate novel environments and structured tasks from multimodal inputs. Our experiments benchmark the performance of several popular LLMs across tasks of increasing complexity, analyzing differences in adaptability, planning, and multi-agent coordination. We also describe our methodology for procedural task generation, task validation, and real-time environment control. VirtualEnv is released as an open-source platform, we aim to advance research at the intersection of AI and gaming, enable standardized evaluation of LLMs in embodied AI settings, and pave the way for future developments in immersive simulations and interactive entertainment.

Method: Task-Decoupled Planning (TDP) decomposes tasks into a directed acyclic graph (DAG) of sub-goals via a Supervisor, then uses a Planner and Executor with scoped contexts to confine reasoning and replanning to the active sub-task only.

Result: TDP outperforms strong baselines on TravelPlanner, ScienceWorld, and HotpotQA benchmarks while reducing token consumption by up to 82%, demonstrating improved robustness and efficiency.

Conclusion: Sub-task decoupling through TDP improves both robustness and efficiency for long-horizon agents by preventing error propagation and enabling local recovery without disrupting the overall workflow.

Abstract: Recent advances in large language models (LLMs) have enabled agents to autonomously execute complex, long-horizon tasks, yet planning remains a primary bottleneck for reliable task execution. Existing methods typically fall into two paradigms: step-wise planning, which is reactive but often short-sighted; and one-shot planning, which generates a complete plan upfront yet is brittle to execution errors. Crucially, both paradigms suffer from entangled contexts, where the agent must reason over a monolithic history spanning multiple sub-tasks. This entanglement increases cognitive load and lets local errors propagate across otherwise independent decisions, making recovery computationally expensive. To address this, we propose Task-Decoupled Planning (TDP), a training-free framework that replaces entangled reasoning with task decoupling. TDP decomposes tasks into a directed acyclic graph (DAG) of sub-goals via a Supervisor. Using a Planner and Executor with scoped contexts, TDP confines reasoning and replanning to the active sub-task. This isolation prevents error propagation and corrects deviations locally without disrupting the workflow. Results on TravelPlanner, ScienceWorld, and HotpotQA show that TDP outperforms strong baselines while reducing token consumption by up to 82%, demonstrating that sub-task decoupling improves both robustness and efficiency for long-horizon agents.

Method: DIAGPaper uses three integrated modules: (1) Customizer module simulates human-defined review criteria and instantiates reviewer agents with criterion-specific expertise, (2) Rebuttal module introduces author agents that engage in structured debate with reviewer agents to validate/refine weaknesses, (3) Prioritizer module learns from large-scale human review practices to assess severity and surfaces top-K severest weaknesses.

Result: Experiments on AAAR and ReviewCritique benchmarks show DIAGPaper substantially outperforms existing methods by producing more valid and paper-specific weaknesses, presented in a user-oriented, prioritized manner.

Conclusion: DIAGPaper addresses key limitations in existing paper weakness identification systems through its integrated multi-agent framework that combines criterion-based expertise, rebuttal validation, and severity prioritization for more effective and user-friendly weakness analysis.

Abstract: Paper weakness identification using single-agent or multi-agent LLMs has attracted increasing attention, yet existing approaches exhibit key limitations. Many multi-agent systems simulate human roles at a surface level, missing the underlying criteria that lead experts to assess complementary intellectual aspects of a paper. Moreover, prior methods implicitly assume identified weaknesses are valid, ignoring reviewer bias, misunderstanding, and the critical role of author rebuttals in validating review quality. Finally, most systems output unranked weakness lists, rather than prioritizing the most consequential issues for users. In this work, we propose DIAGPaper, a novel multi-agent framework that addresses these challenges through three tightly integrated modules. The customizer module simulates human-defined review criteria and instantiates multiple reviewer agents with criterion-specific expertise. The rebuttal module introduces author agents that engage in structured debate with reviewer agents to validate and refine proposed weaknesses. The prioritizer module learns from large-scale human review practices to assess the severity of validated weaknesses and surfaces the top-K severest ones to users. Experiments on two benchmarks, AAAR and ReviewCritique, demonstrate that DIAGPaper substantially outperforms existing methods by producing more valid and more paper-specific weaknesses, while presenting them in a user-oriented, prioritized manner.

Method: Extends SALT benchmark by linking multi-table transactional data with an Operational Business Knowledge Graph (OBKG) that captures field-level descriptions, relational dependencies, and business object types. This creates a semantics-aware benchmark for evaluating models on tabular prediction tasks enhanced with metadata knowledge.

Result: Metadata-derived features yield modest improvements in classical prediction metrics, but consistently reveal gaps in models’ ability to leverage semantics in relational context. The benchmark provides empirical evidence of current limitations in semantics-aware tabular reasoning.

Conclusion: SALT-KG establishes a benchmark for advancing tabular foundation models grounded in declarative knowledge, representing the first empirical step toward semantically linked tables in enterprise-scale structured data by reframing tabular prediction as semantics-conditioned reasoning.

Abstract: Building upon the SALT benchmark for relational prediction (Klein et al., 2024), we introduce SALT-KG, a benchmark for semantics-aware learning on enterprise tables. SALT-KG extends SALT by linking its multi-table transactional data with a structured Operational Business Knowledge represented in a Metadata Knowledge Graph (OBKG) that captures field-level descriptions, relational dependencies, and business object types. This extension enables evaluation of models that jointly reason over tabular evidence and contextual semantics, an increasingly critical capability for foundation models on structured data. Empirical analysis reveals that while metadata-derived features yield modest improvements in classical prediction metrics, these metadata features consistently highlight gaps in the ability of models to leverage semantics in relational context. By reframing tabular prediction as semantics-conditioned reasoning, SALT-KG establishes a benchmark to advance tabular foundation models grounded in declarative knowledge, providing the first empirical step toward semantically linked tables in structured data at enterprise scale.

Method: Extends Chen et al. (2025) work by testing LRMs on multiple choice questions with hints, directly asking them to reflect on unusual prompt content, and allowing them to use hints while measuring actual hint usage.

Result: LRMs flatly deny relying on provided hints in answering questions, even when directly questioned about unusual prompt content and even though experiments demonstrate they are actually using the hints.

Conclusion: Results have discouraging implications for CoT monitoring and interpretability, showing LRMs will lie about their reasoning processes rather than just omitting information.

Abstract: It has been shown that Large Reasoning Models (LRMs) may not say what they think: they do not always volunteer information about how certain parts of the input influence their reasoning. But it is one thing for a model to omit such information and another, worse thing to lie about it. Here, we extend the work of Chen et al. (2025) to show that LRMs will do just this: they will flatly deny relying on hints provided in the prompt in answering multiple choice questions – even when directly asked to reflect on unusual (i.e. hinted) prompt content, even when allowed to use hints, and even though experiments show them to be using the hints. Our results thus have discouraging implications for CoT monitoring and interpretability.

Method: The study applies supervised learning algorithms including K-Nearest Neighbors (KNN), Quadratic Discriminant Analysis (QDA), Linear Discriminant Analysis (LDA), and Gaussian Process Classifiers. To address class imbalance and improve performance, the researchers used Synthetic Minority Over-sampling Technique (SMOTE) and Term Frequency-Inverse Document Frequency (TF-IDF) vectorization.

Result: Among all tested models, Linear Discriminant Analysis (LDA) achieved the highest testing accuracy of 98%. The study also identified important features correlated with dementia, including the presence of the APOE-epsilon4 allele and chronic conditions like diabetes.

Conclusion: The research highlights the importance of model interpretability in dementia prediction and advocates for future machine learning innovations, particularly in integrating explainable AI approaches, to further improve predictive capabilities in dementia care.

Abstract: Dementia is a complex syndrome impacting cognitive and emotional functions, with Alzheimer’s disease being the most common form. This study focuses on enhancing dementia prediction using machine learning (ML) techniques on patient health data. Supervised learning algorithms are applied in this study, including K-Nearest Neighbors (KNN), Quadratic Discriminant Analysis (QDA), Linear Discriminant Analysis (LDA), and Gaussian Process Classifiers. To address class imbalance and improve model performance, techniques such as Synthetic Minority Over-sampling Technique (SMOTE) and Term Frequency-Inverse Document Frequency (TF-IDF) vectorization were employed. Among the models, LDA achieved the highest testing accuracy of 98%. This study highlights the importance of model interpretability and the correlation of dementia with features such as the presence of the APOE-epsilon4 allele and chronic conditions like diabetes. This research advocates for future ML innovations, particularly in integrating explainable AI approaches, to further improve predictive capabilities in dementia care.

Method: Used real-world journalctl data from Linux production servers to evaluate nine small language models (SLMs) and small reasoning language models (SRLMs) under zero-shot, few-shot, and retrieval-augmented generation (RAG) prompting. Measured both accuracy and inference efficiency.

Result: Strong stratification among models: Qwen3-4B achieved highest accuracy (95.64% with RAG), Gemma3-1B improved from 20.25% (few-shot) to 85.28% (RAG), while several SRLMs degraded with RAG. Efficiency varied widely - most Gemma/Llama variants completed inference in <1.2 seconds, while Phi-4-Mini-Reasoning took >228 seconds with <10% accuracy.

Conclusion: Severity classification serves as a lens for evaluating model competence and real-time deployability. Architectural design, training objectives, and ability to integrate retrieved context under strict output constraints jointly determine performance. This benchmark aligns with real-time requirements of digital twin systems and has implications for root cause analysis.

Abstract: System logs are crucial for monitoring and diagnosing modern computing infrastructure, but their scale and complexity require reliable and efficient automated interpretation. Since severity levels are predefined metadata in system log messages, having a model merely classify them offers limited standalone practical value, revealing little about its underlying ability to interpret system logs. We argue that severity classification is more informative when treated as a benchmark for probing runtime log comprehension rather than as an end task. Using real-world journalctl data from Linux production servers, we evaluate nine small language models (SLMs) and small reasoning language models (SRLMs) under zero-shot, few-shot, and retrieval-augmented generation (RAG) prompting. The results reveal strong stratification. Qwen3-4B achieves the highest accuracy at 95.64% with RAG, while Gemma3-1B improves from 20.25% under few-shot prompting to 85.28% with RAG. Notably, the tiny Qwen3-0.6B reaches 88.12% accuracy despite weak performance without retrieval. In contrast, several SRLMs, including Qwen3-1.7B and DeepSeek-R1-Distill-Qwen-1.5B, degrade substantially when paired with RAG. Efficiency measurements further separate models: most Gemma and Llama variants complete inference in under 1.2 seconds per log, whereas Phi-4-Mini-Reasoning exceeds 228 seconds per log while achieving <10% accuracy. These findings suggest that (1) architectural design, (2) training objectives, and (3) the ability to integrate retrieved context under strict output constraints jointly determine performance. By emphasizing small, deployable models, this benchmark aligns with real-time requirements of digital twin (DT) systems and shows that severity classification serves as a lens for evaluating model competence and real-time deployability, with implications for root cause analysis (RCA) and broader DT integration.

Method: Inspired by GANs’ generator/discriminator framework, CGR uses a critic model that periodically probes its own reasoning to assess confidence. Reasoning continues until a target certainty threshold is met, allowing early termination when confident and extended reasoning when uncertain.

Result: CGR improves baseline accuracy on AIME2024 and AIME2025 datasets while reducing token usage. Multi-seed evaluations (64 runs) show stable performance with reduced variance, improved exam-like performance under penalty-based grading, and aggregate token savings of millions.

Conclusion: Certainty is a powerful signal for reasoning sufficiency. CGR makes large reasoning language models more adaptive, trustworthy, and resource-efficient, enabling practical deployment in domains where both accuracy and computational cost matter.

Abstract: The rise of large reasoning language models (LRLMs) has unlocked new potential for solving complex tasks. These models operate with a thinking budget, that is, a predefined number of reasoning tokens used to arrive at a solution. We propose a novel approach, inspired by the generator/discriminator framework in generative adversarial networks, in which a critic model periodically probes its own reasoning to assess whether it has reached a confident conclusion. If not, reasoning continues until a target certainty threshold is met. This mechanism adaptively balances efficiency and reliability by allowing early termination when confidence is high, while encouraging further reasoning when uncertainty persists. Through experiments on the AIME2024 and AIME2025 datasets, we show that Certainty-Guided Reasoning (CGR) improves baseline accuracy while reducing token usage. Importantly, extended multi-seed evaluations over 64 runs demonstrate that CGR is stable, reducing variance across seeds and improving exam-like performance under penalty-based grading. Additionally, our token savings analysis shows that CGR can eliminate millions of tokens in aggregate, with tunable trade-offs between certainty thresholds and efficiency. Together, these findings highlight certainty as a powerful signal for reasoning sufficiency. By integrating confidence into the reasoning process, CGR makes large reasoning language models more adaptive, trustworthy, and resource efficient, paving the way for practical deployment in domains where both accuracy and computational cost matter.

Method: FedKD uses knowledge distillation where a low-dimensional student model mimics the score distribution of a high-dimensional teacher model using KL divergence loss. It adaptively learns a temperature for positive triples, uses predefined temperature for negative triples to mitigate teacher over-confidence, and dynamically adjusts KD loss weight.

Result: Extensive experiments on three datasets demonstrate the effectiveness of FedKD in compressing embeddings while maintaining performance.

Conclusion: FedKD provides an effective solution for embedding compression in federated knowledge graph embedding settings, addressing storage, inference speed, and communication efficiency challenges.

Abstract: Federated Knowledge Graph Embedding (FKGE) aims to facilitate collaborative learning of entity and relation embeddings from distributed Knowledge Graphs (KGs) across multiple clients, while preserving data privacy. Training FKGE models with higher dimensions is typically favored due to their potential for achieving superior performance. However, high-dimensional embeddings present significant challenges in terms of storage resource and inference speed. Unlike traditional KG embedding methods, FKGE involves multiple client-server communication rounds, where communication efficiency is critical. Existing embedding compression methods for traditional KGs may not be directly applicable to FKGE as they often require multiple model trainings which potentially incur substantial communication costs. In this paper, we propose a light-weight component based on Knowledge Distillation (KD) which is titled FedKD and tailored specifically for FKGE methods. During client-side local training, FedKD facilitates the low-dimensional student model to mimic the score distribution of triples from the high-dimensional teacher model using KL divergence loss. Unlike traditional KD way, FedKD adaptively learns a temperature to scale the score of positive triples and separately adjusts the scores of corresponding negative triples using a predefined temperature, thereby mitigating teacher over-confidence issue. Furthermore, we dynamically adjust the weight of KD loss to optimize the training process. Extensive experiments on three datasets support the effectiveness of FedKD.

Method: The study systematically added controlled detection errors (false positives and false negatives) to simulated data to examine how these errors affect ULM image quality. They measured impact using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) metrics.

Result: Both FP and FN rates similarly impact PSNR, but FN rates cause much greater degradation in SSIM (45% drop from 0% to 20% FN) compared to FP rates (7% drop). Dense microbubble regions are more resilient to detection errors, while sparse regions show high sensitivity to errors.

Conclusion: False negatives have a more detrimental effect on ULM image quality than false positives, highlighting the need for robust microbubble detection frameworks that minimize detection errors, particularly in sparse vascular regions, to enhance super-resolution ultrasound imaging quality.

Abstract: Super-resolution ultrasound imaging with ultrasound localization microscopy (ULM) offers a high-resolution view of microvascular structures. Yet, ULM image quality heavily relies on precise microbubble (MB) detection. Despite the crucial role of localization algorithms, there has been limited focus on the practical pitfalls in MB detection tasks such as setting the detection threshold. This study examines how False Positives (FPs) and False Negatives (FNs) affect ULM image quality by systematically adding controlled detection errors to simulated data. Results indicate that while both FP and FN rates impact Peak Signal-to-Noise Ratio (PSNR) similarly, increasing FP rates from 0% to 20% decreases Structural Similarity Index (SSIM) by 7%, whereas same FN rates cause a greater drop of around 45%. Moreover, dense MB regions are more resilient to detection errors, while sparse regions show high sensitivity, showcasing the need for robust MB detection frameworks to enhance super-resolution imaging.

Method: A3 introduces an “essential-state” based procedural evaluation method that uses MLLMs as reward models to progressively verify task completion and process achievement. It includes a benchmark of 100 tasks from 20 widely-used dynamic online apps across 20 Google Play Store categories, plus a toolkit for Android device interaction, environment reset, and data collection.

Result: The paper presents a complete A3 system with benchmark and tools that will be publicly released, providing a robust foundation for future mobile GUI agent research and development. The system addresses limitations of traditional function-based evaluation methods for online dynamic apps.

Conclusion: A3 fills a critical gap in mobile GUI agent evaluation by providing a comprehensive system for assessing agent performance in dynamic, real-world online mobile environments, enabling more realistic and effective development of mobile AI agents.

Abstract: The advancement of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) has catalyzed the development of mobile graphic user interface (GUI) AI agents, which is designed to autonomously perform tasks on mobile devices. However, a significant gap persists in mobile GUI agent evaluation, where existing benchmarks predominantly rely on either static frame assessments such as AndroidControl or offline static apps such as AndroidWorld and thus fail to capture agent performance in dynamic, real-world online mobile apps. To address this gap, we present Android Agent Arena (A3), a novel “essential-state” based procedural evaluation system for mobile GUI agents. A3 introduces a benchmark of 100 tasks derived from 20 widely-used, dynamic online apps across 20 categories from the Google Play Store, ensuring evaluation comprehension. A3 also presents a novel “essential-state” based procedural evaluation method that leverages MLLMs as reward models to progressively verify task completion and process achievement. This evaluation approach address the limitations of traditional function based evaluation methods on online dynamic apps. Furthermore, A3 includes a toolkit to streamline Android device interaction, reset online environment and apps and facilitate data collection from both human and agent demonstrations. The complete A3 system, including the benchmark and tools, will be publicly released to provide a robust foundation for future research and development in mobile GUI agents.

Method: The survey categorizes lifelong learning techniques for LLM agents into three core modules: perception module for multimodal input integration, memory module for storing/retrieving evolving knowledge, and action module for grounded interactions with dynamic environments.

Result: The survey provides a systematic framework for developing lifelong learning capabilities in LLM agents, highlighting how these modules collectively enable continuous adaptation, mitigate catastrophic forgetting, and improve long-term performance.

Conclusion: This first-of-its-kind survey offers a roadmap for researchers and practitioners, providing insights into emerging trends, evaluation metrics, and application scenarios for lifelong learning in LLM agents, with available literature and resources.

Abstract: Lifelong learning, also known as continual or incremental learning, is a crucial component for advancing Artificial General Intelligence (AGI) by enabling systems to continuously adapt in dynamic environments. While large language models (LLMs) have demonstrated impressive capabilities in natural language processing, existing LLM agents are typically designed for static systems and lack the ability to adapt over time in response to new challenges. This survey is the first to systematically summarize the potential techniques for incorporating lifelong learning into LLM-based agents. We categorize the core components of these agents into three modules: the perception module for multimodal input integration, the memory module for storing and retrieving evolving knowledge, and the action module for grounded interactions with the dynamic environment. We highlight how these pillars collectively enable continuous adaptation, mitigate catastrophic forgetting, and improve long-term performance. This survey provides a roadmap for researchers and practitioners working to develop lifelong learning capabilities in LLM agents, offering insights into emerging trends, evaluation metrics, and application scenarios. Relevant literature and resources are available at \href{this url}{https://github.com/qianlima-lab/awesome-lifelong-llm-agent}.

Method: LEMMANAID uses a fine-tuned LLM to generate lemma templates that describe the shape of lemmas, then employs symbolic methods to fill in the details. It combines neural and symbolic approaches in a neuro-symbolic framework.

Result: LEMMANAID outperforms both neural-only and symbolic-only methods: discovers 50% (HOL) and 28% (AFP) of gold standard lemmas (8-13% better than neural-only). With ensembling, performance increases to 55% and 34%. In an Octonions case study, discovers 79% of lemmas vs 62% neural-only and 23% symbolic.

Conclusion: LEMMANAID successfully conjectures significant numbers of interesting lemmas across diverse mathematical domains, demonstrating that neuro-symbolic approaches can effectively automate lemma discovery beyond basic benchmark tasks.

Abstract: Mathematicians and computer scientists are increasingly using proof assistants to formalize and check correctness of complex proofs. This is a non-trivial task in itself, however, with high demands on human expertise. Can we lower the bar by introducing automation for conjecturing helpful, interesting and novel lemmas? We present the first neuro-symbolic lemma conjecturing tool, LEMMANAID, designed to discover conjectures by drawing analogies between mathematical theories. LEMMANAID uses a fine-tuned LLM to generate lemma templates that describe the shape of a lemma, and symbolic methods to fill in the details. We compare LEMMANAID against the same LLM fine-tuned to generate complete lemma statements (a purely neural method), as well as a fully symbolic conjecturing method. LEMMANAID consistently outperforms both neural and symbolic methods on test sets from Isabelle’s HOL library and from its Archive of Formal Proofs (AFP). Using DeepSeek-coder-6.7B as a backend, LEMMANAID discovers 50% (HOL) and 28% (AFP) of the gold standard reference lemmas, 8-13% more than the corresponding neural-only method. Ensembling two LEMMANAID versions with different prompting strategies further increases performance to 55% and 34% respectively. In a case study on the formalization of Octonions, LEMMANAID discovers 79% of the gold standard lemmas, compared to 62% for neural-only and 23% for the state of the art symbolic tool. Our result show that LEMMANAID is able to conjecture a significant number of interesting lemmas across a wide range of domains covering formalizations over complex concepts in both mathematics and computer science, going far beyond the basic concepts of standard benchmarks such as miniF2F, PutnamBench and ProofNet.

Method: Uses frontier models to analyze errors from weaker LMMs, proposes new examples to correct reasoning failures, then filters for quality. Generated a 553k example multimodal instruction tuning dataset.

Result: Models trained on this targeted synthetic data outperform those trained on equivalent amounts of additional real data across multiple downstream tasks.

Conclusion: Targeted synthetic data generation addressing specific reasoning failure modes is highly valuable for improving LMM performance, more effective than generic synthetic or additional real data.

Abstract: Training models on synthetic data has emerged as an increasingly important strategy for improving the performance of generative AI. This approach is particularly helpful for large multimodal models (LMMs) due to the relative scarcity of high-quality paired image-text data compared to language-only data. While a variety of methods have been proposed for generating large multimodal datasets, they do not tailor the synthetic data to address specific deficiencies in the reasoning abilities of LMMs which will be trained with the generated dataset. In contrast, humans often learn in a more efficient manner by seeking out examples related to the types of reasoning where they have failed previously. Inspired by this observation, we propose a new approach for synthetic data generation which is grounded in the analysis of an existing LMM’s reasoning failures. Our methodology leverages frontier models to automatically analyze errors produced by a weaker LMM and propose new examples which can be used to correct the reasoning failure via additional training, which are then further filtered to ensure high quality. We generate a large multimodal instruction tuning dataset containing over 553k examples using our approach and conduct extensive experiments demonstrating its utility for improving the performance of LMMs on multiple downstream tasks. Our results show that models trained on our synthetic data can even exceed the performance of LMMs trained on an equivalent amount of additional real data, demonstrating the high value of generating synthetic data targeted to specific reasoning failure modes in LMMs. We will make our dataset and code publicly available.

Method: Created STARK benchmark with 26 distinct spatiotemporal tasks across three hierarchical levels: state estimation, spatiotemporal reasoning over states, and world-knowledge-aware reasoning. Includes 14,552 challenges with diverse sensor modalities, evaluated 3 LRMs and 8 LLMs using direct answering or Python Code Interpreter.

Result: LLMs show limited success in geometric reasoning tasks (multilateration/triangulation), especially as complexity increases. LRMs perform robustly across difficulty levels, often competing with/surpassing traditional methods. Performance gap narrows on world-knowledge tasks, with some LLMs surpassing LRMs, but LRM o3 maintains leading performance overall.

Conclusion: STARK provides a structured framework to identify spatiotemporal reasoning limitations in LLMs/LRMs, motivating future innovations in model architectures and reasoning paradigms for intelligent CPS. LRM size appears to be a key factor in performance.

Abstract: Spatiotemporal reasoning plays a key role in Cyber-Physical Systems (CPS). Despite advances in Large Language Models (LLMs) and Large Reasoning Models (LRMs), their capacity to reason about complex spatiotemporal signals remains underexplored. This paper proposes a hierarchical SpatioTemporal reAsoning benchmaRK, STARK, to systematically evaluate LLMs across three levels of reasoning complexity: state estimation (e.g., predicting field variables, localizing and tracking events in space and time), spatiotemporal reasoning over states (e.g., inferring spatial-temporal relationships), and world-knowledge-aware reasoning that integrates contextual and domain knowledge (e.g., intent prediction, landmark-aware navigation). We curate 26 distinct spatiotemporal tasks with diverse sensor modalities, comprising 14,552 challenges where models answer directly or by Python Code Interpreter. Evaluating 3 LRMs and 8 LLMs, we find LLMs achieve limited success in tasks requiring geometric reasoning (e.g., multilateration or triangulation), particularly as complexity increases. Surprisingly, LRMs show robust performance across tasks with various levels of difficulty, often competing or surpassing traditional first-principle-based methods. Our results show that in reasoning tasks requiring world knowledge, the performance gap between LLMs and LRMs narrows, with some LLMs even surpassing LRMs. However, the LRM o3 model continues to achieve leading performance across all evaluated tasks, a result attributed primarily to the larger size of the reasoning models. STARK motivates future innovations in model architectures and reasoning paradigms for intelligent CPS by providing a structured framework to identify limitations in the spatiotemporal reasoning of LLMs and LRMs.

Method: Systematic analysis of tree-search LLM planners (depth-first, breadth-first, Monte Carlo Tree Search, bidirectional search) within a unified framework. Uses explicit search trees to expose intermediate decisions, node evaluations, and failure modes, enabling controlled ablations of planner behavior.

Result: Existing tree-based LLM planners often fail to find cost-optimal plans; additional search computation doesn’t reliably improve optimality. Bidirectional search achieves best overall efficiency and success rate. MCTS achieves highest optimality on short-horizon tasks. Tree-search planners are valuable for studying LLM planning due to explicit reasoning steps.

Conclusion: Improving LLM planning under resource constraints requires new search algorithms rather than solely scaling inference-time compute. Tree-search planners provide explicit reasoning that helps understand LLM planning limitations compared to black-box prompting approaches.

Abstract: Planning under resource constraints is central to real-world decision making, yet most large language model (LLM) planners assume uniform action costs. We systematically analyze whether tree-search LLM planners are cost-aware and whether they efficiently generate budget-feasible plans. In contrast to black-box prompting, explicit search trees expose intermediate decisions, node evaluations, and failure modes, which allows for controlled ablations of planner behavior. We study depth-first search, breadth-first search, Monte Carlo Tree Search, and bidirectional search within a unified framework. Our experiments show that existing tree-based LLM planners often struggle to find cost-optimal plans, and that additional search computation does not reliably improve optimality. Among the methods evaluated, bidirectional search achieves the best overall efficiency and success rate. MCTS achieves the highest optimality on short-horizon tasks. Tree-search planners are especially valuable for studying LLM planning because their reasoning steps are explicit, in contrast to plain LLMs that internalize planning dynamics through post-training trajectories. Our findings suggest that improving LLM planning under resource constraints will likely require new search algorithms, rather than solely scaling inference-time compute.

Method: Created FairMedQA benchmark with 4,806 counterfactual question pairs from 801 clinical vignettes. Evaluated 12 representative LLMs to measure accuracy disparities across sensitive demographic groups.

Result: Found substantial accuracy disparities of 3-19 percentage points across demographic groups. FairMedQA exposed biases at least 12 percentage points larger than the CPV benchmark, showing superior sensitivity.

Conclusion: Urgent need for targeted debiasing techniques and identity-aware validation before LLMs can be safely integrated into clinical decision-support systems due to significant demographic biases.

Abstract: Large language models (LLMs) are approaching expert-level performance in medical question answering (QA), demonstrating strong potential to improve public healthcare. However, underlying biases related to sensitive attributes such as sex and race pose life-critical risks. The extent to which such sensitive attributes affect diagnosis remains an open question and requires comprehensive empirical investigation. Additionally, even the latest Counterfactual Patient Variations (CPV) benchmark can hardly distinguish the bias levels of different LLMs. To further explore these dynamics, we propose a new benchmark, FairMedQA, and benchmark 12 representative LLMs. FairMedQA contains 4,806 counterfactual question pairs constructed from 801 clinical vignettes. Our results reveal substantial accuracy disparity ranging from 3 to 19 percentage points across sensitive demographic groups. Notably, FairMedQA exposes biases that are at least 12 percentage points larger than those identified by the latest CPV benchmark, presenting superior benchmarking sensitivity. Our results underscore an urgent need for targeted debiasing techniques and more rigorous, identity-aware validation protocols before LLMs can be safely integrated into practical clinical decision-support systems.

Method: Introduces an agent-based autoscaling framework with four scaling agents: Active Inference, Deep Q Network, Analysis of Structural Knowledge, and Deep Active Inference. Tested on two real-world processing services (YOLOv8 for visual recognition and OpenCV for QR code detection) running in parallel.

Result: All agents achieve acceptable SLO performance with varying convergence patterns. Deep Q Network benefits from pre-training, structural analysis converges quickly, and deep active inference combines theoretical foundations with practical scalability advantages.

Conclusion: Provides evidence for the viability of multi-dimensional agent-based autoscaling for edge environments and encourages future work in this research direction.

Abstract: Edge computing breaks with traditional autoscaling due to strict resource constraints, thus, motivating more flexible scaling behaviors using multiple elasticity dimensions. This work introduces an agent-based autoscaling framework that dynamically adjusts both hardware resources and internal service configurations to maximize requirements fulfillment in constrained environments. We compare four types of scaling agents: Active Inference, Deep Q Network, Analysis of Structural Knowledge, and Deep Active Inference, using two real-world processing services running in parallel: YOLOv8 for visual recognition and OpenCV for QR code detection. Results show all agents achieve acceptable SLO performance with varying convergence patterns. While the Deep Q Network benefits from pre-training, the structural analysis converges quickly, and the deep active inference agent combines theoretical foundations with practical scalability advantages. Our findings provide evidence for the viability of multi-dimensional agent-based autoscaling for edge environments and encourage future work in this research direction.

Method: Introduces Tool-Environment-Agent (TEA) protocol modeling environments, agents, and tools as first-class resources with explicit lifecycles and versioned interfaces. Builds AgentOrchestra framework with central planner orchestrating specialized sub-agents for web navigation, data analysis, and file operations.

Result: AgentOrchestra consistently outperforms strong baselines on three challenging benchmarks, achieving 89.04% on GAIA, establishing state-of-the-art performance.

Conclusion: TEA protocol and hierarchical orchestration improve scalability and generality in multi-agent systems, providing principled foundation for lifecycle management and enabling continual self-evolution through closed feedback loops.

Abstract: Recent advances in LLM-based agent systems have shown promise in tackling complex, long-horizon tasks. However, existing LLM-based agentprotocols (e.g., A2A and MCP) under-specify cross-entity lifecycle and context management, version tracking, and ad-hoc environment integration, which in turn encourages fixed, monolithic agent compositions and brittle glue code. To address these limitations, we introduce the Tool-Environment-Agent (TEA) protocol, a unified abstraction that models environments, agents, and tools as first-class resources with explicit lifecycles and versioned interfaces. TEA provides a principled foundation for end-to-end lifecycle and version management, and for associating each run with its context and outputs across components, improving traceability and reproducibility. Moreover, TEA enables continual self-evolution of agent-associated components through a closed feedback loop, producing improved versions while supporting version selection and rollback. Building on TEA, we present AgentOrchestra, a hierarchical multi-agent framework in which a central planner orchestrates specialized sub-agents for web navigation, data analysis, and file operations, and supports continual adaptation by dynamically instantiating, retrieving, and refining tools online during execution. We evaluate AgentOrchestra on three challenging benchmarks, where it consistently outperforms strong baselines and achieves 89.04% on GAIA, establishing state-of-the-art performance to the best of our knowledge. Overall, our results provide evidence that TEA and hierarchical orchestration improve scalability and generality in multi-agent systems.

Method: Created a novel form-filling benchmark with 432 fields across 55 documents and 3 tasks, requiring knowledge of 236 user features. Developed FieldFinder tool to assist LLMs in identifying text placement locations on forms.

Result: Baseline VLAs achieved <1% accuracy due to poor localization. GUI agents scored 10.6-68.0% with high cost/latency. With FieldFinder, all models achieved equal or better performance, with maximum improvement from 2% to 56%.

Conclusion: FieldFinder effectively addresses the localization challenge in form filling, significantly improving LLM performance on form completion tasks in pure-image domains.

Abstract: Completing paperwork is a challenging and time-consuming problem. Form filling is especially challenging in the pure-image domain without access to OCR, typeset PDF text, or a DOM. For computer agents, it requires multiple abilities, including multi-modal understanding, information retrieval, and tool-use. We present a novel form-filling benchmark consisting of 432 fields spread across 55 documents and 3 tasks, requiring knowledge of 236 features per user. We find that baseline VLAs achieve less than 1% accuracy in most cases, primarily due to poor localization ability. GUI agents also struggle, scoring between 10.6-68.0% despite high cost and latency. Therefore, we also contribute FieldFinder, a tool to assist LLMs in identifying where to place text on a form. With FieldFinder, all models achieve equal or better performance in all six study conditions, with a maximum increase from 2% to 56%.

Method: MoPPS models each prompt’s success rate as a latent variable, performs streaming Bayesian inference, and uses posterior sampling in a multi-armed bandit framework to enable sample-efficient prompt selection without requiring actual LLM evaluations.

Result: Extensive experiments across mathematics, planning, and vision-based geometry tasks show that MoPPS reliably predicts prompt difficulty and accelerates training with significantly reduced LLM rollouts.

Conclusion: MoPPS provides an effective Bayesian risk-predictive framework for prompt selection that reduces computational costs in RL finetuning of LLMs while maintaining performance across diverse reasoning tasks.

Abstract: Recent advances have witnessed the effectiveness of reinforcement learning (RL) finetuning in enhancing the reasoning capabilities of large language models (LLMs). The optimization process often requires numerous iterations to achieve satisfactory performance, resulting in high computational costs due to the need for frequent prompt evaluations under intensive LLM interactions and repeated policy updates. Appropriate online prompt selection methods reduce iteration steps by prioritizing informative prompts during training, while the pipeline’s reliance on exhaustive prompt evaluation and subset selection for optimization still incurs substantial computational overhead due to frequent LLM inference calls. Distinguished from these direct evaluate-then-select schemes, this work investigates iterative approximate evaluation for arbitrary prompts and introduces Model Predictive Prompt Selection (MoPPS), a Bayesian risk-predictive framework that online estimates prompt difficulty without requiring costly LLM interactions. Technically, MoPPS models each prompt’s success rate as a latent variable, performs streaming Bayesian inference, and employs posterior sampling in a constructed multi-armed bandit machine, enabling sample efficient and adaptive prompt selection. Extensive experiments across mathematics, planning, and vision-based geometry tasks show that MoPPS reliably predicts prompt difficulty and accelerates training with significantly reduced LLM rollouts. Our code is available at https://github.com/thu-rllab/MoPPS.

Method: Proposes a data distribution lens hypothesis: CoT reasoning reflects structured inductive bias learned from in-distribution data. Introduces DataAlchemy, a fully controllable environment to train LLMs from scratch and systematically probe them under various distribution conditions across task, length, and format dimensions.

Result: CoT reasoning is effective only when test queries align with training distribution. When pushed beyond training distributions, CoT reasoning fails, revealing it as a “brittle mirage” rather than genuine, generalizable reasoning.

Conclusion: CoT reasoning is not robust general reasoning but rather a learned pattern that breaks under distribution shifts, emphasizing the ongoing challenge of achieving genuine and generalizable reasoning in LLMs.

Abstract: Chain-of-Thought (CoT) prompting has been shown to be effective in eliciting structured reasoning (i.e., CoT reasoning) from large language models (LLMs). Regardless of its popularity, recent studies expose its failures in some reasoning tasks, raising fundamental questions about the nature of CoT reasoning. In this work, we propose a data distribution lens to understand when and why CoT reasoning succeeds or fails. We hypothesize that CoT reasoning reflects a structured inductive bias learned from in-distribution data, enabling models to conditionally generate reasoning trajectories that approximate those observed during training. As such, the effectiveness of CoT reasoning is fundamentally governed by the nature and degree of distribution discrepancy between training data and test queries. Guided by this lens, we dissect CoT reasoning via three dimensions: task, length, and format. To test the hypothesis, we introduce DataAlchemy, an abstract and fully controllable environment that trains LLMs from scratch and systematically probes them under various distribution conditions. Through rigorous controlled experiments, we reveal that CoT reasoning is a brittle mirage when it is pushed beyond training distributions, emphasizing the ongoing challenge of achieving genuine and generalizable reasoning.

Method: Formulates collaboration between cut selection and branching as a Stackelberg game, uses two-phase learning: data-communicated policy pretraining followed by orchestrated policy learning for multiple modules.

Result: Extensive experiments on synthetic and real-world datasets show jointly learned policies significantly improve solving performance and demonstrate excellent generalization across different instance sets.

Conclusion: Collaborative policy optimization through multi-agent learning framework effectively addresses interdependence between MILP solver modules, leading to better performance and generalization.

Abstract: Mixed-integer linear programming (MILP) has been a fundamental problem in combinatorial optimization. Conventional MILP solving mainly relies on carefully designed heuristics embedded in the branch-and-bound framework. Driven by the strong capabilities of neural networks, recent research is exploring the value of machine learning alongside conventional MILP solving. Although learning-based MILP methods have shown great promise, existing works typically learn policies for individual modules in MILP solvers in isolation, without considering their interdependence, which limits both solving efficiency and solution quality. To address this limitation, we propose Collab-Solver, a novel multi-agent-based policy learning framework for MILP that enables collaborative policy optimization for multiple modules. Specifically, we formulate the collaboration between cut selection and branching in MILP solving as a Stackelberg game. Under this formulation, we develop a two-phase learning paradigm to stabilize collaborative policy learning: the first phase performs data-communicated policy pretraining, and the second phase further orchestrates the policy learning for various modules. Extensive experiments on both synthetic and large-scale real-world MILP datasets demonstrate that the jointly learned policies significantly improve solving performance. Moreover, the policies learned by Collab-Solver have also demonstrated excellent generalization abilities across different instance sets.

Method: Proposes Distributed Guided Local Search (DGLS) with three key improvements: 1) adaptive violation condition to selectively penalize high-cost constraints, 2) penalty evaporation mechanism to control penalization magnitude, and 3) synchronization scheme for coordinated penalty updates.

Result: Theoretical analysis shows penalty values are bounded and agents play a potential game in DGLS. Extensive empirical results demonstrate DGLS’s superiority over state-of-the-art baselines, achieving competitive performance on general-valued problems and significant improvements on structured problems.

Conclusion: DGLS effectively addresses GDBA’s limitations through systematic improvements, making it a superior local search framework for DCOPs with both theoretical guarantees and strong empirical performance.

Abstract: Local search is an important class of incomplete algorithms for solving Distributed Constraint Optimization Problems (DCOPs) but it often converges to poor local optima. While Generalized Distributed Breakout Algorithm (GDBA) provides a comprehensive rule set to escape premature convergence, its empirical benefits remain marginal on general-valued problems. In this work, we systematically examine GDBA and identify three factors that potentially lead to its inferior performance, i.e., over-aggressive constraint violation conditions, unbounded penalty accumulation, and uncoordinated penalty updates. To address these issues, we propose Distributed Guided Local Search (DGLS), a novel GLS framework for DCOPs that incorporates an adaptive violation condition to selectively penalize constraints with high cost, a penalty evaporation mechanism to control the magnitude of penalization, and a synchronization scheme for coordinated penalty updates. We theoretically show that the penalty values are bounded, and agents play a potential game in DGLS. Extensive empirical results on various benchmarks demonstrate the great superiority of DGLS over state-of-the-art baselines. Compared to Damped Max-sum with high damping factors, our DGLS achieves competitive performance on general-valued problems, and outperforms by significant margins on structured problems in terms of anytime results.

Method: Introduce a general method for constructing benchmarks and present BRINK (Benchmark for Reasoning under Incomplete Knowledge) to systematically assess KG-RAG methods under knowledge incompleteness conditions.

Result: Current KG-RAG methods show limited reasoning ability under missing knowledge, often rely on internal memorization, and exhibit varying degrees of generalization depending on their design.

Conclusion: There is a need for better benchmarks like BRINK to properly evaluate KG-RAG reasoning capabilities, as current methods struggle with true reasoning under incomplete knowledge.

Abstract: Knowledge Graph-based Retrieval-Augmented Generation (KG-RAG) is an increasingly explored approach for combining the reasoning capabilities of large language models with the structured evidence of knowledge graphs. However, current evaluation practices fall short: existing benchmarks often include questions that can be directly answered using existing triples in KG, making it unclear whether models perform reasoning or simply retrieve answers directly. Moreover, inconsistent evaluation metrics and lenient answer matching criteria further obscure meaningful comparisons. In this work, we introduce a general method for constructing benchmarks and present BRINK (Benchmark for Reasoning under Incomplete Knowledge) to systematically assess KG-RAG methods under knowledge incompleteness. Our empirical results show that current KG-RAG methods have limited reasoning ability under missing knowledge, often rely on internal memorization, and exhibit varying degrees of generalization depending on their design.

Method: Introduces Intrinsic Memory Agents with agent-specific memories that evolve intrinsically with agent outputs, maintaining role-aligned memory that preserves specialized perspectives while focusing on task-relevant information. Uses a generic memory template applicable to new problems without hand-crafted prompts.

Result: Benchmarked on PDDL, FEVER, and ALFWorld datasets, showing state-of-the-art or comparable performance across all three with highest consistency. On complex data pipeline design task, produces higher quality designs across 5 metrics: scalability, reliability, usability, cost-effectiveness, and documentation.

Conclusion: Addressing memory limitations through intrinsic approaches can improve multi-agent LLM system capabilities on structured planning tasks.

Abstract: Multi-agent systems built on Large Language Models (LLMs) show exceptional promise for complex collaborative problem-solving, yet they face fundamental challenges stemming from context window limitations that impair memory consistency, role adherence, and procedural integrity. This paper introduces Intrinsic Memory Agents, a novel framework that addresses these limitations through agent-specific memories that evolve intrinsically with agent outputs. Specifically, our method maintains role-aligned memory that preserves specialized perspectives while focusing on task-relevant information. Our approach utilises a generic memory template applicable to new problems without the need to hand-craft specific memory prompts. We benchmark our approach on the PDDL, FEVER, and ALFWorld datasets, comparing its performance to existing state-of-the-art multi-agentic memory approaches and showing state-of-the-art or comparable performance across all three, with the highest consistency. An additional evaluation is performed on a complex data pipeline design task, and we demonstrate that our approach produces higher quality designs across 5 metrics: scalability, reliability, usability, cost-effectiveness, and documentation, plus additional qualitative evidence of the improvements. Our findings suggest that addressing memory limitations through intrinsic approaches can improve the capabilities of multi-agent LLM systems on structured planning tasks.

Method: The paper analyzes attention behaviors in masked diffusion models, identifying the Attention Floating phenomenon. It examines how attention patterns evolve across denoising steps and layers, revealing a Shallow Structure-Aware, Deep Content-Focused attention mechanism.

Result: MDMs exhibit dynamic, dispersed attention anchors (Attention Floating) that shift across steps and layers, unlike ARMs’ fixed attention sinks. Shallow layers use floating tokens to build global structure, while deeper layers focus on semantic content. This mechanism explains MDMs’ ability to double ARM performance in knowledge-intensive tasks.

Conclusion: The Attention Floating phenomenon provides a mechanistic explanation for MDMs’ strong in-context learning capabilities. This distinctive attention pattern allows MDMs to outperform ARMs in knowledge-intensive tasks, offering insights into why MDMs are narrowing the performance gap with ARMs.

Abstract: Masked diffusion models (MDMs), which leverage bidirectional attention and a denoising process, are narrowing the performance gap with autoregressive models (ARMs). However, their internal attention mechanisms remain under-explored. This paper investigates the attention behaviors in MDMs, revealing the phenomenon of Attention Floating. Unlike ARMs, where attention converges to a fixed sink, MDMs exhibit dynamic, dispersed attention anchors that shift across denoising steps and layers. Further analysis reveals its Shallow Structure-Aware, Deep Content-Focused attention mechanism: shallow layers utilize floating tokens to build a global structural framework, while deeper layers allocate more capability toward capturing semantic content. Empirically, this distinctive attention pattern provides a mechanistic explanation for the strong in-context learning capabilities of MDMs, allowing them to double the performance compared to ARMs in knowledge-intensive tasks. All codes are available at https://github.com/NEUIR/Uncode.

Method: Designed structured prompts using chain-of-thought and agentic reasoning for three SO categories: individual chance-constrained, joint chance-constrained, and two-stage stochastic mixed-integer linear programming models. Introduced a novel soft-scoring metric to evaluate structural quality and partial correctness.

Result: GPT-4-Turbo achieved better partial scores than GPT-3.5 variants except for individual chance-constrained problems. Structured prompts significantly outperformed simple prompting, reducing extra-element generation and improving objective matching, though extra-element generation remained challenging.

Conclusion: With well-engineered prompts and multi-agent collaboration, LLMs can facilitate Stochastic Optimization formulations, paving the way for practical language-driven modeling pipelines for SO problems.

Abstract: This paper presents an integrated systematic study of the performance of large language models (LLMs), specifically ChatGPT, for automatically formulating and solving Stochastic Optimization (SO) problems from natural language descriptions. Focusing on three key categories, individual chance-constrained models, joint chance-constrained models, and two-stage stochastic mixed-integer linear programming models, we design several prompts that guide ChatGPT through structured tasks using chain-of-thought and agentic reasoning. We introduce a novel soft-scoring metric that evaluates the structural quality and partial correctness of generated models, addressing the limitations of canonical and execution-based accuracy metrics. Across a diverse set of SO problems, GPT-4-Turbo achieves better partial scores than GPT-3.5 variants except for individual chance-constrained problems. Structured prompts significantly outperform simple prompting, reducing extra-element generation and improving objective matching, although extra-element generation remains a nontrivial task. Our findings reveal that with well-engineered prompts and multi-agent collaboration, LLMs can facilitate SO formulations, paving the way for intelligent, language-driven modeling pipelines for SO in practice.

Method: The framework combines multi-layered validation pipelines, stack-specific orchestration, and model-agnostic architecture across three reference stacks. It uses systematic validation to ensure application viability and quality.

Result: Evaluation on 30 generation tasks shows comprehensive validation achieves 73.3% viability rate with 30% reaching perfect quality scores. Open-weights models achieve 80.8% of closed-model performance when provided structured environments. Over 3,000 applications have been generated using the framework.

Conclusion: Scaling reliable AI agents requires scaling environments, not just models. The work provides empirical insights and complete reference implementations for production-oriented agent systems through an open-source framework that has gained community adoption.

Abstract: We present app.build (https://github.com/neondatabase/appdotbuild-agent), an open-source framework that improves LLM-based application generation through systematic validation and structured environments. Our approach combines multi-layered validation pipelines, stack-specific orchestration, and model-agnostic architecture, implemented across three reference stacks. Through evaluation on 30 generation tasks, we demonstrate that comprehensive validation achieves 73.3% viability rate with 30% reaching perfect quality scores, while open-weights models achieve 80.8% of closed-model performance when provided structured environments. The open-source framework has been adopted by the community, with over 3,000 applications generated to date. This work demonstrates that scaling reliable AI agents requires scaling environments, not just models – providing empirical insights and complete reference implementations for production-oriented agent systems.

Method: WiA-LLM trains LLMs as explicit language-based world models using natural language to simulate game state evolution. Two-stage training: supervised fine-tuning on human reasoning traces, then reinforcement learning with outcome-based rewards based on predicted vs. actual state alignment.

Result: In Honor of Kings, WiA-LLM achieves 74.2% accuracy in forecasting game-state changes (27% improvement over base model) and demonstrates strategic behavior more closely aligned with expert players than reactive LLMs.

Conclusion: WiA-LLM enhances LLMs’ foresight and expert-like decision-making in complex environments by using explicit language-based world modeling and what-if analysis, showing promise for high-stakes decision-making applications.

Abstract: LLMs struggle with decision-making in high-stakes environments like MOBA games, primarily due to a lack of proactive reasoning and limited understanding of complex game dynamics. To address this, we propose What-if Analysis LLM (WiA-LLM), a framework that trains an LLM as an explicit, language-based world model. Instead of representing the environment in latent vectors, WiA-LLM uses natural language to simulate how the game state evolves over time in response to candidate actions, and provides textual justifications for these predicted outcomes. WiA-LLM is trained in two stages: supervised fine-tuning on human-like reasoning traces, followed by reinforcement learning with outcome-based rewards based on the alignment between predicted and actual future states. In the Honor of Kings (HoK) environment, WiA-LLM attains 74.2% accuracy (27%$\uparrow$ vs. base model) in forecasting game-state changes. In addition, WiA-LLM demonstrate strategic behavior more closely aligned with expert players than purely reactive LLMs, indicating enhanced foresight and expert-like decision-making.

Method: ToolBrain provides a framework supporting reinforcement learning algorithms (GRPO, DPO), supervised learning, custom reward functions on execution traces, automated LLM-as-a-judge reward generation, knowledge distillation, automatic task generation from tool descriptions, seamless tool retrieval, efficient fine-tuning with QLoRA through Unsloth, and quantized inference via bitsandbytes.

Result: Demonstrated through an Email Search Agent case study showing measurable improvements in tool-use skills under realistic workflows while maintaining a simple and extensible codebase.

Conclusion: ToolBrain eases barriers for researchers and practitioners to adapt LLM-based agents to specific domains, providing a publicly available framework for effective tool-use training.

Abstract: Effective tool use is essential for agentic AI, yet training agents to utilize tools remains challenging due to manually designed rewards, limited training data, and poor multi-tool selection, resulting in slow adaptation, wasted computational resources, and suboptimal performance. We introduce ToolBrain, a lightweight and user-friendly framework for training tool use in agentic models with flexible reinforcement learning, thereby easing the barriers for researchers and practitioners to adapt LLM-based agents to specific domains. It supports a wide range of training strategies, including reinforcement learning algorithms such as GRPO and DPO, as well as supervised learning. ToolBrain enables custom reward callables directly on an agent’s execution traces or simply utilizes an automated LLM-as-a-judge system for reward generation. It is packed with useful capabilities, including knowledge distillation from large to small models, automatic task generation from tool descriptions, seamless tool retrieval, efficient fine-tuning pipelines with QLoRA through Unsloth, and quantized inference via bitsandbytes. We demonstrate ToolBrain through an Email Search Agent case study, showing measurable improvements in tool-use skills under a realistic workflow, while keeping the codebase simple and extensible. Our framework is publicly available at https://toolbrain.org/.

Method: Tool-Augmented Policy Optimization (TAPO) adapts Dynamic Sampling Policy Optimization (DAPO) for tool invocation scenarios, enabling models to dynamically interleave complex reasoning with on-demand tool usage. Introduces two datasets: TAPO-easy-60K and TAPO-hard-18K for training and evaluation.

Result: Experiments on Qwen2.5-3B and Qwen2.5-7B models achieve state-of-the-art performance on tasks requiring external knowledge and mathematical computation among comparable parameter methods. TAPO achieves more efficient tool utilization than baselines while preventing excessive calls from reward hacking.

Conclusion: Combining advanced reasoning with tool usage significantly enhances model performance in knowledge-intensive and computationally demanding tasks, demonstrating efficient tool utilization without over-calling.

Abstract: Recent advances in large language models (LLMs) have popularized test-time scaling, where models generate additional reasoning tokens before producing final answers. These approaches have demonstrated significant performance improvements on benchmarks involving mathematical reasoning. However, language models relying solely on direct inference still struggle with tasks demanding up-to-date knowledge or computational tools such as calculators and code interpreters for complex arithmetic operations. To overcome these limitations, we propose Tool-Augmented Policy Optimization (TAPO), a novel reinforcement learning framework that systematically integrates multi-hop reasoning with adaptive tool-calling capabilities. Our approach employs a modified version of Dynamic Sampling Policy Optimization (DAPO), a recently developed RL paradigm, which we adapt specifically for tool invocation scenarios, enabling models to dynamically interleave complex reasoning with on-demand tool usage (including search APIs and Python interpreters). To support this research, we introduce two new datasets: TAPO-easy-60K and TAPO-hard-18K, specifically designed to train and evaluate both fact-based reasoning and mathematical calculation capabilities. Our experiments on Qwen2.5-3B and Qwen2.5-7B models demonstrate the effectiveness of our approach, with both models achieving state-of-the-art performance on tasks requiring external knowledge and mathematical computation among methods with comparable parameters. Notably, TAPO achieves more efficient tool utilization than baseline methods while preventing excessive calls caused by reward hacking. These results highlight the significant potential of combining advanced reasoning with tool usage to enhance model performance in knowledge-intensive and computationally demanding tasks.

Method: A multi-agent framework that actively constructs and evolves dynamic structured knowledge flows to drive subtask execution and reasoning, with strategic planning, parallel exploration, hierarchical decomposition, and real-time adjustment based on feedback.

Result: Achieves competitive performance on general and scientific benchmarks including GAIA, HLE, GPQA and TRQA, demonstrating effectiveness in multi-disciplinary research scenarios.

Conclusion: FlowSearch shows potential to advance scientific discovery by effectively handling complex research tasks through dynamic knowledge flow construction and evolution.

Abstract: Deep research is an inherently challenging task that demands both breadth and depth of thinking. It involves navigating diverse knowledge spaces and reasoning over complex, multi-step dependencies, which presents substantial challenges for agentic systems. To address this, we propose FlowSearch, a multi-agent framework that actively constructs and evolves a dynamic structured knowledge flow to drive subtask execution and reasoning. FlowSearch is capable of strategically planning and expanding the knowledge flow to enable parallel exploration and hierarchical task decomposition, while also adjusting the knowledge flow in real time based on feedback from intermediate reasoning outcomes and insights. FlowSearch achieves competitive performance on both general and scientific benchmarks, including GAIA, HLE, GPQA and TRQA, demonstrating its effectiveness in multi-disciplinary research scenarios and its potential to advance scientific discovery. The code is available at https://github.com/InternScience/InternAgent.

Method: Proposed RIPRAG framework with Reinforcement Learning from Black-box Feedback (RLBF). Treats target RAG system as black box, uses generation model for poisoned documents with two rewards: similarity reward (for document relevance) and attack reward (for successful poisoning).

Result: Method effectively executes poisoning attacks against complex RAG systems, achieving up to 0.72 attack success rate (ASR) improvement compared to baseline methods.

Conclusion: Highlights deficiencies in current RAG defensive methods and provides critical insights for LLM security research, demonstrating vulnerability of RAG systems to sophisticated black-box attacks.

Abstract: Retrieval-Augmented Generation (RAG) systems based on Large Language Models (LLMs) have become a core technology for tasks such as question-answering (QA) and content generation. RAG poisoning is an attack method to induce LLMs to generate the attacker’s expected text by injecting poisoned documents into the database of RAG systems. Existing research can be broadly divided into two classes: white-box methods and black-box methods. White-box methods utilize gradient information to optimize poisoned documents, and black-box methods use a pre-trained LLM to generate them. However, existing white-box methods require knowledge of the RAG system’s internal composition and implementation details, whereas black-box methods are unable to utilize interactive information. In this work, we propose the RIPRAG attack framework, an end-to-end attack pipeline that treats the target RAG system as a black box and leverages our proposed Reinforcement Learning from Black-box Feedback (RLBF) method to optimize the generation model for poisoned documents. We designed two kinds of rewards: similarity reward and attack reward. Experimental results demonstrate that this method can effectively execute poisoning attacks against most complex RAG systems, achieving an attack success rate (ASR) improvement of up to 0.72 compared to baseline methods. This highlights prevalent deficiencies in current defensive methods and provides critical insights for LLM security research.

Method: Memory-as-Action (MemAct) framework formulates context management as learnable policy actions using in-place editing operations (deletion, insertion). Uses Dynamic Context Policy Optimization for efficient training while maintaining reasoning integrity through end-to-end reinforcement learning.

Result: MemAct-RL-14B matches accuracy of models 16× larger while reducing average context length by 51%. Learned strategies adapt to model capabilities and generalize across task complexities.

Conclusion: Treating working memory management as learnable actions enables efficient long-context reasoning, with MemAct demonstrating significant context reduction while maintaining performance through adaptive, generalizable strategies.

Abstract: Long-context Large Language Models, despite their expanded capacity, require careful working memory management to mitigate attention dilution during long-horizon tasks. Yet existing approaches rely on external mechanisms that lack awareness of the agent’s reasoning state, leading to suboptimal decisions. We propose Memory-as-Action (MemAct), a framework that treats working memory management as learnable policy actions. By formulating context management as in-place editing operations (deletion, insertion), MemAct enables joint optimization of information retention and task performance through end-to-end reinforcement learning. To address the computational challenges of dynamic context updates, we introduce Dynamic Context Policy Optimization, which restores training efficiency without compromising reasoning integrity. Experiments show that MemAct-RL-14B matches the accuracy of models $16\times$ larger while reducing average context length by 51%, with learned strategies that adapt to model capabilities and generalize across task complexities.

Method: MASC uses two complementary designs: (1) Next-Execution Reconstruction predicts next step embeddings from query and history to capture causal consistency, and (2) Prototype-Guided Enhancement learns prototype priors over normal-step embeddings to stabilize reconstruction and anomaly scoring under sparse context. When anomalies are detected, a correction agent revises the acting agent’s output before downstream propagation.

Result: On the Who&When benchmark, MASC outperforms all baselines, improving step-level error detection by up to 8.47% AUC-ROC. When integrated into diverse MAS frameworks, it delivers consistent end-to-end gains across architectures.

Conclusion: Metacognitive monitoring and targeted correction can effectively mitigate error propagation in multi-agent systems with minimal overhead, making LLM-based MAS more robust to cascading failures.

Abstract: Large Language Model based multi-agent systems (MAS) excel at collaborative problem solving but remain brittle to cascading errors: a single faulty step can propagate across agents and disrupt the trajectory. In this paper, we present MASC, a metacognitive framework that endows MAS with real-time, unsupervised, step-level error detection and self-correction. MASC rethinks detection as history-conditioned anomaly scoring via two complementary designs: (1) Next-Execution Reconstruction, which predicts the embedding of the next step from the query and interaction history to capture causal consistency, and (2) Prototype-Guided Enhancement, which learns a prototype prior over normal-step embeddings and uses it to stabilize reconstruction and anomaly scoring under sparse context (e.g., early steps). When an anomaly step is flagged, MASC triggers a correction agent to revise the acting agent’s output before information flows downstream. On the Who&When benchmark, MASC consistently outperforms all baselines, improving step-level error detection by up to 8.47% AUC-ROC ; When plugged into diverse MAS frameworks, it delivers consistent end-to-end gains across architectures, confirming that our metacognitive monitoring and targeted correction can mitigate error propagation with minimal overhead.

Method: Proposes NeuroGenPoisoning framework that: 1) Identifies Poison-Responsive Neurons whose activation correlates with contextual poisoning knowledge, 2) Uses genetic algorithm to evolve adversarial passages that maximally activate these neurons, 3) Enables massive-scale generation by reusing promising but initially unsuccessful knowledge variants via attribution signals.

Result: Consistently achieves high Population Overwrite Success Rate (POSR) of over 90% across models and datasets while preserving fluency. Effectively resolves knowledge conflicts between poisoned external knowledge and model’s internal parametric knowledge.

Conclusion: NeuroGenPoisoning demonstrates that neuron-level analysis combined with genetic optimization creates highly effective RAG poisoning attacks, revealing vulnerabilities in current RAG systems and providing insights into knowledge conflict resolution mechanisms.

Abstract: Retrieval-Augmented Generation (RAG) empowers Large Language Models (LLMs) to dynamically integrate external knowledge during inference, improving their factual accuracy and adaptability. However, adversaries can inject poisoned external knowledge to override the model’s internal memory. While existing attacks iteratively manipulate retrieval content or prompt structure of RAG, they largely ignore the model’s internal representation dynamics and neuron-level sensitivities. The underlying mechanism of RAG poisoning has not been fully studied and the effect of knowledge conflict with strong parametric knowledge in RAG is not considered. In this work, we propose NeuroGenPoisoning, a novel attack framework that generates adversarial external knowledge in RAG guided by LLM internal neuron attribution and genetic optimization. Our method first identifies a set of Poison-Responsive Neurons whose activation strongly correlates with contextual poisoning knowledge. We then employ a genetic algorithm to evolve adversarial passages that maximally activate these neurons. Crucially, our framework enables massive-scale generation of effective poisoned RAG knowledge by identifying and reusing promising but initially unsuccessful external knowledge variants via observed attribution signals. At the same time, Poison-Responsive Neurons guided poisoning can effectively resolves knowledge conflict. Experimental results across models and datasets demonstrate consistently achieving high Population Overwrite Success Rate (POSR) of over 90% while preserving fluency. Empirical evidence shows that our method effectively resolves knowledge conflict.

Method: Deep Value Benchmark (DVB) uses controlled confounding between deep values and shallow features in training, then breaks correlations in testing to measure Deep Value Generalization Rate (DVGR).

Result: Average DVGR across 9 models is only 0.30 (below chance), with larger models having slightly lower DVGR than smaller ones, indicating poor value generalization.

Conclusion: Current LLMs fail to learn fundamental human values, relying instead on superficial patterns, highlighting a critical alignment challenge that needs addressing.

Abstract: We introduce the Deep Value Benchmark (DVB), an evaluation framework that directly tests whether large language models (LLMs) learn fundamental human values or merely surface-level preferences. This distinction is critical for AI alignment: Systems that capture deeper values are likely to generalize human intentions robustly, while those that capture only superficial patterns in preference data risk producing misaligned behavior. The DVB uses a novel experimental design with controlled confounding between deep values (e.g., moral principles) and shallow features (e.g., superficial attributes). In the training phase, we expose LLMs to human preference data with deliberately correlated deep and shallow features – for instance, where a user consistently prefers (non-maleficence, formal language) options over (justice, informal language) alternatives. The testing phase then breaks these correlations, presenting choices between (justice, formal language) and (non-maleficence, informal language) options. This design allows us to precisely measure a model’s Deep Value Generalization Rate (DVGR) – the probability of generalizing based on the underlying value rather than the shallow feature. Across 9 different models, the average DVGR is just 0.30. All models generalize deep values less than chance. Larger models have a (slightly) lower DVGR than smaller models. We are releasing our dataset, which was subject to three separate human validation experiments. DVB provides an interpretable measure of a core feature of alignment.

Method: Structured Cognitive Loop (SCL) architecture with five modular phases: Retrieval, Cognition, Control, Action, and Memory (R-CCAM). Core innovation is Soft Symbolic Control - adaptive governance applying symbolic constraints to probabilistic inference while preserving neural flexibility.

Result: Empirical validation shows SCL achieves zero policy violations, eliminates redundant tool calls, and maintains complete decision traceability on multi-step conditional reasoning tasks. Outperforms existing frameworks in reliability and explainability.

Conclusion: SCL provides a practical path toward reliable, explainable, and governable AI agents by connecting expert system principles with modern LLM capabilities. Establishes three design principles for trustworthy agents: modular decomposition, adaptive symbolic governance, and transparent state management.

Abstract: Large language model agents suffer from fundamental architectural problems: entangled reasoning and execution, memory volatility, and uncontrolled action sequences. We introduce Structured Cognitive Loop (SCL), a modular architecture that explicitly separates agent cognition into five phases: Retrieval, Cognition, Control, Action, and Memory (R-CCAM). At the core of SCL is Soft Symbolic Control, an adaptive governance mechanism that applies symbolic constraints to probabilistic inference, preserving neural flexibility while restoring the explainability and controllability of classical symbolic systems. Through empirical validation on multi-step conditional reasoning tasks, we demonstrate that SCL achieves zero policy violations, eliminates redundant tool calls, and maintains complete decision traceability. These results address critical gaps in existing frameworks such as ReAct, AutoGPT, and memory-augmented approaches. Our contributions are threefold: (1) we situate SCL within the taxonomy of hybrid intelligence, differentiating it from prompt-centric and memory-only approaches; (2) we formally define Soft Symbolic Control and contrast it with neuro-symbolic AI; and (3) we derive three design principles for trustworthy agents: modular decomposition, adaptive symbolic governance, and transparent state management. We provide a complete open-source implementation demonstrating the R-CCAM loop architecture, alongside a live GPT-4o-powered travel planning agent. By connecting expert system principles with modern LLM capabilities, this work offers a practical and theoretically grounded path toward reliable, explainable, and governable AI agents.

Method: Extends the Jeffrey-Bolker decision framework to model value refinement, proves a value-of-information theorem for axiological refinement, and analyzes multi-agent settings including game theory transformations.

Result: Established that mutual refinement transforms zero-sum games into positive-sum interactions and yields Pareto-improvements in Nash bargaining. Proved value-of-information theorems for axiological refinement.

Conclusion: A framework of rational choice can be extended to model value refinement, unifying epistemic and axiological refinement under a single formalism, broadening the conceptual foundations of rational choice and illuminating the normative status of ethical deliberation.

Abstract: Standard decision frameworks address uncertainty about facts but assume fixed options and values. We extend the Jeffrey-Bolker framework to model refinements in values and prove a value-of-information theorem for axiological refinement. In multi-agent settings, we establish that mutual refinement will characteristically transform zero-sum games into positive-sum interactions and yield Pareto-improvements in Nash bargaining. These results show that a framework of rational choice can be extended to model value refinement. By unifying epistemic and axiological refinement under a single formalism, we broaden the conceptual foundations of rational choice and illuminate the normative status of ethical deliberation.

Method: Empirical analysis of activation steering across 50 behaviors spanning persona archetypes, personality traits, misalignment behaviors, style cues, and impersonation of public figures. Comprehensive experiments on coefficient optimization, vector properties, and data requirements.

Result: Steering effectiveness varies significantly by behavior type with distinct response patterns. Trait expression follows an inverted-U curve with steering coefficient strength. Vector separation metrics do not predict steering success, but larger training datasets enable more aggressive steering.

Conclusion: Steering effectiveness is heavily influenced by behavior type, providing empirically grounded guidance for implementing activation steering in LLMs.

Abstract: Large language models (LLMs) require precise behavior control for safe and effective deployment across diverse applications. Activation steering offers a promising approach for LLMs’ behavioral control. We focus on the question of how steering effectiveness varies across different behavior types and whether the nature of target behaviors can predict steering success. We address this through empirical analysis of activation steering across 50 behaviors that span persona archetypes, personality traits, misalignment behaviors, style cues, and impersonation of public figures. We present a set of comprehensive experiments on coefficient optimization, vector properties, and data requirements to provide comprehensive guidance for the implementation of activation steering. Our analysis demonstrates that steering effectiveness varies significantly by behavior type, with different behavioral categories exhibiting distinct response patterns to intervention strength. We find that trait expression follows an inverted-U curve with a steering coefficient strength. We also show that vector separation metrics do not predict steering success, but larger training datasets enable more aggressive steering. These findings provide empirically grounded guidance for implementing activation steering and demonstrate that steering effectiveness is heavily influenced by behavior type.

Method: The study examined clinical correlates of pain spikes using a range of AI approaches, including machine learning models and large language models, analyzing data from wearable devices monitoring pain variability and clinical correlates like perceived stress.

Result: Machine learning models achieved relatively high accuracy (>0.7) in predicting pain spikes, while LLMs were limited in providing insights on pain spikes. Real-time monitoring through wearables combined with advanced AI could facilitate early detection and personalized interventions.

Conclusion: The findings highlight the need to develop LLMs that can provide actionable insights in the OUD/CP context, as current LLM performance is limited despite the potential of wearable-AI integration for improving care integration and reducing opioid relapse risk.

Abstract: Chronic pain (CP) and opioid use disorder (OUD) are common and interrelated chronic medical conditions. Currently, there is a paucity of evidence-based integrated treatments for CP and OUD among individuals receiving medication for opioid use disorder (MOUD). Wearable devices have the potential to monitor complex patient information and inform treatment development for persons with OUD and CP, including pain variability (e.g., exacerbations of pain or pain spikes) and clinical correlates (e.g., perceived stress). However, the application of large language models (LLMs) with wearable data for understanding pain spikes, remains unexplored. Consequently, the aim of this pilot study was to examine the clinical correlates of pain spikes using a range of AI approaches. We found that machine learning models achieved relatively high accuracy (>0.7) in predicting pain spikes, while LLMs were limited in providing insights on pain spikes. Real-time monitoring through wearable devices, combined with advanced AI models, could facilitate early detection of pain spikes and support personalized interventions that may help mitigate the risk of opioid relapse, improve adherence to MOUD, and enhance the integration of CP and OUD care. Given overall limited LLM performance, these findings highlight the need to develop LLMs which can provide actionable insights in the OUD/CP context.

Method: Two-stage framework: 1) Fine-tune RoBERTa-large to classify messages as conspiratorial (F1=0.866). 2) Build signed belief graph with message nodes and edge signs reflecting belief alignment, weighted by textual similarity. Introduce Signed Belief Graph Neural Network (SiBeGNN) with Sign Disentanglement Loss to separate ideological alignment from stylistic features.

Result: Identified seven narrative archetypes across 553,648 messages: legal topics, medical concerns, media discussions, finance, contradictions in authority, group moderation, and general chat. SiBeGNN achieved superior clustering quality (cDBI=8.38 vs baselines 13.60-67.27) with 88% inter-rater expert agreement.

Conclusion: Conspiratorial messages appear within routine discussions of finance, law, and everyday matters, not just in skepticism/distrust clusters. This challenges online radicalization assumptions by showing conspiratorial discourse operates within ordinary social interaction. Framework advances belief-driven discourse analysis with applications for stance detection and content moderation.

Abstract: Conspiratorial discourse is increasingly embedded within digital communication ecosystems, yet its structure and spread remain difficult to study. This work analyzes conspiratorial narratives in Singapore-based Telegram groups, showing that such content is woven into everyday discussions rather than confined to isolated echo chambers. We propose a two-stage computational framework. First, we fine-tune RoBERTa-large to classify messages as conspiratorial or not, achieving an F1-score of 0.866 on 2,000 expert-labeled messages. Second, we build a signed belief graph in which nodes represent messages and edge signs reflect alignment in belief labels, weighted by textual similarity. We introduce a Signed Belief Graph Neural Network (SiBeGNN) that uses a Sign Disentanglement Loss to learn embeddings that separate ideological alignment from stylistic features. Using hierarchical clustering on these embeddings, we identify seven narrative archetypes across 553,648 messages: legal topics, medical concerns, media discussions, finance, contradictions in authority, group moderation, and general chat. SiBeGNN yields stronger clustering quality (cDBI = 8.38) than baseline methods (13.60 to 67.27), supported by 88 percent inter-rater agreement in expert evaluations. Our analysis shows that conspiratorial messages appear not only in clusters focused on skepticism or distrust, but also within routine discussions of finance, law, and everyday matters. These findings challenge common assumptions about online radicalization by demonstrating that conspiratorial discourse operates within ordinary social interaction. The proposed framework advances computational methods for belief-driven discourse analysis and offers applications for stance detection, political communication studies, and content moderation policy.

Method: Two-stage training: self-supervised pretraining on 6.03 million sequences from 3,724 USGS stations learns hydrological representations, followed by fine-tuning with synthetic anomalies. Uses hybrid TCN-Transformer architecture (14.2M parameters) to capture local temporal patterns and long-range dependencies, with hierarchical normalization for wide discharge ranges.

Result: Achieves F1 = 0.792 for detection and 68.7% reconstruction-error reduction (36.3% improvement over existing methods). Zero-shot transfer to 100 Canadian stations yields F1 = 0.586, exceeding all baselines and demonstrating cross-national generalization.

Conclusion: HydroGEM provides effective quality control suggestions for streamflow data, designed for human-in-the-loop workflows where outputs require expert review rather than autonomous corrections, enabling scalable data quality management across continental monitoring networks.

Abstract: Real-time streamflow monitoring networks generate millions of observations annually, yet maintaining data quality across thousands of remote sensors remains labor-intensive. We introduce HydroGEM (Hydrological Generalizable Encoder for Monitoring), a foundation model for continental-scale streamflow quality control. HydroGEM uses two-stage training: self-supervised pretraining on 6.03 million sequences from 3,724 USGS stations learns hydrological representations, followed by fine-tuning with synthetic anomalies for detection and reconstruction. A hybrid TCN-Transformer architecture (14.2M parameters) captures local temporal patterns and long-range dependencies, while hierarchical normalization handles six orders of magnitude in discharge. On held-out synthetic tests comprising 799 stations with 18 expert-validated anomaly types, HydroGEM achieves F1 = 0.792 for detection and 68.7% reconstruction-error reduction, a 36.3% improvement over existing methods. Zero-shot transfer to 100 Environment and Climate Change Canada stations yields F1 = 0.586, exceeding all baselines and demonstrating cross-national generalization. The model maintains consistent detection across correction magnitudes and aligns with operational seasonal patterns. HydroGEM is designed for human-in-the-loop workflows - outputs are quality control suggestions requiring expert review, not autonomous corrections.

Method: Systematic evaluation using a unified benchmarking harness comparing CDCL, CP-SAT, and MILP across varying grid sizes and constraint densities. Developed a novel “Warm-Start” hybrid architecture that uses CDCL to generate feasibility hints which are then injected into CP-SAT optimizer.

Result: CDCL demonstrates unrivaled dominance in feasibility detection, solving highly constrained instances orders of magnitude faster than competing paradigms, but struggles with optimization objectives due to cost-blind branching. The Warm-Start hybrid successfully accelerates exact optimization by bridging rapid satisfiability with proven optimality.

Conclusion: CDCL is highly effective for feasibility detection in discrete layout problems, and the proposed Warm-Start hybrid architecture leveraging CDCL’s strengths for feasibility hints followed by CP-SAT optimization provides a practical solution that accelerates exact optimization while maintaining optimality guarantees.

Abstract: Discrete facility layout design involves placing physical entities to minimize handling costs while adhering to strict safety and spatial constraints. This combinatorial problem is typically addressed using Mixed Integer Linear Programming (MILP) or Constraint Programming (CP), though these methods often face scalability challenges as constraint density increases. This study systematically evaluates the potential of Conflict-Driven Clause Learning (CDCL) with VSIDS heuristics as an alternative computational engine for discrete layout problems. Using a unified benchmarking harness, we conducted a controlled comparison of CDCL, CP-SAT, and MILP across varying grid sizes and constraint densities. Experimental results reveal a distinct performance dichotomy: while CDCL struggles with optimization objectives due to cost-blind branching, it demonstrates unrivaled dominance in feasibility detection, solving highly constrained instances orders of magnitude faster than competing paradigms. Leveraging this finding, we developed a novel “Warm-Start” hybrid architecture that utilizes CDCL to rapidly generate valid feasibility hints, which are then injected into a CP-SAT optimizer. Our results confirm that this layered approach successfully accelerates exact optimization, using SAT-driven pruning to bridge the gap between rapid satisfiability and proven optimality.

Method: Proposes Vision-Language Simulation Model (VLSM) that synthesizes executable FlexScript from layout sketches and natural-language prompts, supported by a new dataset of 120,000+ prompt-sketch-code triplets and three novel evaluation metrics (SVR, PMR, ESR).

Result: Models achieve near-perfect structural accuracy and high execution robustness through systematic ablation studies across vision encoders, connectors, and code-pretrained language backbones.

Conclusion: Establishes a foundation for generative digital twins that integrate visual reasoning and language understanding into executable industrial simulation systems.

Abstract: We propose a Vision-Language Simulation Model (VLSM) that unifies visual and textual understanding to synthesize executable FlexScript from layout sketches and natural-language prompts, enabling cross-modal reasoning for industrial simulation systems. To support this new paradigm, the study constructs the first large-scale dataset for generative digital twins, comprising over 120,000 prompt-sketch-code triplets that enable multimodal learning between textual descriptions, spatial structures, and simulation logic. In parallel, three novel evaluation metrics, Structural Validity Rate (SVR), Parameter Match Rate (PMR), and Execution Success Rate (ESR), are proposed specifically for this task to comprehensively evaluate structural integrity, parameter fidelity, and simulator executability. Through systematic ablation across vision encoders, connectors, and code-pretrained language backbones, the proposed models achieve near-perfect structural accuracy and high execution robustness. This work establishes a foundation for generative digital twins that integrate visual reasoning and language understanding into executable industrial simulation systems. Project page: https://danielhsu2014.github.io/GDT-VLSM-project/

Method: Proposed MCPAgentBench with dataset containing authentic tasks and simulated MCP tools. Uses dynamic sandbox environment with candidate tool lists containing distractors to test tool selection and discrimination. Introduces comprehensive metrics for task completion rates and execution efficiency.

Result: Experiments on various latest mainstream LLMs reveal significant performance differences in handling complex, multi-step tool invocations. Code is open-sourced on GitHub.

Conclusion: MCPAgentBench addresses limitations of current MCP evaluation sets and provides a robust framework for assessing LLM agents’ tool-use capabilities, revealing important performance variations across models.

Abstract: Large Language Models (LLMs) are increasingly serving as autonomous agents, and their utilization of external tools via the Model Context Protocol (MCP) is considered a future trend. Current MCP evaluation sets suffer from issues such as reliance on external MCP services and a lack of difficulty awareness. To address these limitations, we propose MCPAgentBench, a benchmark based on real-world MCP definitions designed to evaluate the tool-use capabilities of agents. We construct a dataset containing authentic tasks and simulated MCP tools. The evaluation employs a dynamic sandbox environment that presents agents with candidate tool lists containing distractors, thereby testing their tool selection and discrimination abilities. Furthermore, we introduce comprehensive metrics to measure both task completion rates and execution efficiency. Experiments conducted on various latest mainstream Large Language Models reveal significant performance differences in handling complex, multi-step tool invocations. All code is open-source at Github.

Method: Developed a software package with simulator for multi-modal social interactions and modular architecture that allows exploring different perceptual features, their encoding and fusion, and different agent types.

Result: Created publicly available open-source software (GPL licensed) that has been demonstrated through experimental protocols based on social navigation tasks.

Conclusion: OpenSocInt provides a valuable open-source tool for the research community to study and train social agents, with demonstrated utility in social navigation scenarios and potential for broader applications in social interaction research.

Abstract: In this paper, we introduce OpenSocInt, an open-source software package providing a simulator for multi-modal social interactions and a modular architecture to train social agents. We described the software package and showcased its interest via an experimental protocol based on the task of social navigation. Our framework allows for exploring the use of different perceptual features, their encoding and fusion, as well as the use of different agents. The software is already publicly available under GPL at https://gitlab.inria.fr/robotlearn/OpenSocInt/.

Method: EntroCoT uses an entropy-based mechanism to segment reasoning traces into steps at uncertain junctures, then employs a Monte Carlo rollout-based mechanism to evaluate the marginal contribution of each step, allowing identification and filtering of deceptive reasoning samples.

Result: Extensive experiments on mathematical benchmarks show that fine-tuning on the subset constructed by EntroCoT consistently outperforms baselines using full-dataset supervision.

Conclusion: EntroCoT effectively addresses the problem of low-quality reasoning traces in CoT datasets by providing a systematic approach to identify and filter deceptive samples, resulting in higher-quality training data that improves mathematical reasoning performance.

Abstract: Chain-of-Thought (CoT) prompting has significantly enhanced the mathematical reasoning capabilities of Large Language Models. We find existing fine-tuning datasets frequently suffer from the “answer right but reasoning wrong” probelm, where correct final answers are derived from hallucinated, redundant, or logically invalid intermediate steps. This paper proposes EntroCoT, a unified framework for automatically identifying and refining low-quality CoT supervision traces. EntroCoT first proposes an entropy-based mechanism to segment the reasoning trace into multiple steps at uncertain junctures, and then introduces a Monte Carlo rollout-based mechanism to evaluate the marginal contribution of each step. By accurately filtering deceptive reasoning samples, EntroCoT constructs a high-quality dataset where every intermediate step in each reasoning trace facilitates the final answer. Extensive experiments on mathematical benchmarks demonstrate that fine-tuning on the subset constructed by EntroCoT consistently outperforms the baseslines of full-dataset supervision.

Method: Proposed BackdoorAgent framework structures attacks into three functional stages (planning, memory, tool-use), instruments agent execution, and creates a standardized benchmark across four agent applications (Agent QA, Agent Code, Agent Web, Agent Drive).

Result: Triggers implanted at a single stage can persist across multiple steps and propagate through intermediate states, with GPT-based backbones showing 43.58% persistence in planning attacks, 77.97% in memory attacks, and 60.28% in tool-stage attacks.

Conclusion: Agent workflows are vulnerable to backdoor threats with cross-stage propagation, highlighting the need for systematic analysis frameworks like BackdoorAgent to understand and mitigate these security risks.

Abstract: Large language model (LLM) agents execute tasks through multi-step workflows that combine planning, memory, and tool use. While this design enables autonomy, it also expands the attack surface for backdoor threats. Backdoor triggers injected into specific stages of an agent workflow can persist through multiple intermediate states and adversely influence downstream outputs. However, existing studies remain fragmented and typically analyze individual attack vectors in isolation, leaving the cross-stage interaction and propagation of backdoor triggers poorly understood from an agent-centric perspective. To fill this gap, we propose \textbf{BackdoorAgent}, a modular and stage-aware framework that provides a unified, agent-centric view of backdoor threats in LLM agents. BackdoorAgent structures the attack surface into three functional stages of agentic workflows, including \textbf{planning attacks}, \textbf{memory attacks}, and \textbf{tool-use attacks}, and instruments agent execution to enable systematic analysis of trigger activation and propagation across different stages. Building on this framework, we construct a standardized benchmark spanning four representative agent applications: \textbf{Agent QA}, \textbf{Agent Code}, \textbf{Agent Web}, and \textbf{Agent Drive}, covering both language-only and multimodal settings. Our empirical analysis shows that \textit{triggers implanted at a single stage can persist across multiple steps and propagate through intermediate states.} For instance, when using a GPT-based backbone, we observe trigger persistence in 43.58% of planning attacks, 77.97% of memory attacks, and 60.28% of tool-stage attacks, highlighting the vulnerabilities of the agentic workflow itself to backdoor threats. To facilitate reproducibility and future research, our code and benchmark are publicly available at GitHub.

Method: Introduces SciIF, a multi-discipline benchmark that pairs university-level problems with a fixed catalog of constraints across three pillars: scientific conditions (boundary checks, assumptions), semantic stability (unit/symbol conventions), and specific processes (required numerical methods). Emphasizes auditability by requiring explicit evidence of constraint satisfaction rather than implicit compliance.

Result: The benchmark enables fine-grained diagnosis of compositional reasoning failures by measuring both solution correctness and multi-constraint adherence, ensuring LLMs can function as reliable agents within scientific logical frameworks.

Conclusion: SciIF addresses the gap in evaluating LLMs’ scientific instruction following capability, providing a rigorous standard that incorporates scientific inquiry norms and enables reliable assessment of models’ ability to adhere to scientific validity constraints.

Abstract: As large language models (LLMs) transition from general knowledge retrieval to complex scientific discovery, their evaluation standards must also incorporate the rigorous norms of scientific inquiry. Existing benchmarks exhibit a critical blind spot: general instruction-following metrics focus on superficial formatting, while domain-specific scientific benchmarks assess only final-answer correctness, often rewarding models that arrive at the right result with the wrong reasons. To address this gap, we introduce scientific instruction following: the capability to solve problems while strictly adhering to the constraints that establish scientific validity. Specifically, we introduce SciIF, a multi-discipline benchmark that evaluates this capability by pairing university-level problems with a fixed catalog of constraints across three pillars: scientific conditions (e.g., boundary checks and assumptions), semantic stability (e.g., unit and symbol conventions), and specific processes(e.g., required numerical methods). Uniquely, SciIF emphasizes auditability, requiring models to provide explicit evidence of constraint satisfaction rather than implicit compliance. By measuring both solution correctness and multi-constraint adherence, SciIF enables finegrained diagnosis of compositional reasoning failures, ensuring that LLMs can function as reliable agents within the strict logical frameworks of science.

Method: DR-LoRA dynamically grows expert LoRA ranks during fine-tuning using Expert Saliency Scoring that combines expert routing frequency and LoRA rank importance to quantify each expert’s capacity needs. Higher-saliency experts get priority for rank expansion.

Result: Experiments on multiple benchmarks show DR-LoRA consistently outperforms standard LoRA and static allocation strategies with the same parameter budget, achieving better task performance with more efficient parameter utilization.

Conclusion: Dynamic rank allocation tailored to task-specific expert demands enables superior parameter efficiency and performance for fine-tuning MoE LLMs compared to uniform rank assignment approaches.

Abstract: Mixture-of-Experts (MoE) has become a prominent paradigm for scaling Large Language Models (LLMs). Parameter-efficient fine-tuning (PEFT), such as LoRA, is widely adopted to adapt pretrained MoE LLMs to downstream tasks. However, existing approaches assign identical LoRA ranks to all experts, overlooking the intrinsic functional specialization within MoE LLMs. This uniform allocation leads to resource mismatch, task-relevant experts are under-provisioned while less relevant ones receive redundant parameters. We propose a Dynamic Rank LoRA framework named DR-LoRA, which dynamically grows expert LoRA ranks during fine-tuning based on task-specific demands. DR-LoRA employs an Expert Saliency Scoring mechanism that integrates expert routing frequency and LoRA rank importance to quantify each expert’s demand for additional capacity. Experts with higher saliency scores are prioritized for rank expansion, enabling the automatic formation of a heterogeneous rank distribution tailored to the target task. Experiments on multiple benchmarks demonstrate that DR-LoRA consistently outperforms standard LoRA and static allocation strategies under the same parameter budget, achieving superior task performance with more efficient parameter utilization.

Method: Dual-agent architecture: Agent 01 handles Automatic Speech Recognition (ASR), Agent 02 manages AI processing using local Large Language Models (LLMs), Model Context Protocol (MCP) tools, and Retrieval-Augmented Generation (RAG). System includes RTSP streaming for voice/video, eye tracking data collection, and RabbitMQ messaging for remote task execution.

Result: Successful implementation demonstrating real-time voice command processing with multilingual support and cross-platform task execution capabilities.

Conclusion: The AI glasses system effectively integrates multiple AI components and networking technologies to create a functional wearable computing platform with real-time processing and cross-network capabilities.

Abstract: This paper presents an AI glasses system that integrates real-time voice processing, artificial intelligence(AI) agents, and cross-network streaming capabilities. The system employs dual-agent architecture where Agent 01 handles Automatic Speech Recognition (ASR) and Agent 02 manages AI processing through local Large Language Models (LLMs), Model Context Protocol (MCP) tools, and Retrieval-Augmented Generation (RAG). The system supports real-time RTSP streaming for voice and video data transmission, eye tracking data collection, and remote task execution through RabbitMQ messaging. Implementation demonstrates successful voice command processing with multilingual support and cross-platform task execution capabilities.

Method: 1) Established first benchmark for Coordinate-MLPs in audio signal representations through combinatorial design of 3 positional encodings and 16 activation functions. 2) Proposed Fourier-ASR framework based on Fourier series theorem and Kolmogorov-Arnold representation theorem, introducing Fourier-KAN networks that leverage periodicity and strong nonlinearity without positional encoding. 3) Developed Frequency-adaptive Learning Strategy (FaLS) to enhance convergence by capturing high-frequency components and preventing low-frequency overfitting.

Result: 1) Well-designed positional encoding and activation functions in Coordinate-MLPs can effectively improve audio representation quality. 2) Fourier-ASR can robustly represent complex audio signals without extensive hyperparameter tuning. 3) Extensive experiments on natural speech and music datasets demonstrate the effectiveness of the proposed approach.

Conclusion: The continuity and infinite resolution of implicit audio representations make this research promising for audio compression, synthesis, and generation tasks. The proposed Fourier-ASR framework provides a robust solution for audio signal representation without the hyperparameter tuning limitations of traditional Coordinate-MLPs.

Abstract: Although Coordinate-MLP-based implicit neural representations have excelled in representing radiance fields, 3D shapes, and images, their application to audio signals remains underexplored. To fill this gap, we investigate existing implicit neural representations, from which we extract 3 types of positional encoding and 16 commonly used activation functions. Through combinatorial design, we establish the first benchmark for Coordinate-MLPs in audio signal representations. Our benchmark reveals that Coordinate-MLPs require complex hyperparameter tuning and frequency-dependent initialization, limiting their robustness. To address these issues, we propose Fourier-ASR, a novel framework based on the Fourier series theorem and the Kolmogorov-Arnold representation theorem. Fourier-ASR introduces Fourier Kolmogorov-Arnold Networks (Fourier-KAN), which leverage periodicity and strong nonlinearity to represent audio signals, eliminating the need for additional positional encoding. Furthermore, a Frequency-adaptive Learning Strategy (FaLS) is proposed to enhance the convergence of Fourier-KAN by capturing high-frequency components and preventing overfitting of low-frequency signals. Extensive experiments conducted on natural speech and music datasets reveal that: (1) well-designed positional encoding and activation functions in Coordinate-MLPs can effectively improve audio representation quality; and (2) Fourier-ASR can robustly represent complex audio signals without extensive hyperparameter tuning. Looking ahead, the continuity and infinite resolution of implicit audio representations make our research highly promising for tasks such as audio compression, synthesis, and generation. The source code will be released publicly to ensure reproducibility. The code is available at https://github.com/lif314/Fourier-ASR.

Method: Proposes an objective evaluator based on a Mixture of Experts (MoE) architecture with Sequential Cross-Attention (SeqCoAttn) to assess semantic fidelity between generated audio and input text.

Result: Achieved first rank in the XACLE Challenge with an SRCC of 0.6402, representing a 30.6% improvement over the challenge baseline on the test dataset.

Conclusion: The proposed MOESCORE model provides an effective objective evaluation method for semantic fidelity in Text-to-Audio systems, outperforming existing approaches and addressing the limitations of subjective human evaluation.

Abstract: Recent advances in generative models have enabled modern Text-to-Audio (TTA) systems to synthesize audio with high perceptual quality. However, TTA systems often struggle to maintain semantic consistency with the input text, leading to mismatches in sound events, temporal tructures, or contextual relationships. Evaluating semantic fidelity in TTA remains a significant challenge. Traditional methods primarily rely on subjective human listening tests, which is time-consuming. To solve this, we propose an objective evaluator based on a Mixture of Experts (MoE) architecture with Sequential Cross-Attention (SeqCoAttn). Our model achieves the first rank in the XACLE Challenge, with an SRCC of 0.6402 (an improvement of 30.6% over the challenge baseline) on the test dataset. Code is available at: https://github.com/S-Orion/MOESCORE.

Method: Proposes a learning-based directional SFANC method using convolutional neural network (CNN) to process multiple reference signals, estimate noise source azimuth and elevation angles, and identify the most appropriate control filter for effective noise cancellation.

Result: The proposed approach achieves superior noise reduction with shorter response times compared to traditional adaptive algorithms, even in the presence of reverberations.

Conclusion: The method successfully addresses the gap in considering noise source direction in reverberant environments, improving SFANC performance for real-world applications.

Abstract: Selective fixed-filter active noise control (SFANC) is a novel approach capable of mitigating noise with varying frequency characteristics. It offers faster response and greater computational efficiency compared to traditional adaptive algorithms. However, spatial factors, particularly the influence of the noise source location, are often overlooked. Some existing studies have explored the impact of the direction-of-arrival (DoA) of the noise source on ANC performance, but they are mostly limited to free-field conditions and do not consider the more complex indoor reverberant environments. To address this gap, this paper proposes a learning-based directional SFANC method that incorporates the DoA of the noise source in reverberant environments. In this framework, multiple reference signals are processed by a convolutional neural network (CNN) to estimate the azimuth and elevation angles of the noise source, as well as to identify the most appropriate control filter for effective noise cancellation. Compared to traditional adaptive algorithms, the proposed approach achieves superior noise reduction with shorter response times, even in the presence of reverberations.

Method: Developed Jailbreak-AudioBench with three components: 1) Toolbox for text-to-audio conversion and audio editing with hidden semantics injection, 2) Curated Dataset of diverse explicit and implicit jailbreak audio examples, and 3) Comprehensive Benchmark for evaluating state-of-the-art LALMs.

Result: Established the most comprehensive jailbreak benchmark for audio modality to date, enabling evaluation of multiple state-of-the-art LALMs and exposing previously unexplored audio-specific vulnerabilities.

Conclusion: Jailbreak-AudioBench provides a foundation for advancing LALMs safety research by exposing powerful audio jailbreak threats (like query-based audio editing) and facilitating development of effective defense mechanisms against audio-specific attacks.

Abstract: Large Language Models (LLMs) demonstrate impressive zero-shot performance across a wide range of natural language processing tasks. Integrating various modality encoders further expands their capabilities, giving rise to Multimodal Large Language Models (MLLMs) that process not only text but also visual and auditory modality inputs. However, these advanced capabilities may also pose significant safety problems, as models can be exploited to generate harmful or inappropriate content through jailbreak attacks. While prior work has extensively explored how manipulating textual or visual modality inputs can circumvent safeguards in LLMs and MLLMs, the vulnerability of audio-specific jailbreak on Large Audio-Language Models (LALMs) remains largely underexplored. To address this gap, we introduce Jailbreak-AudioBench, which consists of the Toolbox, curated Dataset, and comprehensive Benchmark. The Toolbox supports not only text-to-audio conversion but also various editing techniques for injecting audio hidden semantics. The curated Dataset provides diverse explicit and implicit jailbreak audio examples in both original and edited forms. Utilizing this dataset, we evaluate multiple state-of-the-art LALMs and establish the most comprehensive Jailbreak benchmark to date for audio modality. Finally, Jailbreak-AudioBench establishes a foundation for advancing future research on LALMs safety alignment by enabling the in-depth exposure of more powerful jailbreak threats, such as query-based audio editing, and by facilitating the development of effective defense mechanisms.

Method: Proposed CompSpoofV2 dataset (over 250k samples, ~283 hours) for component-level audio anti-spoofing, and a separation-enhanced joint learning framework. Also launched the ESDD2 challenge focusing on component-level spoofing detection.

Result: Created a large-scale curated dataset and framework specifically designed for component-level audio deepfake detection, addressing the gap in existing detection systems.

Conclusion: Component-level audio manipulation presents a more challenging detection scenario requiring specialized datasets and approaches. The proposed dataset, framework, and challenge aim to advance research in this area for more realistic deepfake audio detection.

Abstract: Audio recorded in real-world environments often contains a mixture of foreground speech and background environmental sounds. With rapid advances in text-to-speech, voice conversion, and other generation models, either component can now be modified independently. Such component-level manipulations are harder to detect, as the remaining unaltered component can mislead the systems designed for whole deepfake audio, and they often sound more natural to human listeners. To address this gap, we have proposed CompSpoofV2 dataset and a separation-enhanced joint learning framework. CompSpoofV2 is a large-scale curated dataset designed for component-level audio anti-spoofing, which contains over 250k audio samples, with a total duration of approximately 283 hours. Based on the CompSpoofV2 and the separation-enhanced joint learning framework, we launch the Environment-Aware Speech and Sound Deepfake Detection Challenge (ESDD2), focusing on component-level spoofing, where both speech and environmental sounds may be manipulated or synthesized, creating a more challenging and realistic detection scenario. The challenge will be held in conjunction with the IEEE International Conference on Multimedia and Expo 2026 (ICME 2026).

Method: Proposes Signal Embedding Energy (SEE) - a method based on structured activation subspaces from model’s internal representations to quantify noise intensity impact on LALM inputs. Uses this metric to analyze LALM robustness and proposes a mitigation strategy for denoising LALM inputs.

Result: SEE shows strong correlation (0.98) with LALM performance. Surprisingly, traditional audio denoising methods are only marginally effective and sometimes increase SEE and impair performance, indicating mismatch between speech-centric denoising and LALM noise sensitivity. The proposed SEE-based mitigation outperforms existing denoising methods.

Conclusion: SEE provides a novel metric for noise quantification in LALMs, offering guidance for robustness improvements in real-world deployments and revealing limitations of traditional denoising approaches for modern LALMs.

Abstract: Large Audio Language Models (LALMs) have been widely applied in real-time scenarios, such as in-car assistants and online meeting comprehension. In practice, audio inputs are often corrupted by device and environmental noise, leading to performance degradation. However, existing LALM studies on noise lack quantitative analysis and rely mainly on intuition and empirical observation, thus failing to understand practical robustness. To address this issue, we introduce Signal Embedding Energy (SEE), a method for quantifying the impact of noise intensity on LALM inputs, enabling the differentiation of LALM robustness in real-world deployments. SEE introduces a perspective based on structured activation subspaces derived from the model’s internal representations, which more accurately captures its perception of noise than raw audio features. Across experiments, SEE exhibits a strong correlation with LALM performance, achieving a correlation of 0.98. Surprisingly, traditional audio denoising methods are only marginally effective for LALMs, and, in some cases, even increase SEE and impair performance. This suggests a mismatch between speech-centric denoising objectives and the noise sensitivity of modern LALMs. Therefore, we propose a mitigation strategy derived from SEE to denoise LALM inputs, outperforming existing denoising methods. This paper introduces a novel metric for noise quantification in LALMs, providing guidance for robustness improvements in real-world deployments.

Method: Proposes FOCAL framework for benchmarking end-to-end reasoning, component-wise error propagation, and error analysis for multi-modal agents. Introduces two novel metrics: Reasoning and Semantic scores to evaluate voice conversation efficacy.

Result: Framework enables automated and human-assisted testing of voice-to-voice + text input agents, providing systematic evaluation of reasoning capabilities and error propagation through cascading pipelines.

Conclusion: FOCAL addresses the need for comprehensive benchmarking of multi-modal voice agents, offering tools to analyze error propagation and evaluate conversation quality through novel metrics.

Abstract: With the recent advancements in reasoning capa- bilities, tool calling using MCP servers and Audio Language Models (ALMs), development and integration of multi-modal agents (with voice and text support) has come to the industry forefront. Cascading pipelines for voice agents still play a central role in the industry owing to their superior reasoning capabilities facilitated by LLMs. Although, cascading pipelines often present error propagation through the pipeline. We propose a framework, FOCAL to benchmark end-to-end reasoning, component-wise error propagation and error analysis for automated as well as human-assisted testing of multi-modal agents (voice to voice + text input). We also share two novel metrics viz. Reasoning and Semantic scores to evaluate efficacy of the agent in having meaningful conversations in voice mode.

Method: Uses ASR model as noise-robust feature extractor to eliminate non-vocal information. Extracts intermediate layer features from ASR model as emotional speech representation, then applies to downstream NSER task.

Result: 1) Outperforms conventional noise reduction methods; 2) Beats self-supervised learning approaches; 3) Even surpasses text-based approaches using ASR transcription or ground truth transcription of noisy speech.

Conclusion: ASR models serve as effective noise-robust feature extractors for noisy speech emotion recognition, providing superior performance across multiple comparison benchmarks in real-world noisy conditions.

Abstract: This paper proposes an efficient attempt to noisy speech emotion recognition (NSER). Conventional NSER approaches have proven effective in mitigating the impact of artificial noise sources, such as white Gaussian noise, but are limited to non-stationary noises in real-world environments due to their complexity and uncertainty. To overcome this limitation, we introduce a new method for NSER by adopting the automatic speech recognition (ASR) model as a noise-robust feature extractor to eliminate non-vocal information in noisy speech. We first obtain intermediate layer information from the ASR model as a feature representation for emotional speech and then apply this representation for the downstream NSER task. Our experimental results show that 1) the proposed method achieves better NSER performance compared with the conventional noise reduction method, 2) outperforms self-supervised learning approaches, and 3) even outperforms text-based approaches using ASR transcription or the ground truth transcription of noisy speech.

Method: Two main contributions: (1) SoundAtlas dataset with 470k high-quality audio captions using agentic pipeline with Vision-to-Language Compression, Junior-Senior Agent Handoff, and Post-hoc Filtering; (2) Omni2Sound unified diffusion model with three-stage multi-task progressive training to convert competition into joint optimization and mitigate modality bias.

Result: Omni2Sound achieves unified state-of-the-art performance across all three tasks (V2A, T2A, VT2A) within a single model using standard DiT backbone, with strong generalization across benchmarks including challenging off-screen tracks.

Conclusion: The work successfully addresses foundational challenges in unified audio generation through high-quality dataset creation and innovative training strategies, enabling flexible multimodal audio synthesis with superior performance across diverse input conditions.

Abstract: Training a unified model integrating video-to-audio (V2A), text-to-audio (T2A), and joint video-text-to-audio (VT2A) generation offers significant application flexibility, yet faces two unexplored foundational challenges: (1) the scarcity of high-quality audio captions with tight A-V-T alignment, leading to severe semantic conflict between multimodal conditions, and (2) cross-task and intra-task competition, manifesting as an adverse V2A-T2A performance trade-off and modality bias in the VT2A task. First, to address data scarcity, we introduce SoundAtlas, a large-scale dataset (470k pairs) that significantly outperforms existing benchmarks and even human experts in quality. Powered by a novel agentic pipeline, it integrates Vision-to-Language Compression to mitigate visual bias of MLLMs, a Junior-Senior Agent Handoff for a 5 times cost reduction, and rigorous Post-hoc Filtering to ensure fidelity. Consequently, SoundAtlas delivers semantically rich and temporally detailed captions with tight V-A-T alignment. Second, we propose Omni2Sound, a unified VT2A diffusion model supporting flexible input modalities. To resolve the inherent cross-task and intra-task competition, we design a three-stage multi-task progressive training schedule that converts cross-task competition into joint optimization and mitigates modality bias in the VT2A task, maintaining both audio-visual alignment and off-screen audio generation faithfulness. Finally, we construct VGGSound-Omni, a comprehensive benchmark for unified evaluation, including challenging off-screen tracks. With a standard DiT backbone, Omni2Sound achieves unified SOTA performance across all three tasks within a single model, demonstrating strong generalization across benchmarks with heterogeneous input conditions. The project page is at https://swapforward.github.io/Omni2Sound.

Method: SIGNL transforms audio visual representations (spectrograms) into spectral and temporal graphs, uses graph convolutional encoders to learn complementary frequency-time features, pre-trains with non-contrastive self-supervised learning on augmented graph pairs, then fine-tunes on minimal labeled data.

Result: Achieves 7.88% EER on ASVspoof 2021 DF, 3.95% EER on ASVspoof 5 using only 5% labeled data, and generalizes well with 10.16% EER on In-The-Wild dataset when trained on CFAD.

Conclusion: SIGNL provides an effective label-efficient solution for audio deepfake detection by bridging the gap between graph non-contrastive learning and audio’s dual-view structure, achieving strong performance with minimal labeled data and good generalization.

Abstract: Audio deepfake detection is increasingly important as synthetic speech becomes more realistic and accessible. Recent methods, including those using graph neural networks (GNNs) to model frequency and temporal dependencies, show strong potential but need large amounts of labeled data, which limits their practical use. Label-efficient alternatives like graph-based non-contrastive learning offer a potential solution, as they can learn useful representations from unlabeled data without using negative samples. However, current graph non-contrastive approaches are built for single-view graph representations and cannot be directly used for audio, which has unique spectral and temporal structures. Bridging this gap requires dual-view graph modeling suited to audio signals. In this work, we introduce SIGNL (Spectral-temporal vIsion Graph Non-contrastive Learning), a label-efficient expert system for detecting audio deepfakes. SIGNL operates on the visual representation of audio, such as spectrograms or other time-frequency encodings, transforming them into spectral and temporal graphs for structured feature extraction. It then employs graph convolutional encoders to learn complementary frequency-time features, effectively capturing the unique characteristics of audio. These encoders are pre-trained using a non-contrastive self-supervised learning strategy on augmented graph pairs, enabling effective representation learning without labeled data. The resulting encoders are then fine-tuned on minimal labelled data for downstream deepfake detection. SIGNL achieves strong performance on multiple audio deepfake detection benchmarks, including 7.88% EER on ASVspoof 2021 DF and 3.95% EER on ASVspoof 5 using only 5% labeled data. It also generalizes well to unseen conditions, reaching 10.16% EER on the In-The-Wild dataset when trained on CFAD.

Method: Proposed CoM2S approach uses synthetic EMG data generated by pre-trained model, then applies phoneme-level confidence filtering to enhance ETS model through self-training techniques. Also created Libri-EMG dataset - open-access, time-aligned, multi-speaker voiced EMG and speech recordings.

Result: Method improves phoneme accuracy, reduces phonological confusion, and lowers word error rate, confirming effectiveness of CoM2S approach for V-ETS.

Conclusion: CoM2S approach successfully addresses data scarcity in V-ETS, with improved performance metrics. The release of Libri-EMG dataset and codes will support future research in this field.

Abstract: Voiced Electromyography (EMG)-to-Speech (V-ETS) models reconstruct speech from muscle activity signals, facilitating applications such as neurolaryngologic diagnostics. Despite its potential, the advancement of V-ETS is hindered by a scarcity of paired EMG-speech data. To address this, we propose a novel Confidence-based Multi-Speaker Self-training (CoM2S) approach, along with a newly curated Libri-EMG dataset. This approach leverages synthetic EMG data generated by a pre-trained model, followed by a proposed filtering mechanism based on phoneme-level confidence to enhance the ETS model through the proposed self-training techniques. Experiments demonstrate our method improves phoneme accuracy, reduces phonological confusion, and lowers word error rate, confirming the effectiveness of our CoM2S approach for V-ETS. In support of future research, we will release the codes and the proposed Libri-EMG dataset-an open-access, time-aligned, multi-speaker voiced EMG and speech recordings.

Method: Two-fold approach: 1) Created AHEAD-DS by repackaging and refining several open-source datasets with standardized, hearing-aid-relevant labels; 2) Developed OpenYAMNet, a sound recognition model optimized for edge device deployment on smartphones connected to hearing devices.

Result: OpenYAMNet achieved 0.86 mean average precision and 0.93 accuracy on AHEAD-DS testing set across 14 auditory scene categories. Real-time deployment on Android smartphone (Google Pixel 3) showed ~50ms latency for model loading plus ~30ms per second of audio processing.

Conclusion: The work provides a standardized dataset (AHEAD-DS) and baseline model (OpenYAMNet) for auditory scene recognition in hearing devices, enabling systematic comparison and demonstrating practical real-time performance on edge devices.

Abstract: Scene recognition is important for hearing devices, however; this is challenging, in part because of the limitations of existing datasets. Datasets often lack public accessibility, completeness, or audiologically relevant labels, hindering systematic comparison of machine learning models. Deploying such models on resource-constrained edge devices presents another challenge.The proposed solution is two-fold, a repack and refinement of several open source datasets to create AHEAD-DS, a dataset designed for auditory scene recognition for hearing devices, and introduce OpenYAMNet, a sound recognition model. AHEAD-DS aims to provide a standardised, publicly available dataset with consistent labels relevant to hearing aids, facilitating model comparison. OpenYAMNet is designed for deployment on edge devices like smartphones connected to hearing devices, such as hearing aids and wireless earphones with hearing aid functionality, serving as a baseline model for sound-based scene recognition. OpenYAMNet achieved a mean average precision of 0.86 and accuracy of 0.93 on the testing set of AHEAD-DS across fourteen categories relevant to auditory scene recognition. Real-time sound-based scene recognition capabilities were demonstrated on edge devices by deploying OpenYAMNet to an Android smartphone. Even with a 2018 Google Pixel 3, a phone with modest specifications, the model processes audio with approximately 50ms of latency to load the model, and an approximate linear increase of 30ms per 1 second of audio. The project website with links to code, data, and models. \href{https://github.com/Australian-Future-Hearing-Initiative}{https://github.com/Australian-Future-Hearing-Initiative}

Method: 1) Systematically extend task definitions to cover comprehensive editing operations (addition, replacement, removal, reordering, attribute modification). 2) Design scalable data synthesis pipeline to construct large-scale paired datasets with fine-grained event-level annotations. 3) Integrate Qwen2-Audio encoder with MMDiT-based generator for precise cross-modal alignment and localized editing.

Result: Experimental results demonstrate superior editing localization accuracy, robust instruction following, and high fidelity in non-edited regions compared to existing methods.

Conclusion: MMEdit addresses key challenges in text-guided audio editing through unified framework design, comprehensive task coverage, scalable data synthesis, and cross-modal architecture integration, achieving state-of-the-art performance.

Abstract: Text-guided audio editing aims to modify specific acoustic events while strictly preserving non-target content. Despite recent progress, existing approaches remain fundamentally limited. Training-free methods often suffer from signal degradation caused by diffusion inversion, while training-based methods, although achieving higher generation quality, are severely constrained by the scarcity of high-quality paired data and task formulations that cover only a narrow subset of editing operations. In addition, standard architectures typically decouple text and audio processing, limiting the ability to align instructions with specific acoustic contexts. To address these challenges, we propose MMEdit, an audio-language-model-driven framework for unified audio editing. We systematically extend task definitions to cover a comprehensive range of editing operations, including addition, replacement, removal, reordering, and attribute modification. Furthermore, we design a scalable data synthesis pipeline to construct large-scale paired datasets with fine-grained event-level annotations. To capture complex editing semantics, we integrate a Qwen2-Audio encoder with an MMDiT-based generator, enabling precise cross-modal alignment and localized editing. Experimental results demonstrate that our method achieves superior editing localization accuracy, robust instruction following, and high fidelity in non-edited regions.

Method: Release fully open-source system with: 1) 116k licensed synthetic songs dataset (auto-generated lyrics + SunoV5 audio), 2) Muse model trained via single-stage supervised finetuning of Qwen-based LM extended with MuCodec audio tokens, without task-specific losses or complex architecture.

Result: Muse achieves competitive performance on phoneme error rate, text-music style similarity, and audio aesthetic quality despite modest data scale and model size. Enables controllable segment-level generation across different musical structures.

Conclusion: Open-sourcing all data, model weights, and pipelines enables reproducible research and continued progress in controllable long-form song generation, addressing the reproducibility gap in academic research.

Abstract: Recent commercial systems such as Suno demonstrate strong capabilities in long-form song generation, while academic research remains largely non-reproducible due to the lack of publicly available training data, hindering fair comparison and progress. To this end, we release a fully open-source system for long-form song generation with fine-grained style conditioning, including a licensed synthetic dataset, training and evaluation pipelines, and Muse, an easy-to-deploy song generation model. The dataset consists of 116k fully licensed synthetic songs with automatically generated lyrics and style descriptions paired with audio synthesized by SunoV5. We train Muse via single-stage supervised finetuning of a Qwen-based language model extended with discrete audio tokens using MuCodec, without task-specific losses, auxiliary objectives, or additional architectural components. Our evaluations find that although Muse is trained with a modest data scale and model size, it achieves competitive performance on phoneme error rate, text–music style similarity, and audio aesthetic quality, while enabling controllable segment-level generation across different musical structures. All data, model weights, and training and evaluation pipelines will be publicly released, paving the way for continued progress in controllable long-form song generation research. The project repository is available at https://github.com/yuhui1038/Muse.

Method: 1) Discover isomorphism between RUM consistency and flow conservation on Boolean lattice; 2) Develop Tree-Preconditioned Conjugate Gradient solver exploiting spanning tree structure to whiten ill-conditioned Hessian; 3) Formulate projection as differentiable layer using Implicit Function Theorem with exact Jacobian propagation.

Result: Achieves superlinear convergence, scales to previously unsolvable problem sizes, eliminates structural overfitting, enables joint training of provably rational models, and generalizes from sparse data where standard approximations fail.

Conclusion: The “Axioms-as-Layers” paradigm successfully embeds axiomatic choice theory directly into deep learning architectures, creating models that are both trainable and provably rational while overcoming computational barriers of traditional methods.

Abstract: This paper introduces a differentiable framework that embeds the axiomatic structure of Random Utility Models (RUM) directly into deep neural networks. Although projecting empirical choice data onto the RUM polytope is NP-hard in general, we uncover an isomorphism between RUM consistency and flow conservation on the Boolean lattice. Leveraging this combinatorial structure, we derive a novel Tree-Preconditioned Conjugate Gradient solver. By exploiting the spanning tree of the constraint graph, our preconditioner effectively “whitens” the ill-conditioned Hessian spectrum induced by the Interior Point Method barrier, achieving superlinear convergence and scaling to problem sizes previously deemed unsolvable. We further formulate the projection as a differentiable layer via the Implicit Function Theorem, where the exact Jacobian propagates geometric constraints during backpropagation. Empirical results demonstrate that this “Axioms-as-Layers” paradigm eliminates the structural overfitting inherent in penalty-based methods, enabling models that are jointly trainable, provably rational, and capable of generalizing from sparse data regimes where standard approximations fail.

Method: Uses a unified LLM-based architecture with text-guided adaptive graph convolutional networks to merge high-level semantic information with traffic network structure, enabling simultaneous traffic prediction and text generation.

Result: Outperforms state-of-the-art methods on BJTT dataset in both traffic forecasting accuracy and text generation quality, demonstrating improved prediction through generative textual context.

Conclusion: CrossTrafficLLM provides a more interpretable and actionable approach to generative traffic intelligence by unifying prediction and description generation, offering significant advantages for modern ITS applications.

Abstract: While accurate traffic forecasting is vital for Intelligent Transportation Systems (ITS), effectively communicating predicted conditions via natural language for human-centric decision support remains a challenge and is often handled separately. To address this, we propose CrossTrafficLLM, a novel GenAI-driven framework that simultaneously predicts future spatiotemporal traffic states and generates corresponding natural language descriptions, specifically targeting conditional abnormal event summaries. We tackle the core challenge of aligning quantitative traffic data with qualitative textual semantics by leveraging Large Language Models (LLMs) within a unified architecture. This design allows generative textual context to improve prediction accuracy while ensuring generated reports are directly informed by the forecast. Technically, a text-guided adaptive graph convolutional network is employed to effectively merge high-level semantic information with the traffic network structure. Evaluated on the BJTT dataset, CrossTrafficLLM demonstrably surpasses state-of-the-art methods in both traffic forecasting performance and text generation quality. By unifying prediction and description generation, CrossTrafficLLM delivers a more interpretable, and actionable approach to generative traffic intelligence, offering significant advantages for modern ITS applications.

Method: Developed a chunked FFT convolution approach using chunking and overlap-add reconstruction to enable 450K length sequence by 450K length filter convolutions on Alveo U200 FPGA with only 2.8 MB BRAM.

Result: Throughput scales proportionally with chunk size with only 7% degradation for longest sequences, demonstrating that careful memory management enables deployment of long-context primitives on edge FPGAs without sacrificing performance.

Conclusion: Memory-optimized chunked FFT convolution enables efficient long-context neural architectures on resource-constrained edge FPGAs, bridging the gap between computational requirements and hardware limitations.

Abstract: The need for long-context reasoning has led to alternative neural network architectures besides Transformers and self-attention, a popular model being Hyena, which employs causal 1D-convolutions implemented with FFTs. Long convolutions enable efficient global context mixing, but requirements for intermediate results exceed the 2-3 MB Block RAM capacity of FPGAs. We present a chunked FFT convolution approach enabling 450K length sequence by 450K length filter convolutions on an Alveo U200 FPGA with 2.8 MB BRAM through chunking and overlap-add reconstruction. We find that throughput scales proportionally with chunk size while degrading minimally by 7% for our longest sequences, demonstrating that careful memory management enables deployment of long-context primitives on edge FPGAs without sacrificing performance.

Method: Computes Hessian-inverse-vector products without storing Hessian or its inverse. Scales linearly in number of layers, similar to Pearlmutter’s algorithm for Hessian-vector products.

Result: Achieves linear scaling in time and storage with respect to number of layers, avoiding quadratic storage and cubic operations of naive approach.

Conclusion: Provides efficient exact algorithm for Hessian-inverse computations in deep networks, enabling practical applications that require solving linear systems with Hessian matrices.

Abstract: We describe an exact algorithm for solving linear systems $Hx=b$ where $H$ is the Hessian of a deep net. The method computes Hessian-inverse-vector products without storing the Hessian or its inverse in time and storage that scale linearly in the number of layers. Compared to the naive approach of first computing the Hessian, then solving the linear system, which takes storage that’s quadratic in the number of parameters and cubically many operations, our Hessian-inverse-vector product method scales roughly like Pearlmutter’s algorithm for computing Hessian-vector products.

Method: Formulate task/context-specific learning as sequential inference of low-dimensional latent adaptation state using linearized state-space model with Gaussian assumptions, leading to Kalman filter recursion with closed-form updates for posterior mean and covariance.

Result: Shows inference-time learning is driven by covariance collapse (rapid posterior uncertainty contraction), establishes filter stability via observability conditions, proves exponential covariance contraction rates, and derives mean-square error bounds. Gradient-based methods emerge as singular limits of Bayesian inference.

Conclusion: Provides unified probabilistic theory for in-context learning, parameter-efficient adaptation, and test-time learning with explicit stability/sample efficiency guarantees, principled prompt informativeness interpretation, and uncertainty dynamics absent in existing accounts.

Abstract: We present a theory-first framework that interprets inference-time adaptation in large language models (LLMs) as online Bayesian state estimation. Rather than modeling rapid adaptation as implicit optimization or meta-learning, we formulate task- and context-specific learning as the sequential inference of a low-dimensional latent adaptation state governed by a linearized state-space model. Under Gaussian assumptions, adaptation follows a Kalman recursion with closed-form updates for both the posterior mean and covariance. This perspective elevates epistemic uncertainty to an explicit dynamical variable. We show that inference-time learning is driven by covariance collapse, i.e., rapid contraction of posterior uncertainty induced by informative tokens, which typically precedes convergence of the posterior mean. Using observability conditions on token-level Jacobians, we establish stability of the Bayesian filter, prove exponential covariance contraction rates, and derive mean-square error bounds. Gradient descent, natural-gradient methods, and meta-learning updates arise as singular, noise-free limits of the filtering dynamics, positioning optimization-based adaptation as a degenerate approximation of Bayesian inference. The resulting theory provides a unified probabilistic account of in-context learning, parameter-efficient adaptation, and test-time learning without parameter updates. It yields explicit guarantees on stability and sample efficiency, offers a principled interpretation of prompt informativeness via information accumulation, and clarifies the role of uncertainty dynamics absent from existing accounts. Minimal illustrative experiments corroborate the qualitative predictions of the theory.

Method: Inject 5 copies of GSM8K/MBPP test items into first 2B tokens of 25B token pretraining dataset using clean Qwen2.5 and Gemma3 checkpoints. Compare contaminated vs clean models after pretraining and after SFT/GRPO post-training without contamination.

Result: 1) Contamination causes performance spikes that diminish with continued pretraining (close to zero after 25B tokens). 2) SFT and GRPO resurface leaked info differently: SFT inflates only contaminated tasks, GRPO also inflates uncontaminated counterparts. 3) Scale amplifies effects: larger SFT models memorize more, larger GRPO models translate leakage into more generalizable capabilities.

Conclusion: Contamination audits needed after post-training; RL-based post-training can help alleviate contamination-related over-estimation problems despite not being immune to contamination effects.

Abstract: We present a controlled study of how dataset contamination interacts with the post-training stages now standard in large language model training pipelines. Starting from clean checkpoints of Qwen2.5 (0.5B/1.5B) and Gemma3 (1B/4B), we inject five copies of GSM8K and MBPP test items into the first 2B tokens of an otherwise 25B token extended pre-training dataset. We then compare the contaminated and clean models both immediately after pre-training and again after two popular post-training methods: supervised fine-tuning (SFT) and reinforcement learning (RL) with group relative policy optimization (GRPO). The applied post-training steps do not have any contamination. Across math and coding benchmarks, we find three consistent patterns: (i) Contamination causes performance spikes that are gradually diminished with continued pre-training. After even 25B tokens the apparent performance inflation of contamination can become close to zero. (ii) Both SFT and GRPO resurface the leaked information, but with different external validity: SFT inflates scores only on the contaminated tasks, whereas GRPO also inflates performance on uncontaminated counterparts (GSMPlus, HumanEval). (iii) Model scale amplifies these tendencies, larger Supervised Fine Tuned models memorize more, while larger GRPO models translate leakage into more generalizable capabilities. Our results underscore the need for contamination audits \emph{after} post-training and suggest that RL-based post-training, although not immune, can help alleviate contamination-related over-estimation problems.

Method: Proposes Time-RA task reformulation from discriminative to generative reasoning paradigm, creates RATs40K benchmark with raw time series, textual context, visual plots, and structured reasoning annotations across 10 domains.

Result: Supervised fine-tuning and visual representations boost diagnostic accuracy and reasoning consistency, with fine-tuned models showing strong “plug-and-play” transferability outperforming traditional baselines on unseen datasets.

Conclusion: Establishes foundation for interpretable, multimodal time series analysis, with open-sourced code and dataset to facilitate future research in reasoning-intensive anomaly detection.

Abstract: Time series anomaly detection (TSAD) has traditionally focused on binary classification and often lacks the fine-grained categorization and explanatory reasoning required for transparent decision-making. To address these limitations, we propose Time-series Reasoning for Anomaly (Time-RA), a novel task that reformulates TSAD from a discriminative into a generative, reasoning-intensive paradigm. To facilitate this, we introduce RATs40K, the first real-world large-scale multimodal benchmark with ~40,000 samples across 10 domains, integrating raw time series, textual context, and visual plots with structured reasoning annotations. Extensive benchmarking shows that while supervised fine-tuning and visual representations boost diagnostic accuracy and reasoning consistency, performance varies across complex scenarios. Notably, fine-tuned models demonstrate strong “plug-and-play” transferability, outperforming traditional baselines on unseen real-world datasets. Our work establishes a foundation for interpretable, multimodal time series analysis. All code (https://github.com/yyysjz1997/Time-RA) and the RATs40K dataset (https://huggingface.co/datasets/Time-RA/RATs40K) are fully open-sourced to facilitate future research.

Method: Integrated historical fire events (NASA FIRMS 2015-2023), daily meteorological observations (Meteostat), and vegetation indices (NDVI from Google Earth Engine). Harmonized datasets using spatial and temporal joins, then evaluated multiple ML models: Random Forest, XGBoost, LightGBM, MLP, and an ensemble classifier under binary classification (low vs high fire risk).

Result: The ensemble approach achieved 87% accuracy in distinguishing between low and high fire risk zones. This demonstrates reliable bushfire intensity prediction capability.

Conclusion: Combining multi-source environmental features with advanced machine learning techniques can produce reliable bushfire intensity predictions, supporting more informed and timely disaster management in Australia.

Abstract: Bushfires are among the most destructive natural hazards in Australia, causing significant ecological, economic, and social damage. Accurate prediction of bushfire intensity is therefore essential for effective disaster preparedness and response. This study examines the predictive capability of spatio-temporal environmental data for identifying high-risk bushfire zones across Australia. We integrated historical fire events from NASA FIRMS, daily meteorological observations from Meteostat, and vegetation indices such as the Normalized Difference Vegetation Index (NDVI) from Google Earth Engine for the period 2015-2023. After harmonizing the datasets using spatial and temporal joins, we evaluated several machine learning models, including Random Forest, XGBoost, LightGBM, a Multi-Layer Perceptron (MLP), and an ensemble classifier. Under a binary classification framework distinguishing ’low’ and ‘high’ fire risk, the ensemble approach achieved an accuracy of 87%. The results demonstrate that combining multi-source environmental features with advanced machine learning techniques can produce reliable bushfire intensity predictions, supporting more informed and timely disaster management.

Method: Developed a dedicated multimodal Judge Model that aggregates multimodal judgments, analyzes quality and reasoning consistency, and generates diagnostic feedback. Built benchmark from carefully sampled public datasets with fixed seeds across text, audio, image, and video modalities. Evaluated on 280 multimodal samples comparing judge model assessments with human annotators.

Result: Strong alignment between Judge Model and human scores across multiple MLLMs (Gemini 2.5, Phi 4, Qwen 2.5), demonstrating the model’s reliability and potential as a scalable evaluation tool.

Conclusion: The multimodal Judge Model provides a scalable, interpretable evaluation pipeline for future multimodal AI research, offering reliable assessment with diagnostic feedback capabilities that align well with human judgment.

Abstract: We propose a dedicated multimodal Judge Model designed to provide reliable, explainable evaluation across a diverse suite of tasks. Our benchmark spans text, audio, image, and video modalities, drawing from carefully sampled public datasets with fixed seeds to ensure reproducibility and minimize train test leakage. Instead of simple scoring, our framework aggregates multimodal judgments, analyzes the quality and reasoning consistency of model outputs, and generates diagnostic feedback. We evaluate several MLLMs, including Gemini 2.5, Phi 4, and Qwen 2.5, across 280 multimodal samples and compare judge model assessments with human annotators. Results show strong alignment between the Judge Model and human scores, demonstrating its potential as a scalable, interpretable evaluation pipeline for future multimodal AI research.

Method: Proposes flow-matching generative decoder under land-then-transport paradigm that integrates physical wireless channel into continuous-time probability flow. For AWGN channels: builds Gaussian smoothing path with noise schedule indexing effective noise levels, derives closed-form teacher velocity field, trains neural network student vector field via conditional flow matching. For Rayleigh/MIMO: maps to AWGN-equivalent channels via linear MMSE equalization and singular-value-domain processing, reusing same probability path without retraining.

Result: Experiments on MNIST, Fashion-MNIST, and DIV2K datasets show consistent gains over JPEG2000+LDPC, DeepJSCC, and diffusion-based baselines across AWGN, Rayleigh, and MIMO channels. Achieves good perceptual quality with only few ODE steps, providing deterministic, physically interpretable, computation-efficient decoding.

Conclusion: LTT framework provides deterministic, physically interpretable, and computation-efficient generative wireless image decoding across diverse channels, solving latency issues of diffusion methods while maintaining perceptual quality.

Abstract: Due to strict rate and reliability demands, wireless image transmission remains difficult for both classical layered designs and joint source-channel coding (JSCC), especially under low latency. Diffusion-based generative decoders can deliver strong perceptual quality by leveraging learned image priors, but iterative stochastic denoising leads to high decoding delay. To enable low-latency decoding, we propose a flow-matching (FM) generative decoder under a new land-then-transport (LTT) paradigm that tightly integrates the physical wireless channel into a continuous-time probability flow. For AWGN channels, we build a Gaussian smoothing path whose noise schedule indexes effective noise levels, and derive a closed-form teacher velocity field along this path. A neural-network student vector field is trained by conditional flow matching, yielding a deterministic, channel-aware ODE decoder with complexity linear in the number of ODE steps. At inference, it only needs an estimate of the effective noise variance to set the ODE starting time. We further show that Rayleigh fading and MIMO channels can be mapped, via linear MMSE equalization and singular-value-domain processing, to AWGN-equivalent channels with calibrated starting times. Therefore, the same probability path and trained velocity field can be reused for Rayleigh and MIMO without retraining. Experiments on MNIST, Fashion-MNIST, and DIV2K over AWGN, Rayleigh, and MIMO demonstrate consistent gains over JPEG2000+LDPC, DeepJSCC, and diffusion-based baselines, while achieving good perceptual quality with only a few ODE steps. Overall, LTT provides a deterministic, physically interpretable, and computation-efficient framework for generative wireless image decoding across diverse channels.

Method: Proposes GroupSegment SHAP (GS-SHAP) which constructs explanatory units as “group-segment players” based on cross-variable dependence and distribution shifts over time, then quantifies each unit’s contribution via Shapley attribution.

Result: GS-SHAP improves deletion-based faithfulness (DeltaAUC) by about 1.7x on average over time-series SHAP baselines, while reducing wall-clock runtime by about 40% on average under matched perturbation budgets. Successfully applied across four domains: human activity recognition, power-system forecasting, medical signal analysis, and financial time series.

Conclusion: GS-SHAP provides more faithful and efficient interpretation of multivariate time-series models by jointly considering cross-variable and temporal interactions, enabling identification of interpretable multivariate-temporal patterns during specific regimes like high market volatility.

Abstract: Multivariate time-series models achieve strong predictive performance in healthcare, industry, energy, and finance, but how they combine cross-variable interactions with temporal dynamics remains unclear. SHapley Additive exPlanations (SHAP) are widely used for interpretation. However, existing time-series variants typically treat the feature and time axes independently, fragmenting structural signals formed jointly by multiple variables over specific intervals. We propose GroupSegment SHAP (GS-SHAP), which constructs explanatory units as group-segment players based on cross-variable dependence and distribution shifts over time, and then quantifies each unit’s contribution via Shapley attribution. We evaluate GS-SHAP across four real-world domains: human activity recognition, power-system forecasting, medical signal analysis, and financial time series, and compare it with KernelSHAP, TimeSHAP, SequenceSHAP, WindowSHAP, and TSHAP. GS-SHAP improves deletion-based faithfulness (DeltaAUC) by about 1.7x on average over time-series SHAP baselines, while reducing wall-clock runtime by about 40 percent on average under matched perturbation budgets. A financial case study shows that GS-SHAP identifies interpretable multivariate-temporal interactions among key market variables during high-volatility regimes.

Method: Variational Decomposition Autoencoding (VDA) framework extends VAEs with structural bias toward signal decomposition. Uses DecVAEs - encoder-only networks combining signal decomposition model, contrastive self-supervised task, and variational prior approximation to learn multiple latent subspaces aligned with time-frequency characteristics.

Result: DecVAEs surpass state-of-the-art VAE-based methods in disentanglement quality, generalization across tasks, and interpretability of latent encodings. Demonstrated effectiveness on simulated data and three scientific datasets: speech recognition, dysarthria severity evaluation, and emotional speech classification.

Conclusion: Decomposition-aware architectures like VDA can serve as robust tools for extracting structured representations from dynamic signals, with potential applications in clinical diagnostics, human-computer interaction, and adaptive neurotechnologies.

Abstract: Understanding the structure of complex, nonstationary, high-dimensional time-evolving signals is a central challenge in scientific data analysis. In many domains, such as speech and biomedical signal processing, the ability to learn disentangled and interpretable representations is critical for uncovering latent generative mechanisms. Traditional approaches to unsupervised representation learning, including variational autoencoders (VAEs), often struggle to capture the temporal and spectral diversity inherent in such data. Here we introduce variational decomposition autoencoding (VDA), a framework that extends VAEs by incorporating a strong structural bias toward signal decomposition. VDA is instantiated through variational decomposition autoencoders (DecVAEs), i.e., encoder-only neural networks that combine a signal decomposition model, a contrastive self-supervised task, and variational prior approximation to learn multiple latent subspaces aligned with time-frequency characteristics. We demonstrate the effectiveness of DecVAEs on simulated data and three publicly available scientific datasets, spanning speech recognition, dysarthria severity evaluation, and emotional speech classification. Our results demonstrate that DecVAEs surpass state-of-the-art VAE-based methods in terms of disentanglement quality, generalization across tasks, and the interpretability of latent encodings. These findings suggest that decomposition-aware architectures can serve as robust tools for extracting structured representations from dynamic signals, with potential applications in clinical diagnostics, human-computer interaction, and adaptive neurotechnologies.

Method: Three main contributions: 1) High-throughput pipeline reformulating Pythagorean triple generation into single-parameter index stream; 2) Hypothesis-driven Negative Dataset (HND) with nine classes of adversarial attacks; 3) Fault-tolerant infrastructure for large-scale training using 10B deterministic samples and 5B adversarial counterexamples.

Result: LightGBM achieves 99.99% accuracy, but feature attribution reveals the model prioritizes underlying quadratic patterns over direct algebraic verification of Pythagorean triples.

Conclusion: Learned heuristics can effectively identify structural representations in numerical data and potentially serve as efficient preprocessors for formal verification methods, though they rely on pattern recognition rather than true algebraic understanding.

Abstract: This paper presents a large-scale stress test of machine learning systems using structured mathematical data as a benchmark. We evaluate the robustness of tree-based classifiers at an unprecedented scale, utilizing ten billion deterministic samples and five billion adversarial counterexamples. Our framework introduces three primary contributions: first, a high-throughput pipeline that reformulates Pythagorean triple generation into a single-parameter index stream, significantly improving computational efficiency over classical methods; second, the Hypothesis-driven Negative Dataset (HND), which categorizes nine classes of adversarial attacks designed to exploit both arithmetic precision and structural patterns; and third, a fault-tolerant infrastructure for reliable large-scale training. Experimental results demonstrate that while LightGBM achieves 99.99% accuracy, feature attribution reveals that the model prioritizes underlying quadratic patterns over direct algebraic verification. These findings suggest that learned heuristics can effectively identify structural representations in numerical data, potentially serving as efficient preprocessors for formal verification methods.

Method: L2CU identifies representative annotator profiles capturing distinct labeling patterns from sparse/noisy annotations, then matches unseen users to these profiles and leverages profile-specific models to complement users.

Result: L2CU demonstrates effectiveness across multiple datasets (CIFAR-10N, CIFAR-10H, Fashion-MNIST-H, Chaoyang, AgNews) as a model-agnostic solution for improving human-AI cooperative classification.

Conclusion: L2CU addresses the generalization challenge in human-AI cooperation by capturing user variability through representative profiles, enabling better complementarity with unseen users.

Abstract: Recent research highlights the potential of machine learning models to learn to complement (L2C) human strengths; however, generalizing this capability to unseen users remains a significant challenge. Existing L2C methods oversimplify interaction between human and AI by relying on a single, global user model that neglects individual user variability, leading to suboptimal cooperative performance. Addressing this, we introduce L2CU, a novel L2C framework for human-AI cooperative classification with unseen users. Given sparse and noisy user annotations, L2CU identifies representative annotator profiles capturing distinct labeling patterns. By matching unseen users to these profiles, L2CU leverages profile-specific models to complement the user and achieve superior joint accuracy. We evaluate L2CU on datasets (CIFAR-10N, CIFAR-10H, Fashion-MNIST-H, Chaoyang and AgNews), demonstrating its effectiveness as a model-agnostic solution for improving human-AI cooperative classification.

Method: Learn a shared representation space that aligns k-v caches of multiple models using adapter modules. These adapters translate each model’s internal state into and out of the shared space without altering the original pre-trained parameters.

Result: Experiments with Gemma-2 models show the approach enables seamless inter-model communication, improves individual model performance, and allows direct transfer of learned skills (like soft prompts) between different models.

Conclusion: This work represents significant progress toward models that can fluidly share knowledge and capabilities through direct internal state communication, moving beyond inefficient text-based collaboration.

Abstract: Solving increasingly complex problems with large language models (LLMs) necessitates a move beyond individual models and towards multi-model systems that can effectively collaborate. While text has traditionally served as the medium for inter-model communication, a richer and more efficient exchange is possible if models can access each other’s internal states directly. In this paper, we propose learning a shared representation space that aligns the k-v caches of multiple models, creating a high-bandwidth channel for collaboration without altering the underlying pre-trained parameters. We do so by augmenting each model with adapters to translate its state into and out of this shared space. Via a suite of experiments with Gemma-2 models, we demonstrate that this approach not only enables seamless inter-model communication but also improves individual model performance. We also show that the shared space allows for the direct transfer of learned skills, such as soft prompts, between different models. Our work represents a significant step towards a future where models can fluidly share knowledge and capabilities.

Method: Four-step pipeline: (1) constructs drivable networks from open geographic data, (2) solves Dijkstra routes minimizing edge traversal time, (3) extracts sparse operational features (signals, stops, crossings, turn counts), and (4) trains a random forest regression ensemble on limited high-quality reference times.

Result: Significant improvements over traversal-time baselines across multiple metrics (MAE, MAPE, MSE, relative bias, explained variance). No significant mean bias under minimally congested conditions, with consistent k-fold stability indicating negligible overfitting.

Conclusion: The approach offers a practical middle ground for transportation engineering: preserves point-to-point fidelity at metropolitan scale, reduces resource requirements, and provides defensible performance estimates where congestion data is inaccessible or cost-prohibitive, supporting planning and network applications under low-traffic conditions.

Abstract: Accurate roadway travel-time prediction is foundational to transportation systems analysis, yet widespread reliance on either data-intensive congestion models or overly naïve heuristics limits scalability and practical adoption in engineering workflows. This paper develops a lightweight estimator for minimally-congested car travel times that integrates open road-network data, speed constraints, and sparse control/turn features within a random forest framework to correct bias from shortest-path traversal-time baselines. Using an urban testbed, the pipeline: (i) constructs drivable networks from volunteered geographic data; (ii) solves Dijkstra routes minimizing edge traversal time; (iii) derives sparse operational features (signals, stops, crossings, yield, roundabouts; left/right/slight/U-turn counts); and (iv) trains a regression ensemble on limited high-quality reference times to generalize predictions beyond the training set. Out-of-sample evaluation demonstrates marked improvements over traversal-time baselines across mean absolute error, mean absolute percentage error, mean squared error, relative bias, and explained variance, with no significant mean bias under minimally congested conditions and consistent k-fold stability indicating negligible overfitting. The resulting approach offers a practical middle ground for transportation engineering: it preserves point-to-point fidelity at metropolitan scale, reduces resource requirements, and supplies defensible performance estimates where congestion feeds are inaccessible or cost-prohibitive, supporting planning, accessibility, and network performance applications under low-traffic operating regimes.

Method: AISCycleGen uses Cycle-Consistent Generative Adversarial Networks (CycleGAN) for unpaired domain translation, with a 1D convolutional generator and adaptive noise injection to preserve spatiotemporal structure of AIS trajectories.

Result: The method outperforms contemporary GAN-based augmentation techniques with PSNR of 30.5 and FID score of 38.9, and improves performance of baseline regression models across various maritime domains.

Conclusion: AISCycleGen is an effective and generalizable solution for augmenting AIS datasets, enhancing downstream model performance in real-world maritime intelligence applications.

Abstract: Automatic Identification System (AIS) data are vital for maritime domain awareness, yet they often suffer from domain shifts, data sparsity, and class imbalance, which hinder the performance of predictive models. In this paper, we propose a robust data augmentation method, AISCycleGen, based on Cycle-Consistent Generative Adversarial Networks (CycleGAN), which is tailored for AIS datasets. Unlike traditional methods, AISCycleGen leverages unpaired domain translation to generate high-fidelity synthetic AIS data sequences without requiring paired source-target data. The framework employs a 1D convolutional generator with adaptive noise injection to preserve the spatiotemporal structure of AIS trajectories, enhancing the diversity and realism of the generated data. To demonstrate its efficacy, we apply AISCycleGen to several baseline regression models, showing improvements in performance across various maritime domains. The results indicate that AISCycleGen outperforms contemporary GAN-based augmentation techniques, achieving a PSNR value of 30.5 and an FID score of 38.9. These findings underscore AISCycleGen’s potential as an effective and generalizable solution for augmenting AIS datasets, improving downstream model performance in real-world maritime intelligence applications.

Method: The paper proposes a taxonomy categorizing existing Online DPRL approaches into four families: Action-Gradient, Q-Weighting, Proximity-Based, and Backpropagation Through Time (BPTT) methods. It conducts extensive experiments on a unified NVIDIA Isaac Lab benchmark with 12 diverse robotic tasks, evaluating algorithms across five dimensions: task diversity, parallelization capability, diffusion step scalability, cross-embodiment generalization, and environmental robustness.

Result: The analysis identifies key findings about fundamental trade-offs in each algorithmic family, particularly concerning sample efficiency and scalability. It reveals critical computational and algorithmic bottlenecks limiting practical deployment of online DPRL. The study provides concrete guidelines for algorithm selection based on operational constraints.

Conclusion: The paper establishes a comprehensive framework for understanding Online DPRL algorithms, identifies current limitations, and provides practical guidelines for algorithm selection while outlining promising future research directions to advance toward more general and scalable robotic learning systems.

Abstract: Diffusion policies have emerged as a powerful approach for robotic control, demonstrating superior expressiveness in modeling multimodal action distributions compared to conventional policy networks. However, their integration with online reinforcement learning remains challenging due to fundamental incompatibilities between diffusion model training objectives and standard RL policy improvement mechanisms. This paper presents the first comprehensive review and empirical analysis of current Online Diffusion Policy Reinforcement Learning (Online DPRL) algorithms for scalable robotic control systems. We propose a novel taxonomy that categorizes existing approaches into four distinct families – Action-Gradient, Q-Weighting, Proximity-Based, and Backpropagation Through Time (BPTT) methods – based on their policy improvement mechanisms. Through extensive experiments on a unified NVIDIA Isaac Lab benchmark encompassing 12 diverse robotic tasks, we systematically evaluate representative algorithms across five critical dimensions: task diversity, parallelization capability, diffusion step scalability, cross-embodiment generalization, and environmental robustness. Our analysis identifies key findings regarding the fundamental trade-offs inherent in each algorithmic family, particularly concerning sample efficiency and scalability. Furthermore, we reveal critical computational and algorithmic bottlenecks that currently limit the practical deployment of online DPRL. Based on these findings, we provide concrete guidelines for algorithm selection tailored to specific operational constraints and outline promising future research directions to advance the field toward more general and scalable robotic learning systems.

Method: DeeperBrain incorporates domain-specific inductive biases: 1) Volume conduction-aware channel encoding using 3D geometry to model spatial mixing, 2) Neurodynamics-aware temporal encoding with oscillatory and exponential bases to capture slow adaptations. For pretraining, it uses dual objectives: Masked EEG Reconstruction (MER) for local fidelity and Neurodynamics Statistics Prediction (NSP) that predicts interpretable order parameters (spectral power, functional connectivity, cross-frequency coupling, dynamic complexity) to align with macroscopic brain states.

Result: DeeperBrain achieves state-of-the-art or highly competitive performance under end-to-end fine-tuning and maintains superior efficacy under rigorous frozen-probing protocols, demonstrating that embedding neuroscientific principles endows learned representations with intrinsic universality essential for universal BCI.

Conclusion: Integrating neuroscientific first principles into foundation model design and learning objectives creates representations with intrinsic universality, enabling effective performance in both fine-tuning and frozen-probing scenarios for universal brain-computer interfaces.

Abstract: Electroencephalography (EEG) foundation models hold significant promise for universal Brain-Computer Interfaces (BCIs). However, existing approaches often rely on end-to-end fine-tuning and exhibit limited efficacy under frozen-probing protocols, lacking the intrinsic universality required for broad generalization. This limitation stems from adapting general-purpose sequence architectures that overlook the biophysical and dynamical principles of neural activity. To bridge this gap, we propose DeeperBrain, a neuro-grounded foundation model integrating domain-specific inductive biases into its model design and learning objectives. Architecturally, DeeperBrain incorporates a volume conduction-aware channel encoding to model spatial mixing via 3D geometry, and a neurodynamics-aware temporal encoding capturing slow adaptations using oscillatory and exponential bases. For pretraining, we introduce a dual-objective strategy combining Masked EEG Reconstruction (MER) for local fidelity and Neurodynamics Statistics Prediction (NSP). NSP enforces alignment with macroscopic brain states by predicting interpretable order parameters, including spectral power, functional connectivity, cross-frequency coupling, and dynamic complexity. Extensive experiments demonstrate that DeeperBrain achieves state-of-the-art or highly competitive performance under end-to-end fine-tuning. Crucially, it maintains superior efficacy under a rigorous frozen-probing protocol, verifying that embedding neuroscientific first principles endows learned representations with the intrinsic universality essential for universal BCI. The code will be publicly available.

Method: ADF uses a query-conditioned, metric-induced attention operator in continuous space, treating spatial aggregation as geometric attention. Scalability is achieved via FAISS-accelerated inverted file indices, integrating approximate nearest-neighbor search as part of the attention mechanism.

Result: Demonstrated through aircraft trajectory analysis in Chengdu region, extracting trajectory-conditioned Zones of Influence (ZOI) that reveal recurrent airspace structures and localized deviations.

Conclusion: ADF provides a scalable geometric attention framework that successfully bridges adaptive kernel methods and attention mechanisms, enabling effective spatial analysis as shown in the aircraft trajectory case study.

Abstract: This work introduces Adaptive Density Fields (ADF), a geometric attention framework that formulates spatial aggregation as a query-conditioned, metric-induced attention operator in continuous space. By reinterpreting spatial influence as geometry-preserving attention grounded in physical distance, ADF bridges concepts from adaptive kernel methods and attention mechanisms. Scalability is achieved via FAISS-accelerated inverted file indices, treating approximate nearest-neighbor search as an intrinsic component of the attention mechanism. We demonstrate the framework through a case study on aircraft trajectory analysis in the Chengdu region, extracting trajectory-conditioned Zones of Influence (ZOI) to reveal recurrent airspace structures and localized deviations.

Method: Propose several practical test-time training (TTT) strategies for fine-tuning normalizing flows to adapt to distribution shifts during inference.

Result: Numerical study demonstrates TTT schemes are remarkably effective for real-time adjustment of flow-based inference for ABM parameters.

Conclusion: TTT provides a practical approach for adapting deep models like normalizing flows to handle distribution shifts in ABM parameter posterior estimation.

Abstract: Agent-Based Models (ABMs) are gaining great popularity in economics and social science because of their strong flexibility to describe the realistic and heterogeneous decisions and interaction rules between individual agents. In this work, we investigate for the first time the practicality of test-time training (TTT) of deep models such as normalizing flows, in the parameters posterior estimations of ABMs. We propose several practical TTT strategies for fine-tuning the normalizing flow against distribution shifts. Our numerical study demonstrates that TTT schemes are remarkably effective, enabling real-time adjustment of flow-based inference for ABM parameters.

Method: Proposes RainBalance, a plug-and-play module that performs clustering for each input sample to obtain cluster probability distribution, maps it into continuous latent space via VAE, and reformulates the task from fitting imbalanced labels to modeling continuous probabilistic label distributions.

Result: Integration into multiple state-of-the-art models shows consistent performance gains. Comprehensive statistical analysis and ablation studies validate the effectiveness of the approach.

Conclusion: RainBalance effectively addresses the dual imbalance problem in precipitation nowcasting by learning in continuous probabilistic space, improving model performance for practical disaster mitigation applications.

Abstract: Global navigation satellite systems (GNSS) station-based Precipitation Nowcasting aims to predict rainfall within the next 0-6 hours by leveraging a GNSS station’s historical observations of precipitation, GNSS-PWV, and related meteorological variables, which is crucial for disaster mitigation and real-time decision-making. In recent years, time-series forecasting approaches have been extensively applied to GNSS station-based precipitation nowcasting. However, the highly imbalanced temporal distribution of precipitation, marked not only by the dominance of non-rainfall events but also by the scarcity of extreme precipitation samples, significantly limits model performance in practical applications. To address the dual imbalance problem in precipitation nowcasting, we propose a continuous probability modeling-based framework, RainBalance. This plug-and-play module performs clustering for each input sample to obtain its cluster probability distribution, which is further mapped into a continuous latent space via a variational autoencoder (VAE). By learning in this continuous probabilistic space, the task is reformulated from fitting single and imbalance-prone precipitation labels to modeling continuous probabilistic label distributions, thereby alleviating the imbalance issue. We integrate this module into multiple state-of-the-art models and observe consistent performance gains. Comprehensive statistical analysis and ablation studies further validate the effectiveness of our approach.

Method: Integrated multimodal framework using cross-modal transformers with graph neural networks and causal representation learning. Combines genomic variation, cardiac MRI, ECG waveforms, wearable streams, and structured EHR data. Includes SHAP feature attribution, counterfactual explanations, causal latent alignment for interpretability, and federated privacy-preserving optimization with convergence, calibration, and uncertainty quantification rules.

Result: Experimental studies on large-scale biobank and multi-institutional datasets show state-of-the-art discrimination and robustness, with fair performance across demographic strata and clinically distinct cohorts.

Conclusion: The study provides a principled approach for clinically trustworthy, interpretable, and privacy-respecting CVD prediction at population level, paving the way for practical clinical implementation.

Abstract: Cardiovascular disease (CVD) continues to be the major cause of death globally, calling for predictive models that not only handle diverse and high-dimensional biomedical signals but also maintain interpretability and privacy. We create a single multimodal learning framework that integrates cross modal transformers with graph neural networks and causal representation learning to measure personalized CVD risk. The model combines genomic variation, cardiac MRI, ECG waveforms, wearable streams, and structured EHR data to predict risk while also implementing causal invariance constraints across different clinical subpopulations. To maintain transparency, we employ SHAP based feature attribution, counterfactual explanations and causal latent alignment for understandable risk factors. Besides, we position the design in a federated, privacy, preserving optimization protocol and establish rules for convergence, calibration and uncertainty quantification under distributional shift. Experimental studies based on large-scale biobank and multi institutional datasets reveal state discrimination and robustness, exhibiting fair performance across demographic strata and clinically distinct cohorts. This study paves the way for a principled approach to clinically trustworthy, interpretable and privacy respecting CVD prediction at the population level.

Method: Introduces a general method for sampling from a product-of-experts of a diffusion/flow matching model and an ’expert’ with binned probability density. Specifically combines neural diffusion processes with LLM token probabilities for regression tasks.

Result: The combined approach exceeds the empirical performance of either neural diffusion processes or LLMs alone on regression tasks, effectively leveraging both textual knowledge and meta-learning capabilities.

Conclusion: The product-of-experts framework successfully integrates neural diffusion processes with LLM knowledge, creating a hybrid approach that outperforms individual methods for regression tasks requiring both meta-learning and prior knowledge incorporation.

Abstract: Meta-learning methods for regression like Neural (Diffusion) Processes achieve impressive results, but with these models it can be difficult to incorporate expert prior knowledge and information contained in metadata. Large Language Models (LLMs) are trained on giant corpora including varied real-world regression datasets alongside their descriptions and metadata, leading to impressive performance on a range of downstream tasks. Recent work has extended this to regression tasks and is able to leverage such prior knowledge and metadata, achieving surprisingly good performance, but this still rarely matches dedicated meta-learning methods. Here we introduce a general method for sampling from a product-of-experts of a diffusion or flow matching model and an `expert’ with binned probability density; we apply this to combine neural diffusion processes with LLM token probabilities for regression (which may incorporate textual knowledge), exceeding the empirical performance of either alone.

Method: Self-supervised pre-training using 2,444 hours of unlabeled CTG recordings with masked pre-training, followed by fine-tuning on the 552-recording CTU-UHB benchmark. Uses PatchTST transformer architecture with channel-asymmetric masking scheme designed for fetal heart rate reconstruction.

Result: Achieved AUROC of 0.83 on full test set and 0.853 on uncomplicated vaginal deliveries, exceeding previously reported results (0.68-0.75). Error analysis shows false-positive alerts correspond to clinically concerning CTG patterns.

Conclusion: Self-supervised pre-training addresses data scarcity in fetal monitoring, offering a path toward reliable decision support in labor rooms. Standardized dataset splits and model weights released for reproducible benchmarking.

Abstract: Intrapartum cardiotocography (CTG) is widely used for fetal monitoring during labor, yet its interpretation suffers from high inter-observer variability and limited predictive accuracy. Deep learning approaches have been constrained by the scarcity of CTG recordings with clinical outcome labels. We present the first application of self-supervised pre-training to intrapartum CTG analysis, leveraging 2,444 hours of unlabeled recordings for masked pre-training followed by fine-tuning on the 552-recording CTU-UHB benchmark. Using a PatchTST transformer architecture with a channel-asymmetric masking scheme designed for fetal heart rate reconstruction, we achieve an area under the receiver operating characteristic curve of 0.83 on the full test set and 0.853 on uncomplicated vaginal deliveries, exceeding previously reported results on this benchmark (0.68-0.75). Error analysis reveals that false-positive alerts typically correspond to CTG patterns judged concerning on retrospective clinical review, suggesting clinically meaningful predictions even when umbilical pH is normal. We release standardized dataset splits and model weights to enable reproducible benchmarking. Our results demonstrate that self-supervised pre-training can address data scarcity in fetal monitoring, offering a path toward reliable decision support in the labor room.

Method: Introduces a dual-metric evaluation framework (ROS for strict parsing reliability, CSS for semantic capability) and PromptPort - a reliability layer combining deterministic canonicalization with a lightweight DistilBERT verifier and safe-override policy to recover format failures and improve extraction quality.

Result: On 37,346-example camera metadata benchmark across six model families, found severe format collapse (e.g., Gemma-2B: ROS 0.116 vs CSS 0.246) and large cross-model portability gaps. PromptPort recovers format failures (+6-8 F1), adds verifier-driven semantic selection (+14-16 F1), and approaches per-field oracle performance (0.890 vs 0.896 zero-shot).

Conclusion: PromptPort enables reliable structured extraction in production by fixing format collapse issues without modifying base models, generalizes to held-out model families, and provides explicit abstention when uncertain, making LLM extraction more robust for real-world deployment.

Abstract: Structured extraction with LLMs fails in production not because models lack understanding, but because output formatting is unreliable across models and prompts. A prompt that returns clean JSON on GPT-4 may produce fenced, prose-wrapped, or malformed output on Llama, causing strict parsers to reject otherwise correct extractions. We formalize this as format collapse and introduce a dual-metric evaluation framework: ROS (strict parsing, measuring operational reliability) and CSS (post-canonicalization, measuring semantic capability). On a 37,346-example camera metadata benchmark across six model families, we find severe format collapse (for example, Gemma-2B: ROS 0.116 versus CSS 0.246) and large cross-model portability gaps (0.4 to 0.6 F1). We then present PromptPort, a reliability layer combining deterministic canonicalization with a lightweight verifier (DistilBERT) and a safe-override policy. PromptPort recovers format failures (plus 6 to 8 F1), adds verifier-driven semantic selection (plus 14 to 16 F1 beyond canonicalization), and approaches per-field oracle performance (0.890 versus 0.896 in zero-shot) without modifying base models. The method generalizes to held-out model families and provides explicit abstention when uncertain, enabling reliable structured extraction in production deployments.

Method: Introduces CITA (Contrastive Instruction-Tuned Alignment) which combines supervised fine-tuning (SFT) with contrastive preference optimization under an explicit geometric anchor to a reference model. This creates a stable Riemannian chart keeping instruction updates within a shared neighborhood for reliable switching. Also introduces ECLIPTICA benchmark with 3000 controlled cases to isolate policy switching from ordinary instruction following.

Result: On Llama-3.1-8B across five evaluation suites (ECLIPTICA, TruthfulQA, Conditional Safety, Length Control, LITMUS), CITA achieves 86.7% instruction-alignment efficiency, significantly outperforming DPO (56.1%), GRPO (36.1%), and PPO (20.4%).

Conclusion: ECLIPTICA successfully demonstrates that alignment can be made instruction-driven and runtime-controllable, enabling on-the-fly behavioral modulation through natural-language alignment instructions that act as explicit behavioral contracts, with CITA providing a stable and effective method for achieving this.

Abstract: Alignment in large language models (LLMs) is still largely static: after training, the policy is frozen. DPO, GRPO methods typically imprint one behavior into the weights, leaving little runtime control beyond prompt hacks or expensive re-alignment. We introduce ECLIPTICA, which treats alignment as instruction-driven and runtime-controllable: natural-language alignment instructions act as an explicit behavioral contract (stance, refusal boundary, verbosity) that modulates behavior on the fly under evolving safety requirements, user roles, and governance constraints. We introduce CITA (Contrastive Instruction-Tuned Alignment), combining SFT with contrastive preference optimization under an explicit geometric anchor to a reference model. This yields a stable Riemannian chart and keeps instruction updates within a shared neighborhood, so regimes stay nearby and traversable for reliable switching. To isolate policy switching from ordinary instruction following, we release the ECLIPTICA benchmark: 3000 controlled cases (300 prompts x 10 instruction types) where the user request is fixed and only the alignment instruction changes. On Llama-3.1-8B across five suites (ECLIPTICA, TruthfulQA, Conditional Safety, Length Control, LITMUS), CITA reaches 86.7% instruction-alignment efficiency, beating DPO (56.1%), GRPO (36.1%), and PPO (20.4%).

Method: Simulate data from literature summary statistics, pretrain random forests on simulated data, then fine-tune on real dataset. Compare with random forests trained only on real data using Monte Carlo Cross Validation (100 iterations).

Result: Study 1: Some pretrained random forests descriptively outperformed standard RF but not significantly (t(99)=0.89, p=0.19). Study 2: Standard RF trained only on real data outperformed pretrained models.

Conclusion: Pretraining on simulated literature data doesn’t reliably improve performance. Challenges include scarcity of informative publications. Need better methodology for leveraging published statistics.

Abstract: In the context of personalized medicine, machine learning algorithms are growing in popularity. These algorithms require substantial information, which can be acquired effectively through the usage of previously gathered data. Open data and the utilization of synthetization techniques have been proposed to address this. In this paper, we propose and evaluate alternative approach that uses additional simulated data based on summary statistics published in the literature. The simulated data are used to pretrain random forests, which are afterwards fine-tuned on a real dataset. We compare the predictive performance of the new approach to random forests trained only on the real data. A Monte Carlo Cross Validation (MCCV) framework with 100 iterations was employed to investigate significance and stability of the results. Since a first study yielded inconclusive results, a second study with improved methodology (i.e., systematic information extraction and different prediction outcome) was conducted. In Study 1, some pretrained random forests descriptively outperformed the standard random forest. However, this improvement was not significant (t(99) = 0.89, p = 0.19). Contrary to expectations, in Study 2 the random forest trained only with the real data outperformed the pretrained random forests. We conclude with a discussion of challenges, such as the scarcity of informative publications, and recommendations for future research.

Method: ScaPre introduces: 1) Conflict-aware stable design with spectral trace regularization and geometry alignment to stabilize optimization and preserve global structure; 2) Informax Decoupler that identifies concept-relevant parameters and adaptively reweights updates to confine unlearning to target subspace; 3) Efficient closed-form solution without auxiliary data or sub-models.

Result: ScaPre effectively removes target concepts while maintaining generation quality, forgetting up to 5× more concepts than the best baseline within acceptable quality limits, achieving state-of-the-art precision and efficiency for large-scale unlearning across objects, styles, and explicit content.

Conclusion: ScaPre provides a unified framework for large-scale concept unlearning that addresses key challenges of conflicting updates, collateral damage, and scalability, offering an efficient solution without requiring additional data or modules while maintaining generation quality.

Abstract: Text-to-image diffusion models have achieved remarkable progress, yet their use raises copyright and misuse concerns, prompting research into machine unlearning. However, extending multi-concept unlearning to large-scale scenarios remains difficult due to three challenges: (i) conflicting weight updates that hinder unlearning or degrade generation; (ii) imprecise mechanisms that cause collateral damage to similar content; and (iii) reliance on additional data or modules, creating scalability bottlenecks. To address these, we propose Scalable-Precise Concept Unlearning (ScaPre), a unified framework tailored for large-scale unlearning. ScaPre introduces a conflict-aware stable design, integrating spectral trace regularization and geometry alignment to stabilize optimization, suppress conflicts, and preserve global structure. Furthermore, an Informax Decoupler identifies concept-relevant parameters and adaptively reweights updates, strictly confining unlearning to the target subspace. ScaPre yields an efficient closed-form solution without requiring auxiliary data or sub-models. Comprehensive experiments on objects, styles, and explicit content demonstrate that ScaPre effectively removes target concepts while maintaining generation quality. It forgets up to $\times \mathbf{5}$ more concepts than the best baseline within acceptable quality limits, achieving state-of-the-art precision and efficiency for large-scale unlearning.

Method: Adds a parent layer that computes three reflex signals (incident detection, overconfidence correction, recovery memory), fuses them into a bounded reliability index [0,1], which continuously modulates the learner’s effective learning rate.

Result: Empirical evaluations show improved calibration, reduced loss variance, and faster recovery compared to standard optimizers, while maintaining computational simplicity.

Conclusion: PGAR serves as a plug-in reliability layer for existing optimization pipelines, providing interpretable reliability traces suitable for safety-relevant applications, with theoretical guarantees under mild assumptions.

Abstract: Parent-Guided Adaptive Reliability (PGAR) is a lightweight behavioural meta-learning framework that adds a supervisory “parent” layer on top of a standard learner to improve stability, calibration, and recovery under disturbances. PGAR computes three reflex-level signals (incident detection, overconfidence correction, and recovery memory) and fuses them into a bounded reliability index in [0,1]. This index continuously modulates the learner’s effective learning rate, reducing update magnitude during instability and restoring it as reliability improves. We provide a Lyapunov-based proof sketch establishing bounded adaptation of the reliability dynamics under mild assumptions (smooth loss, descent direction, and bounded reflex outputs). Empirical evaluations on representative learning tasks show improved calibration, reduced loss variance, and faster recovery compared to standard optimizers, while retaining computational simplicity. PGAR functions as a plug-in reliability layer for existing optimization and learning pipelines, supporting interpretable reliability traces in safety-relevant settings.

Method: Mixed Logit Direct Preference Optimization (MixDPO) generalizes DPO by modeling preference strength variation across training examples using learned strength distributions.

Result: MixDPO improves aggregate alignment performance by +11.2 points on Pythia-2.8B across three datasets, with largest gains in high heterogeneity settings, while preserving subgroup preferences.

Conclusion: Modeling preference heterogeneity through MixDPO enables more faithful human judgment capture and better alignment performance, making preference strength variation explicit and learnable.

Abstract: Preference based alignment objectives implicitly assume that all human preferences are expressed with equal strength. In practice, however, preference strength varies across individuals and contexts – a phenomenon established in behavioral economics and discrete choice theory. This mismatch limits the ability of existing objectives to faithfully capture heterogeneous human judgments. Inspired by this literature, we introduce Mixed Logit Direct Preference Optimization (MixDPO), a generalization of Direct Preference Optimization that models variation in preference strength. MixDPO enables alignment objectives to capture heterogeneity in how strongly preferences are expressed across training examples. We evaluate MixDPO on three preference datasets using two open-weight language models. Across datasets, MixDPO improves aggregate alignment performance (+11.2 points on Pythia-2.8B) while preserving subgroup level preferences, with the largest gains appearing in settings with higher inferred preference heterogeneity. MixDPO makes preference heterogeneity explicit through learned strength distributions. We release our code for reproducibility.

Method: Organizes a community challenge with three tracks: compression (compact representations), forecasting (predicting future states), and sensing (inferring unmeasured states). Provides standardized success metrics, evaluation tools, baseline implementations (one classical and one ML baseline per challenge), and uses blind tests on withheld data for final assessment.

Result: The challenge framework is established and open for participation, with outcomes to be disseminated through an AIAA Journal Virtual Collection and invited presentations at AIAA conferences.

Conclusion: This community challenge aims to provide fair comparisons of data-driven methods for aerospace flow analysis, encouraging broad participation including negative results and careful limitation analyses to advance the field.

Abstract: We introduce a community challenge designed to facilitate direct comparisons between data-driven methods for compression, forecasting, and sensing of complex aerospace flows. The challenge is organized into three tracks that target these complementary capabilities: compression (compact representations for large datasets), forecasting (predicting future flow states from a finite history), and sensing (inferring unmeasured flow states from limited measurements). Across these tracks, multiple challenges span diverse flow datasets and use cases, each emphasizing different model requirements. The challenge is open to anyone, and we invite broad participation to build a comprehensive and balanced picture of what works and where current methods fall short. To support fair comparisons, we provide standardized success metrics, evaluation tools, and baseline implementations, with one classical and one machine-learning baseline per challenge. Final assessments use blind tests on withheld data. We explicitly encourage negative results and careful analyses of limitations. Outcomes will be disseminated through an AIAA Journal Virtual Collection and invited presentations at AIAA conferences.

Method: Propose statistical detector using Mahalanobis distance and forward-backward detection fractions to adjust supervised training labels for LSTM networks. Tested on digital twin simulations of ground-stage propulsion system.

Result: Statistical data relabeling improved LSTM classifier precision by 7% and recall by 22% on propulsion system simulations with 20.8 minutes per trial and O(10^8) training timesteps.

Conclusion: The proposed statistical relabeling method effectively enhances LSTM-based anomaly detection for launch vehicle telemetry, addressing limitations of traditional approaches and improving classification performance.

Abstract: Supporting Go/No-Go decisions prior to launch requires assessing real-time telemetry data against redline limits established during the design qualification phase. Family data from ground testing or previous flights is commonly used to detect initiating failure modes and their timing; however, this approach relies heavily on engineering judgment and is more error-prone for new launch vehicles. To address these limitations, we utilize Long-Term Short-Term Memory (LSTM) networks for supervised classification of time-series anomalies. Although, initial training labels derived from simulated anomaly data may be suboptimal due to variations in anomaly strength, anomaly settling times, and other factors. In this work, we propose a novel statistical detector based on the Mahalanobis distance and forward-backward detection fractions to adjust the supervised training labels. We demonstrate our method on digital twin simulations of a ground-stage propulsion system with 20.8 minutes of operation per trial and O(10^8) training timesteps. The statistical data relabeling improved precision and recall of the LSTM classifier by 7% and 22% respectively.

Method: Proposes TG-DCMADDPG: 1) Temporal Graph Neural Network (TimeGNN) models/predicts multi-dimensional server states to reduce online interactions; 2) Multi-agent deterministic policy gradient (DC-MADDPG) in discrete-continuous hybrid action space optimizes task partitioning ratios, transmission power, and priority scheduling collaboratively.

Result: Simulations show TG-DCMADDPG achieves faster policy convergence, superior energy-latency optimization, higher task completion rates than state-of-the-art methods, demonstrating robust scalability and practical effectiveness in dynamic MEC scenarios.

Conclusion: TG-DCMADDPG effectively addresses MEC challenges through temporal modeling and multi-agent collaborative optimization, offering a practical solution for dynamic edge computing environments with constrained resources.

Abstract: With the rapid growth of IoT devices and latency-sensitive applications, the demand for both real-time and energy-efficient computing has surged, placing significant pressure on traditional cloud computing architectures. Mobile edge computing (MEC), an emerging paradigm, effectively alleviates the load on cloud centers and improves service quality by offloading computing tasks to edge servers closer to end users. However, the limited computing resources, non-continuous power provisioning (e.g., battery-powered nodes), and highly dynamic systems of edge servers complicate efficient task scheduling and resource allocation. To address these challenges, this paper proposes a multi-agent deep reinforcement learning algorithm, TG-DCMADDPG, and constructs a collaborative computing framework for multiple edge servers, aiming to achieve joint optimization of fine-grained task partitioning and offloading. This approach incorporates a temporal graph neural network (TimeGNN) to model and predict time series of multi-dimensional server state information, thereby reducing the frequency of online interactions and improving policy predictability. Furthermore, a multi-agent deterministic policy gradient algorithm (DC-MADDPG) in a discrete-continuous hybrid action space is introduced to collaboratively optimize task partitioning ratios, transmission power, and priority scheduling strategies. Extensive simulation experiments confirm that TG-DCMADDPG achieves markedly faster policy convergence, superior energy-latency optimization, and higher task completion rates compared with existing state-of-the-art methods, underscoring its robust scalability and practical effectiveness in dynamic and constrained MEC scenarios.

Method: Created MLB benchmark with 5 dimensions: Medical Knowledge, Safety/Ethics, Medical Record Understanding, Smart Services, and Smart Healthcare. Used 22 datasets (17 new) from Chinese clinical sources across 64 specialties, curated by 300 physicians. Developed specialized judge model via SFT on 19k expert annotations.

Result: Top model Kimi-K2-Instruct scored 77.3% overall but showed large performance gap: 87.8% in structured tasks vs 61.3% in patient-facing scenarios. Baichuan-M2-32B achieved 90.6% safety score despite smaller size. Judge model achieved 92.1% accuracy, 94.37% F1, and 81.3% Cohen’s Kappa for human-AI consistency.

Conclusion: MLB provides a rigorous framework to guide development of clinically viable LLMs, revealing critical translational gaps between knowledge and practical application that must be addressed for real-world healthcare deployment.

Abstract: The proliferation of Large Language Models (LLMs) presents transformative potential for healthcare, yet practical deployment is hindered by the absence of frameworks that assess real-world clinical utility. Existing benchmarks test static knowledge, failing to capture the dynamic, application-oriented capabilities required in clinical practice. To bridge this gap, we introduce a Medical LLM Benchmark MLB, a comprehensive benchmark evaluating LLMs on both foundational knowledge and scenario-based reasoning. MLB is structured around five core dimensions: Medical Knowledge (MedKQA), Safety and Ethics (MedSE), Medical Record Understanding (MedRU), Smart Services (SmartServ), and Smart Healthcare (SmartCare). The benchmark integrates 22 datasets (17 newly curated) from diverse Chinese clinical sources, covering 64 clinical specialties. Its design features a rigorous curation pipeline involving 300 licensed physicians. Besides, we provide a scalable evaluation methodology, centered on a specialized judge model trained via Supervised Fine-Tuning (SFT) on expert annotations. Our comprehensive evaluation of 10 leading models reveals a critical translational gap: while the top-ranked model, Kimi-K2-Instruct (77.3% accuracy overall), excels in structured tasks like information extraction (87.8% accuracy in MedRU), performance plummets in patient-facing scenarios (61.3% in SmartServ). Moreover, the exceptional safety score (90.6% in MedSE) of the much smaller Baichuan-M2-32B highlights that targeted training is equally critical. Our specialized judge model, trained via SFT on a 19k expert-annotated medical dataset, achieves 92.1% accuracy, an F1-score of 94.37%, and a Cohen’s Kappa of 81.3% for human-AI consistency, validating a reproducible and expert-aligned evaluation protocol. MLB thus provides a rigorous framework to guide the development of clinically viable LLMs.

Method: Proposes EntroLnn framework using entropy-guided transformable liquid neural networks. Introduces entropy-based features derived from online temperature fields (first application in battery analytics) combined with customized LNNs that effectively model temporal battery dynamics. The framework enhances both static and dynamic adaptability of LNNs.

Result: Achieves robust and generalizable capacity fade trajectory refinement across different batteries and operating conditions. Provides high fidelity battery health model with lightweight computation, achieving mean absolute errors of only 0.004577 for CFT and 18 cycles for EoL prediction.

Conclusion: Establishes foundation for entropy-informed learning in battery analytics and enables self-adaptive, lightweight, and interpretable battery health prediction for practical battery management systems.

Abstract: Battery capacity degradation prediction has long been a central topic in battery health analytics, and most studies focus on state of health (SoH) estimation and end of life (EoL) prediction. This study extends the scope to online refinement of the entire capacity fade trajectory (CFT) through EntroLnn, a framework based on entropy-guided transformable liquid neural networks (LNNs). EntroLnn treats CFT refinement as an integrated process rather than two independent tasks for pointwise SoH and EoL. We introduce entropy-based features derived from online temperature fields, applied for the first time in battery analytics, and combine them with customized LNNs that model temporal battery dynamics effectively. The framework enhances both static and dynamic adaptability of LNNs and achieves robust and generalizable CFT refinement across different batteries and operating conditions. The approach provides a high fidelity battery health model with lightweight computation, achieving mean absolute errors of only 0.004577 for CFT and 18 cycles for EoL prediction. This work establishes a foundation for entropy-informed learning in battery analytics and enables self-adaptive, lightweight, and interpretable battery health prediction in practical battery management systems.

Method: MB-ICL uses latent representations from frozen LLMs to model local manifold structure and class-aware prototype geometry. It selects demonstrations based on proximity to learned prototypes rather than lexical or embedding similarity alone.

Result: Outperforms standard ICL selection baselines on factual verification (FEVER) and hallucination detection (HaluEval) benchmarks, with strong gains on dialogue and summarization tasks. Shows robustness under temperature perturbations and model variation.

Conclusion: Manifold-based prototype selection provides a reliable, training-light approach for hallucination detection without modifying LLM parameters, offering a principled direction for improved ICL demonstration selection.

Abstract: Large language models (LLMs) frequently generate factually incorrect or unsupported content, commonly referred to as hallucinations. Prior work has explored decoding strategies, retrieval augmentation, and supervised fine-tuning for hallucination detection, while recent studies show that in-context learning (ICL) can substantially influence factual reliability. However, existing ICL demonstration selection methods often rely on surface-level similarity heuristics and exhibit limited robustness across tasks and models. We propose MB-ICL, a manifold-based demonstration sampling framework for selecting in-context demonstrations that leverages latent representations extracted from frozen LLMs. By jointly modeling local manifold structure and class-aware prototype geometry, MB-ICL selects demonstrations based on their proximity to learned prototypes rather than lexical or embedding similarity alone. Across factual verification (FEVER) and hallucination detection (HaluEval) benchmarks, MB-ICL outperforms standard ICL selection baselines in the majority of evaluated settings, with particularly strong gains on dialogue and summarization tasks. The method remains robust under temperature perturbations and model variation, indicating improved stability compared to heuristic retrieval strategies. While lexical retrieval can remain competitive in certain question-answering regimes, our results demonstrate that manifold-based prototype selection provides a reliable and training light approach for hallucination detection without modifying LLM parameters, offering a principled direction for improved ICL demonstration selection.

Method: Two key innovations: 1) Structure refinement module trained via mask mutation modeling on wild-type structures to generate inaccessible mutant structures, and 2) Probability density cloud network (PDC-Net) to capture 3D dynamic variations and encode atomic uncertainty in PPIs.

Result: Comprehensive experiments on SKEMPI.v2 benchmark show Refine-PPI outperforms all existing tools for predicting free energy change (ΔΔG) of mutations in protein-protein interactions.

Conclusion: The framework effectively addresses the absence of mutant protein structures through structure hallucination and successfully models geometric uncertainty in dynamic PPIs, demonstrating superior performance in mutation effect prediction.

Abstract: Protein-protein interaction (PPI) represents a central challenge within the biology field, and accurately predicting the consequences of mutations in this context is crucial for drug design and protein engineering. Deep learning (DL) has shown promise in forecasting the effects of such mutations, but is hindered by two primary constraints. First, the structures of mutant proteins are often elusive to acquire. Secondly, PPI takes place dynamically, which is rarely integrated into the DL architecture design. To address these obstacles, we present a novel framework named Refine-PPI with two key enhancements. First, we introduce a structure refinement module trained by a mask mutation modeling (MMM) task on available wild-type structures, which is then transferred to produce the inaccessible mutant structures. Second, we employ a new kind of geometric network, called the probability density cloud network (PDC-Net), to capture 3D dynamic variations and encode the atomic uncertainty associated with PPI. Comprehensive experiments on SKEMPI.v2 substantiate the superiority of Refine-PPI over all existing tools for predicting free energy change. These findings underscore the effectiveness of our hallucination strategy and the PDC module in addressing the absence of mutant protein structure and modeling geometric uncertainty.

Method: Combines Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose vibration signals into intrinsic mode functions, followed by Multiscale Convolutional Neural Network (MSCNN) for deep hierarchical feature extraction and classification.

Result: Achieves 98.95% F1 score on real-world datasets, demonstrating superior performance in both detection accuracy and computational speed compared to existing approaches.

Conclusion: The hybrid CEEMDAN-MSCNN framework offers a balanced solution for reliable and efficient fault diagnosis in wind turbine systems, addressing the challenges of complex vibration signal analysis.

Abstract: Wind turbines play a critical role in the shift toward sustainable energy generation. Their operation relies on multiple interconnected components, and a failure in any of these can compromise the entire system’s functionality. Detecting faults accurately is challenging due to the intricate, non-linear, and non-stationary nature of vibration signals, influenced by dynamic loading, environmental variations, and mechanical interactions. As such, effective signal processing techniques are essential for extracting meaningful features to enhance diagnostic accuracy. This study presents a hybrid approach for fault detection in wind turbine gearboxes, combining Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and a Multiscale Convolutional Neural Network (MSCNN). CEEMDAN is employed to decompose vibration signals into intrinsic mode functions, isolating critical features at different time-frequency scales. These are then input into the MSCNN, which performs deep hierarchical feature extraction and classification. The proposed method achieves an F1 Score of 98.95%, evaluated on real-world datasets, and demonstrates superior performance in both detection accuracy and computational speed compared to existing approaches. This framework offers a balanced solution for reliable and efficient fault diagnosis in wind turbine systems.

Method: ZeroRouter uses a universal latent space - a model-agnostic representation of query difficulty that decouples query characterization from model profiling. It features a context-aware predictor to map queries to this space and a dual-mode optimizer to balance accuracy, cost, and latency.

Result: ZeroRouter consistently outperforms all baselines, delivering higher accuracy at lower cost and latency. It enables zero-shot onboarding of new models without full-scale retraining.

Conclusion: ZeroRouter breaks the model lock-in problem in LLM ecosystems by providing a scalable, adaptable routing framework that eliminates costly retraining while maintaining performance advantages over existing approaches.

Abstract: The rapid proliferation of Large Language Models (LLMs) has led to a fragmented and inefficient ecosystem, a state of ``model lock-in’’ where seamlessly integrating novel models remains a significant bottleneck. Current routing frameworks require exhaustive, costly retraining, hindering scalability and adaptability. We introduce ZeroRouter, a new paradigm for LLM routing that breaks this lock-in. Our approach is founded on a universal latent space, a model-agnostic representation of query difficulty that fundamentally decouples the characterization of a query from the profiling of a model. This allows for zero-shot onboarding of new models without full-scale retraining. ZeroRouter features a context-aware predictor that maps queries to this universal space and a dual-mode optimizer that balances accuracy, cost, and latency. Our framework consistently outperforms all baselines, delivering higher accuracy at lower cost and latency.

Method: Proposes LDTC which combines dimensionality reduction and temporal clustering in an end-to-end deep unsupervised framework using a specifically designed autoencoder. It employs dynamic model expansion and rehearsal-based techniques to learn new tasks sequentially while preventing catastrophic forgetting.

Result: Experiments on seven real-world multivariate time series datasets demonstrate that LDTC is effective and efficient for temporal clustering, achieving high-quality clustering results while handling dynamic data in sequential learning.

Conclusion: LDTC is a promising method for temporal clustering that successfully addresses the challenges of sequential task learning, dynamic data handling, and catastrophic forgetting through its integrated framework and lifelong learning capabilities.

Abstract: Clustering temporal and dynamically changing multivariate time series from real-world fields, called temporal clustering for short, has been a major challenge due to inherent complexities. Although several deep temporal clustering algorithms have demonstrated a strong advantage over traditional methods in terms of model learning and clustering results, the accuracy of the few algorithms are not satisfactory. None of the existing algorithms can continuously learn new tasks and deal with the dynamic data effectively and efficiently in the sequential tasks learning. To bridge the gap and tackle these issues, this paper proposes a novel algorithm \textbf{L}ifelong \textbf{D}eep \textbf{T}emporal \textbf{C}lustering (\textbf{LDTC}), which effectively integrates dimensionality reduction and temporal clustering into an end-to-end deep unsupervised learning framework. Using a specifically designed autoencoder and jointly optimizing for both the latent representation and clustering objective, the LDTC can achieve high-quality clustering results. Moreover, unlike any previous work, the LDTC is uniquely equipped with the fully dynamic model expansion and rehearsal-based techniques to effectively learn new tasks and to tackle the dynamic data in the sequential tasks learning without the catastrophic forgetting or degradation of the model accuracy. Experiments on seven real-world multivariate time series datasets show that the LDTC is a promising method for dealing with temporal clustering issues effectively and efficiently.

Method: GLOSS (Global Toxic Subspace Suppression) identifies and eliminates global toxic subspaces from Feed-Forward Network (FFN) parameters. The method focuses on finding robust toxic subspaces that cannot be easily reconstructed from non-toxic components, addressing limitations of previous vector-based approaches.

Result: Experiments on LLMs (including Qwen3) show GLOSS achieves state-of-the-art detoxification performance while preserving general model capabilities. The method accomplishes this without requiring large-scale retraining.

Conclusion: GLOSS provides an effective lightweight solution for LLM detoxification by targeting global toxic subspaces in model parameters, overcoming limitations of previous approaches and enabling safer deployment of LLMs without compromising their general capabilities.

Abstract: Large language models (LLMs) exhibit exceptional performance but pose inherent risks of generating toxic content, restricting their safe deployment. While traditional methods (e.g., alignment) adjust output preferences, they fail to eliminate underlying toxic regions in parameters, leaving models vulnerable to adversarial attacks. Prior mechanistic studies characterize toxic regions as “toxic vectors” or “layer-wise subspaces”, yet our analysis identifies critical limitations: i) Removed toxic vectors can be reconstructed via linear combinations of non-toxic vectors, demanding targeting of entire toxic subspace; ii) Contrastive objective over limited samples inject noise into layer-wise subspaces, hindering stable extraction. These highlight the challenge of identifying robust toxic subspace and removing them. Therefore, we propose GLOSS (GLobal tOxic Subspace Suppression), a lightweight method that mitigates toxicity by identifying and eliminating this global subspace from FFN parameters. Experiments on LLMs (e.g., Qwen3) show GLOSS achieves SOTA detoxification while preserving general capabilities without requiring large-scale retraining. WARNING: This paper contains context which is toxic in nature.

Method: DLNet applies Euler discretization to make liquid neural networks compatible with embedded systems, then performs dual-stage knowledge distillation to transfer temporal behavior from a high-capacity teacher model to compact student models, followed by Pareto-guided selection based on joint error-cost objectives.

Result: The deployed student model achieves 0.0066 error predicting battery health over next 100 cycles (15.4% lower than teacher), reduces model size from 616 kB to 94 kB (84.7% reduction), and takes 21 ms per inference on Arduino Nano 33 BLE Sense with int8 deployment.

Conclusion: Small models can match or exceed large teachers for edge-based prognostics with proper supervision and selection, and the DLNet framework can extend to other industrial analytics tasks with strict hardware constraints.

Abstract: Battery management systems increasingly require accurate battery health prognostics under strict on-device constraints. This paper presents DLNet, a practical framework with dual-stage distillation of liquid neural networks that turns a high-capacity model into compact and edge-deployable models for battery health prediction. DLNet first applies Euler discretization to reformulate liquid dynamics for embedded compatibility. It then performs dual-stage knowledge distillation to transfer the teacher model’s temporal behavior and recover it after further compression. Pareto-guided selection under joint error-cost objectives retains student models that balance accuracy and efficiency. We evaluate DLNet on a widely used dataset and validate real-device feasibility on an Arduino Nano 33 BLE Sense using int8 deployment. The final deployed student achieves a low error of 0.0066 when predicting battery health over the next 100 cycles, which is 15.4% lower than the teacher model. It reduces the model size from 616 kB to 94 kB with 84.7% reduction and takes 21 ms per inference on the device. These results support a practical smaller wins observation that a small model can match or exceed a large teacher for edge-based prognostics with proper supervision and selection. Beyond batteries, the DLNet framework can extend to other industrial analytics tasks with strict hardware constraints.

Method: Train a simple ANN on minterm values, partition it into cells based on ReLU nodes, convert to a 3D bit tensor, then apply Formal Concept Analysis to extract concepts represented as logic trees that capture attribute interactions.

Result: The derived symbolic logic trees preserve the classification power of the original ANN model while providing interpretable representations of attribute interactions.

Conclusion: The proposed method successfully bridges the gap between ANN’s high accuracy and symbolic methods’ interpretability by extracting meaningful logic trees from ANN models without sacrificing classification performance.

Abstract: An artificial neural network (ANN) is a numerical method used to solve complex classification problems. Due to its high classification power, the ANN method often outperforms other classification methods in terms of accuracy. However, an ANN model lacks interpretability compared to methods that use the symbolic paradigm. Our idea is to derive a symbolic representation from a simple ANN model trained on minterm values of input objects. Based on ReLU nodes, the ANN model is partitioned into cells. We convert the ANN model into a cell-based, three-dimensional bit tensor. The theory of Formal Concept Analysis applied to the tensor yields concepts that are represented as logic trees, expressing interpretable attribute interactions. Their evaluations preserve the classification power of the initial ANN model.

Method: SPINAL measures alignment effects by tracing localized structural change layer by layer using two scores: contraction score (spectrum decay rate) and transport score (token distribution shift between layers). It encodes checkpoints as depth traces over (layer index, contraction, transport).

Result: DPO produces layerwise calibration concentrated in final decoder blocks (layers 21-30). Aligned checkpoints show late-layer ramp-up in contraction and smooth reduction in transport, while unaligned models show higher-curvature, more entropic paths.

Conclusion: Alignment is geometrically localized in final layers, which encode dominant preference-induced corrections. SPINAL provides practical audit signals for quantifying where alignment concentrates, its strength, and training stability.

Abstract: Direct Preference Optimization (DPO) is a principled, scalable alternative to RLHF for aligning large language models from pairwise preferences, but its internal geometric footprint remains undercharacterized, limiting audits, checkpoint comparisons, and failure prediction. We introduce SPINAL (Scaling-law and Preference Integration in Neural Alignment Layers), a diagnostic that measures how alignment reshapes representations across depth by tracing localized structural change layer by layer. Across model families, DPO produces a layerwise calibration effect concentrated in the final decoder blocks (often layers 21-30), where preference gradients most directly affect the next-token distribution. SPINAL encodes each checkpoint as a depth trace over (layer index, contraction score, transport score). The contraction score summarizes how quickly the tail of a layer’s spectrum decays (how fast small modes vanish); higher values indicate stronger contraction into fewer effective directions. The transport score summarizes how much the token distribution shifts between adjacent layers using a bounded overlap measure; lower values indicate shorter, smoother steps through representation space. Aligned checkpoints show a late-layer ramp-up in contraction and a smooth reduction in transport, consistent with tightened and stabilized policy mass, while unaligned models trace higher-curvature, more entropic, and geometrically incoherent depth paths. Overall, alignment is geometrically localized: the final layers encode the dominant preference-induced corrections. SPINAL turns this localization into a practical audit signal, quantifying where alignment concentrates, how strongly it manifests, and when it begins to destabilize during training.

Method: 1) Decomposes inference into analytically modelable primitives (GEMM, attention, communication, memory ops) while capturing framework-specific scheduling dynamics; 2) Uses calibrated kernel-level performance database across hardware platforms and models; 3) Provides abstraction layer that automatically resolves optimal launch parameters for target backends.

Result: Improves performance by up to 40% for dense models (Qwen3-32B) and 50% for MoE architectures (DeepSeek-V3), while completing configuration searches within 30 seconds on average.

Conclusion: AIConfigurator enables rapid exploration of vast LLM inference design spaces (from cluster topology to engine-specific flags) without GPU profiling, providing framework-agnostic optimization for production systems.

Abstract: Optimizing Large Language Model (LLM) inference in production systems is increasingly difficult due to dynamic workloads, stringent latency/throughput targets, and a rapidly expanding configuration space. This complexity spans not only distributed parallelism strategies (tensor/pipeline/expert) but also intricate framework-specific runtime parameters such as those concerning the enablement of CUDA graphs, available KV-cache memory fractions, and maximum token capacity, which drastically impact performance. The diversity of modern inference frameworks (e.g., TRT-LLM, vLLM, SGLang), each employing distinct kernels and execution policies, makes manual tuning both framework-specific and computationally prohibitive. We present AIConfigurator, a unified performance-modeling system that enables rapid, framework-agnostic inference configuration search without requiring GPU-based profiling. AIConfigurator combines (1) a methodology that decomposes inference into analytically modelable primitives - GEMM, attention, communication, and memory operations while capturing framework-specific scheduling dynamics; (2) a calibrated kernel-level performance database for these primitives across a wide range of hardware platforms and popular open-weights models (GPT-OSS, Qwen, DeepSeek, LLama, Mistral); and (3) an abstraction layer that automatically resolves optimal launch parameters for the target backend, seamlessly integrating into production-grade orchestration systems. Evaluation on production LLM serving workloads demonstrates that AIConfigurator identifies superior serving configurations that improve performance by up to 40% for dense models (e.g., Qwen3-32B) and 50% for MoE architectures (e.g., DeepSeek-V3), while completing searches within 30 seconds on average. Enabling the rapid exploration of vast design spaces - from cluster topology down to engine specific flags.

Method: Proposes SourceNet, a Transformer-based framework treating sensor arrays as flexible sets for arbitrary geometries. Introduces Physics-Structured Domain Randomization (PSDR) that randomizes physical dynamics (velocity structures, propagation effects, sensor availability) to force robust representations invariant to environmental heterogeneity.

Result: Achieves state-of-the-art precision on real-world data after pre-training on 100,000 synthetic events and fine-tuning on ~2,000 real events. Demonstrates exceptional data efficiency, matches classical solvers while enabling real-time processing. Model autonomously discovers geometric information bottlenecks and learns attention policies that prioritize sparse sensor placements.

Conclusion: SourceNet successfully addresses challenges in AI for Science by combining flexible Transformer architectures with physics-informed domain randomization. The model not only achieves practical performance but also exhibits scientific-agent-like behavior, recovering principles of optimal experimental design from data alone.

Abstract: Inferring high-dimensional physical states from sparse, ad-hoc sensor arrays is a fundamental challenge across AI for Science, as they are complicated by irregular geometries and the profound Sim-to-Real gap in physical modeling. Taking earthquake source characterization as a representative challenge, we address limitations in conventional deep learning: CNNs demand fixed grids, while pooling-based architectures (e.g., DeepSets) struggle to capture the relational wave physics. Here, we propose SourceNet, a Transformer-based framework that treats the sensor array as a flexible set to model arbitrary geometries. To bridge the reality gap, we introduce Physics-Structured Domain Randomization (PSDR). Instead of forcing feature alignment, PSDR randomizes the governing physical dynamics by varying velocity structures, propagation effects, and sensor availability, to force the model to learn robust representations invariant to unmodeled environmental heterogeneity. By pre-training on 100,000 synthetic events and fine-tuning on ~2,000 real world events, SourceNet achieves state-of-the-art precision on held-out real data. This demonstrates exceptional data efficiency, and matches classical solvers while enabling real-time processing. Remarkably, interpretability analysis reveals that the model shows scientific-agent-like features: it autonomously discovers geometric information bottlenecks and learns an attention policy that prioritizes sparse sensor placements, effectively recovering principles of optimal experimental design from data alone.

Method: Extends reinforcement learning with verifiable rewards to temporally resolved prediction, training language models to make probabilistic forecasts under causally masked information with retrospective evaluation using proper scoring rules.

Result: Qwen3-32B trained with Foresight Learning improves Brier score by 27% and halves calibration error relative to baseline, outperforms Qwen3-235B on future-event prediction tasks and Metaculus benchmark despite 7x fewer parameters.

Conclusion: Foresight Learning effectively addresses delayed supervision in real-world forecasting, enabling language models to make accurate probabilistic predictions using only post-resolution outcomes.

Abstract: Many real-world prediction problems lack labels observable at prediction time, creating a temporal gap between prediction and outcome that yields supervision only after events resolve. To address this setting, we extend reinforcement learning with verifiable rewards to temporally resolved real-world prediction, and use it to train language models to make probabilistic forecasts under causally masked information with retrospective evaluation using proper scoring rules. Supervision is derived solely from post-resolution outcomes, preserving delayed-reward semantics. On real-world forecasting benchmarks, Qwen3-32B trained using Foresight Learning improves Brier score by 27% and halves calibration error relative to its pretrained baseline, and outperforms Qwen3-235B on both constructed future-event prediction tasks and the Metaculus benchmark despite a 7x parameter disadvantage.

Method: Developed a comprehensive benchmark suite evaluating robustness across multiple perturbation types: general text edits (punctuation, whitespace), formatting changes (JSON, YAML), multilingual/cross-lingual inputs, and positional instruction variations. Evaluated 11 models ranging from 4B to 120B+ parameters.

Result: Minor perturbations reduce performance by up to 40 percentage points on key enterprise metrics. The relationship between model size and robustness is nuanced: an 8B parameter model (Ministral 3 8B) outperforms most larger models, while another 8B model (Llama 3.1 8B) performs worst overall.

Conclusion: Enterprise LLM applications face significant robustness challenges from minor prompt variations, and model size alone doesn’t guarantee robustness - careful model selection and comprehensive testing are essential for reliable enterprise deployment.

Abstract: Enterprise LLM applications require consistently high quality and reliable performance across diverse scenarios, demanding robustness to minor variations. Existing research shows that even small prompt changes can lead to substantial differences in output, but has mainly focused on a narrow set of perturbations with small academic datasets, limiting their relevance to real-world applications. To address this, we present a comprehensive benchmark suite that evaluates robustness across multiple perturbation types, including general text edits (e.g., punctuation, whitespace), formatting changes (e.g., JSON, YAML), multilingual and cross-lingual inputs, and positional variations in instructions. Evaluating 11 models ranging from 4B to 120B+ parameters, we find that minor perturbations reduce performance by up to 40 percentage points on key enterprise metrics. Critically, we demonstrate that the relationship between model size and robustness is more nuanced than conventional assumptions suggest: an 8B parameter model (Ministral 3 8B) outperforms most larger models, while another 8B model (Llama 3.1 8B) performs worst overall.

Method: Three key methodological contributions: (1) reproduction of RHFL+ and benchmark algorithms under unified evaluation framework; (2) extension of RHFL+ to real-world medical imaging datasets (CBIS-DDSM, BreastMNIST, BHI); (3) novel implementation using NVFlare (NVIDIA’s production FL framework) for modular, scalable deployment.

Result: Extensive ablation studies, algorithmic comparisons under various noise conditions, and scalability experiments across increasing numbers of clients were conducted to validate effectiveness, though specific numerical results are not provided in the abstract.

Conclusion: The work provides a comprehensive evaluation of RHFL+’s robustness to class imbalances, extends it to practical medical applications, and delivers a production-ready implementation using NVFlare, advancing federated learning for real-world deployment scenarios.

Abstract: Federated Learning (FL) enables collaborative model training across decentralized devices while preserving data privacy. However, real-world FL deployments face critical challenges such as data imbalances, including label noise and non-IID distributions. RHFL+, a state-of-the-art method, was proposed to address these challenges in settings with heterogeneous client models. This work investigates the robustness of RHFL+ under class imbalances through three key contributions: (1) reproduction of RHFL+ along with all benchmark algorithms under a unified evaluation framework; (2) extension of RHFL+ to real-world medical imaging datasets, including CBIS-DDSM, BreastMNIST and BHI; (3) a novel implementation using NVFlare, NVIDIA’s production-level federated learning framework, enabling a modular, scalable and deployment-ready codebase. To validate effectiveness, extensive ablation studies, algorithmic comparisons under various noise conditions and scalability experiments across increasing numbers of clients are conducted.

Method: Proposes Assignment-Based Anticlustering (ABA) algorithm designed for Euclidean spaces where objects are D-dimensional feature vectors and distances are squared Euclidean distances. ABA uses an assignment-based approach that scales better than exchange-based heuristics like fast_anticlustering.

Result: ABA outperforms fast_anticlustering in both solution quality and running time, scales to instances with millions of objects and hundreds of thousands of anticlusters within short running times, and is superior to METIS for balanced K-cut partitioning in both solution quality and running time.

Conclusion: ABA provides a scalable, efficient solution for large-scale anticlustering problems, addressing limitations of existing methods and enabling applications with massive datasets and large numbers of clusters. The algorithm is publicly available on GitHub.

Abstract: The anticlustering problem is to partition a set of objects into K equal-sized anticlusters such that the sum of distances within anticlusters is maximized. The anticlustering problem is NP-hard. We focus on anticlustering in Euclidean spaces, where the input data is tabular and each object is represented as a D-dimensional feature vector. Distances are measured as squared Euclidean distances between the respective vectors. Applications of Euclidean anticlustering include social studies, particularly in psychology, K-fold cross-validation in which each fold should be a good representative of the entire dataset, the creation of mini-batches for gradient descent in neural network training, and balanced K-cut partitioning. In particular, machine-learning applications involve million-scale datasets and very large values of K, making scalable anticlustering algorithms essential. Existing algorithms are either exact methods that can solve only small instances or heuristic methods, among which the most scalable is the exchange-based heuristic fast_anticlustering. We propose a new algorithm, the Assignment-Based Anticlustering algorithm (ABA), which scales to very large instances. A computational study shows that ABA outperforms fast_anticlustering in both solution quality and running time. Moreover, ABA scales to instances with millions of objects and hundreds of thousands of anticlusters within short running times, beyond what fast_anticlustering can handle. As a balanced K-cut partitioning method for tabular data, ABA is superior to the well-known METIS method in both solution quality and running time. The code of the ABA algorithm is available on GitHub.

Method: Treats existing adapters in Transformer blocks (query, key, value, up, down projections) as implicit experts. Routes tokens among them using k-means clustering with exponentially moving averaged cluster centers, requiring no gradients or learned parameters.

Result: Achieves competitive performance with mixture-of-experts-based methods while using 7-29x fewer trainable parameters, up to 48% lower memory consumption, and 1.5-2x faster training across 14 text, 14 image, and 19 video benchmarks.

Conclusion: Monkey Jump provides an efficient, architecture-agnostic approach to achieve mixture-of-experts-style specialization in parameter-efficient fine-tuning without introducing additional trainable parameters, maintaining the core benefits of parameter efficiency.

Abstract: Mixture-of-experts variants of parameter-efficient fine-tuning enable per-token specialization, but they introduce additional trainable routers and expert parameters, increasing memory usage and training cost. This undermines the core goal of parameter-efficient fine-tuning. We propose Monkey Jump, a method that brings mixture-of-experts-style specialization to parameter-efficient fine-tuning without introducing extra trainable parameters for experts or routers. Instead of adding new adapters as experts, Monkey Jump treats the adapters already present in each Transformer block (such as query, key, value, up, and down projections) as implicit experts and routes tokens among them. Routing is performed using k-means clustering with exponentially moving averaged cluster centers, requiring no gradients and no learned parameters. We theoretically show that token-wise routing increases expressivity and can outperform shared adapters by avoiding cancellation effects. Across multi-task experiments covering 14 text, 14 image, and 19 video benchmarks, Monkey Jump achieves competitive performance with mixture-of-experts-based parameter-efficient fine-tuning methods while using 7 to 29 times fewer trainable parameters, up to 48 percent lower memory consumption, and 1.5 to 2 times faster training. Monkey Jump is architecture-agnostic and can be applied to any adapter-based parameter-efficient fine-tuning method.

Method: Combined TCGA RNA-seq data with STRING protein-protein interaction network, used Chebyshev graph convolutions (K=2) with weighted pooling layers that aggregate gene clusters into ‘supernodes’ across multiple coarsening levels, and applied saliency methods for interpretation.

Result: Increasing convolution and pooling layers did not enhance classification performance; highest F1-macro score (0.978) achieved with single pooling layer, with deeper layers causing over-smoothing and performance degradation. However, model proved highly interpretable, identifying known cancer-related genes and enriched biological processes.

Conclusion: While deeper GNN architectures didn’t improve performance, the hierarchical pooling structure provided valuable biological insights, making GNNs promising for cancer biomarker discovery and interpretation, with potential for developing new explainable architectures.

Abstract: This study explores the use of graph neural networks (GNNs) with hierarchical pooling and multiple convolution layers for cancer classification based on RNA-seq data. We combine gene expression data from The Cancer Genome Atlas (TCGA) with a precomputed STRING protein-protein interaction network to classify tissue origin and distinguish between normal and tumor samples. The model employs Chebyshev graph convolutions (K=2) and weighted pooling layers, aggregating gene clusters into ‘supernodes’ across multiple coarsening levels. This approach enables dimensionality reduction while preserving meaningful interactions. Saliency methods were applied to interpret the model by identifying key genes and biological processes relevant to cancer. Our findings reveal that increasing the number of convolution and pooling layers did not enhance classification performance. The highest F1-macro score (0.978) was achieved with a single pooling layer. However, adding more layers resulted in over-smoothing and performance degradation. However, the model proved highly interpretable through gradient methods, identifying known cancer-related genes and highlighting enriched biological processes, and its hierarchical structure can be used to develop new explainable architectures. Overall, while deeper GNN architectures did not improve performance, the hierarchical pooling structure provided valuable insights into tumor biology, making GNNs a promising tool for cancer biomarker discovery and interpretation

Method: Proposes a one-shot hierarchical FC framework with client-end distribution exploration and server-end distribution aggregation via one-way prototype-level communication. Uses fine partition mechanism to generate successive clusterlets describing complex client cluster landscapes, and multi-granular learning mechanism to fuse clusterlets with inconsistent granularities.

Result: The method efficiently explores complex cluster distributions across clients. Extensive experiments on ten public datasets demonstrate superiority over state-of-the-art methods.

Conclusion: The proposed one-shot hierarchical federated clustering framework effectively addresses challenges of exploring complex cluster distributions in heterogeneous federated environments while maintaining efficiency and privacy protection.

Abstract: Driven by the growth of Web-scale decentralized services, Federated Clustering (FC) aims to extract knowledge from heterogeneous clients in an unsupervised manner while preserving the clients’ privacy, which has emerged as a significant challenge due to the lack of label guidance and the Non-Independent and Identically Distributed (non-IID) nature of clients. In real scenarios such as personalized recommendation and cross-device user profiling, the global cluster may be fragmented and distributed among different clients, and the clusters may exist at different granularities or even nested. Although Hierarchical Clustering (HC) is considered promising for exploring such distributions, the sophisticated recursive clustering process makes it more computationally expensive and vulnerable to privacy exposure, thus relatively unexplored under the federated learning scenario. This paper introduces an efficient one-shot hierarchical FC framework that performs client-end distribution exploration and server-end distribution aggregation through one-way prototype-level communication from clients to the server. A fine partition mechanism is developed to generate successive clusterlets to describe the complex landscape of the clients’ clusters. Then, a multi-granular learning mechanism on the server is proposed to fuse the clusterlets, even when they have inconsistent granularities generated from different clients. It turns out that the complex cluster distributions across clients can be efficiently explored, and extensive experiments comparing state-of-the-art methods on ten public datasets demonstrate the superiority of the proposed method.

Method: Two-stage approach: 1) Fully optimize DLM’s generative ability, then freeze model; 2) Train specialized correction head using Future-Context Augmentation (FCA) that augments samples with ground-truth tokens to generalize error training distribution.

Result: DSC substantially reduces quality degradation with larger generation steps, allowing DLMs to achieve both high generation speed and strong output fidelity on mathematical reasoning and code generation benchmarks.

Conclusion: The decoupled self-correction framework enables models to jointly generate and revise tokens, effectively mitigating error accumulation in parallel multi-token generation while preserving peak SFT performance.

Abstract: Masked Diffusion Language Models (DLMs) achieve significant speed by generating multiple tokens in parallel. However, this parallel sampling approach, especially when using fewer inference steps, will introduce strong dependency errors and cause quality to deteriorate rapidly as the generation step size grows. As a result, reliable self-correction becomes essential for maintaining high-quality multi-token generation. To address this, we propose Decoupled Self-Correction (DSC), a novel two-stage methodology. DSC first fully optimizes the DLM’s generative ability before freezing the model and training a specialized correction head. This decoupling preserves the model’s peak SFT performance and ensures the generated errors used for correction head training are of higher quality. Additionally, we introduce Future-Context Augmentation (FCA) to maximize the correction head’s accuracy. FCA generalizes the error training distribution by augmenting samples with ground-truth tokens, effectively training the head to utilize a richer, future-looking context. This mechanism is used for reliably detecting the subtle errors of the high-fidelity base model. Our DSC framework enables the model, at inference time, to jointly generate and revise tokens, thereby correcting errors introduced by multi-token generation and mitigating error accumulation across steps. Experiments on mathematical reasoning and code generation benchmarks demonstrate that our approach substantially reduces the quality degradation associated with larger generation steps, allowing DLMs to achieve both high generation speed and strong output fidelity.

Method: 1) Shape-aware adapter that adaptively aggregates multiscale discriminative subsequences (shapes) into class tokens, selecting relevant subsequence scales for interpretability. 2) Prototype-based pretraining module to jointly learn instance- and shape-level representations for transferable shape patterns. 3) Pre-trained on large-scale multi-domain dataset (1.89M samples).

Result: Superior generalization across diverse target domains. Achieves state-of-the-art classification performance on 128 UCR datasets and 30 additional time series datasets. Interpretability and ablation analyses validate effectiveness.

Conclusion: UniShape bridges the gap between time series foundation models and classification tasks by incorporating shape-aware mechanisms for interpretable and transferable shape pattern learning, demonstrating strong performance and generalization capabilities.

Abstract: Foundation models pre-trained on large-scale source datasets are reshaping the traditional training paradigm for time series classification. However, existing time series foundation models primarily focus on forecasting tasks and often overlook classification-specific challenges, such as modeling interpretable shapelets that capture class-discriminative temporal features. To bridge this gap, we propose UniShape, a unified shape-aware foundation model designed for time series classification. UniShape incorporates a shape-aware adapter that adaptively aggregates multiscale discriminative subsequences (shapes) into class tokens, effectively selecting the most relevant subsequence scales to enhance model interpretability. Meanwhile, a prototype-based pretraining module is introduced to jointly learn instance- and shape-level representations, enabling the capture of transferable shape patterns. Pre-trained on a large-scale multi-domain time series dataset comprising 1.89 million samples, UniShape exhibits superior generalization across diverse target domains. Experiments on 128 UCR datasets and 30 additional time series datasets demonstrate that UniShape achieves state-of-the-art classification performance, with interpretability and ablation analyses further validating its effectiveness.

Method: Proposes a Newton-style update framework that quantifies client data influence propagation, uses curvature information of loss with respect to target data, approximates second-order information via Fisher information matrices, perturbs updates with calibrated noise, and broadcasts corrective updates through the network.

Result: The approach theoretically satisfies certified unlearning definition (unlearned model is difficult to distinguish from retrained model without deleted data) and maintains utility bounds close to retraining from scratch. Extensive experiments demonstrate effectiveness across diverse decentralized settings.

Conclusion: The proposed framework enables certified unlearning in decentralized federated learning by addressing the challenge of removing implicitly embedded client influence across distributed networks without centralized coordination.

Abstract: Driven by the right to be forgotten (RTBF), machine unlearning has become an essential requirement for privacy-preserving machine learning. However, its realization in decentralized federated learning (DFL) remains largely unexplored. In DFL, clients exchange local updates only with neighbors, causing model information to propagate and mix across the network. As a result, when a client requests data deletion, its influence is implicitly embedded throughout the system, making removal difficult without centralized coordination. We propose a novel certified unlearning framework for DFL based on Newton-style updates. Our approach first quantifies how a client’s data influence propagates during training. Leveraging curvature information of the loss with respect to the target data, we then construct corrective updates using Newton-style approximations. To ensure scalability, we approximate second-order information via Fisher information matrices. The resulting updates are perturbed with calibrated noise and broadcast through the network to eliminate residual influence across clients. We theoretically prove that our approach satisfies the formal definition of certified unlearning, ensuring that the unlearned model is difficult to distinguish from a retrained model without the deleted data. We also establish utility bounds showing that the unlearned model remains close to retraining from scratch. Extensive experiments across diverse decentralized settings demonstrate the effectiveness and efficiency of our framework.

Method: Uses Gumbel-Softmax trick to enable discrete yet differentiable selection among a predefined set of activation functions during training. The method dynamically learns the optimal activation function independently of the input.

Result: Experiments on synthetic datasets show the model consistently selects the most suitable activation function, demonstrating effectiveness in connecting theoretical advances with practical utility.

Conclusion: The framework paves the way for more adaptive and modular neural architectures in complex learning scenarios by enabling differentiable selection of optimal activation functions.

Abstract: Learning activation functions has emerged as a promising direction in deep learning, allowing networks to adapt activation mechanisms to task-specific demands. In this work, we introduce a novel framework that employs the Gumbel-Softmax trick to enable discrete yet differentiable selection among a predefined set of activation functions during training. Our method dynamically learns the optimal activation function independently of the input, thereby enhancing both predictive accuracy and architectural flexibility. Experiments on synthetic datasets show that our model consistently selects the most suitable activation function, underscoring its effectiveness. These results connect theoretical advances with practical utility, paving the way for more adaptive and modular neural architectures in complex learning scenarios.

Method: Extends policy-driven tree-based reinforcement learning to continuous spaces. Integrates population-level decision trees with surrogate-guided directional sampling, reward shaping, and hierarchical switching between global exploration and local exploitation.

Result: Superior or comparable performance to standard global optimization baselines on test functions. Successfully applied to crystal structure optimization, fitting interatomic potentials, and constrained engineering design with high fidelity and evaluation efficiency.

Conclusion: Physics-informed tree search establishes a scalable, interpretable paradigm for computational design, bridging discrete decision-making with continuous search in scientific workflows while preserving physical constraints.

Abstract: High-dimensional design spaces underpin a wide range of physics-based modeling and computational design tasks in science and engineering. These problems are commonly formulated as constrained black-box searches over rugged objective landscapes, where function evaluations are expensive, and gradients are unavailable or unreliable. Conventional global search engines and optimizers struggle in such settings due to the exponential scaling of design spaces, the presence of multiple local basins, and the absence of physical guidance in sampling. We present a physics-informed Monte Carlo Tree Search (MCTS) framework that extends policy-driven tree-based reinforcement concepts to continuous, high-dimensional scientific optimization. Our method integrates population-level decision trees with surrogate-guided directional sampling, reward shaping, and hierarchical switching between global exploration and local exploitation. These ingredients allow efficient traversal of non-convex, multimodal landscapes where physically meaningful optima are sparse. We benchmark our approach against standard global optimization baselines on a suite of canonical test functions, demonstrating superior or comparable performance in terms of convergence, robustness, and generalization. Beyond synthetic tests, we demonstrate physics-consistent applicability to (i) crystal structure optimization from clusters to bulk, (ii) fitting of classical interatomic potentials, and (iii) constrained engineering design problems. Across all cases, the method converges with high fidelity and evaluation efficiency while preserving physical constraints. Overall, our work establishes physics-informed tree search as a scalable and interpretable paradigm for computational design and high-dimensional scientific optimization, bridging discrete decision-making frameworks with continuous search in scientific design workflows.

Method: Gecko inherits Mega/Megalodon’s exponential moving average with gated attention, and adds three key components: timestep decay normalization, sliding chunk attention mechanism, and adaptive working memory to better capture long-range dependencies.

Result: In 7B parameter pretraining with 2 trillion tokens, Gecko achieves training loss of 1.68, outperforming Llama2-7B (1.75) and Megalodon-7B (1.70), close to Llama2-13B (1.67). It handles sequences up to 4 million tokens and retrieves information from contexts 4× longer than its attention window.

Conclusion: Gecko demonstrates superior efficiency and long-context scalability compared to existing models, with inherent long-context processing capabilities without requiring context-extension techniques.

Abstract: Designing a unified neural network to efficiently and inherently process sequential data with arbitrary lengths is a central and challenging problem in sequence modeling. The design choices in Transformer, including quadratic complexity and weak length extrapolation, have limited their ability to scale to long sequences. In this work, we propose Gecko, a neural architecture that inherits the design of Mega and Megalodon (exponential moving average with gated attention), and further introduces multiple technical components to improve its capability to capture long range dependencies, including timestep decay normalization, sliding chunk attention mechanism, and adaptive working memory. In a controlled pretraining comparison with Llama2 and Megalodon in the scale of 7 billion parameters and 2 trillion training tokens, Gecko achieves better efficiency and long-context scalability. Gecko reaches a training loss of 1.68, significantly outperforming Llama2-7B (1.75) and Megalodon-7B (1.70), and landing close to Llama2-13B (1.67). Notably, without relying on any context-extension techniques, Gecko exhibits inherent long-context processing and retrieval capabilities, stably handling sequences of up to 4 million tokens and retrieving information from contexts up to $4\times$ longer than its attention window. Code: https://github.com/XuezheMax/gecko-llm

Method: Formulates operator learning as a min-max optimization problem where the model is trained against worst-case input perturbations through adversarial training. Uses a robust self-supervised neural operator framework that enhances stability while preserving accuracy.

Result: The method achieves good performance on standard inputs while maintaining high fidelity under adversarial perturbed inputs. Demonstrates consistent performance under both normal and adversarial conditions.

Conclusion: Highlights the importance of stability-aware training in operator learning and provides a foundation for developing reliable neural PDE solvers for real-world applications where input noise and uncertainties are inevitable.

Abstract: Learning solution operators for differential equations with neural networks has shown great potential in scientific computing, but ensuring their stability under input perturbations remains a critical challenge. This paper presents a robust self-supervised neural operator framework that enhances stability through adversarial training while preserving accuracy. We formulate operator learning as a min-max optimization problem, where the model is trained against worst-case input perturbations to achieve consistent performance under both normal and adversarial conditions. We demonstrate that our method not only achieves good performance on standard inputs, but also maintains high fidelity under adversarial perturbed inputs. The results highlight the importance of stability-aware training in operator learning and provide a foundation for developing reliable neural PDE solvers in real-world applications, where input noise and uncertainties are inevitable.

Method: Use a single-layer encoder with log-sum-exp objective and InfoMax regularization for volume control. Analyze gradient-responsibility identity, variance effects, and decorrelation mechanisms.

Result: Gradient-responsibility identity holds exactly; LSE alone causes collapse; variance prevents dead components; decorrelation prevents redundancy. Model shows EM-like dynamics where lower loss doesn’t mean better features, and adaptive optimizers offer no advantage.

Conclusion: Implicit EM theory can prescribe effective architectures. The decoder-free model learns interpretable mixture components, confirming theoretical predictions about gradient descent on LSE objectives.

Abstract: Gradient descent on log-sum-exp (LSE) objectives performs implicit expectation–maximization (EM): the gradient with respect to each component output equals its responsibility. The same theory predicts collapse without volume control analogous to the log-determinant in Gaussian mixture models. We instantiate the theory in a single-layer encoder with an LSE objective and InfoMax regularization for volume control. Experiments confirm the theory’s predictions. The gradient–responsibility identity holds exactly; LSE alone collapses; variance prevents dead components; decorrelation prevents redundancy. The model exhibits EM-like optimization dynamics in which lower loss does not correspond to better features and adaptive optimizers offer no advantage. The resulting decoder-free model learns interpretable mixture components, confirming that implicit EM theory can prescribe architectures.

Method: ArenaRL shifts from pointwise scalar scoring to intra-group relative ranking using: 1) process-aware pairwise evaluation with multi-level rubrics for fine-grained relative scores, 2) intra-group adversarial arena with tournament-based ranking (seeded single-elimination scheme) that achieves O(N) complexity while maintaining accuracy comparable to O(N²) full pairwise comparisons.

Result: The seeded single-elimination scheme achieves nearly equivalent advantage estimation accuracy to full pairwise comparisons with only O(N) complexity. ArenaRL substantially outperforms standard RL baselines on new benchmarks Open-Travel and Open-DeepResearch, enabling LLM agents to generate more robust solutions for complex real-world tasks.

Conclusion: ArenaRL effectively addresses discrimination collapse in open-ended agent tasks by replacing pointwise scoring with efficient relative ranking, providing a better balance between efficiency and precision for RL training on complex tasks with vast solution spaces.

Abstract: Reinforcement learning has substantially improved the performance of LLM agents on tasks with verifiable outcomes, but it still struggles on open-ended agent tasks with vast solution spaces (e.g., complex travel planning). Due to the absence of objective ground-truth for these tasks, current RL algorithms largely rely on reward models that assign scalar scores to individual responses. We contend that such pointwise scoring suffers from an inherent discrimination collapse: the reward model struggles to distinguish subtle advantages among different trajectories, resulting in scores within a group being compressed into a narrow range. Consequently, the effective reward signal becomes dominated by noise from the reward model, leading to optimization stagnation. To address this, we propose ArenaRL, a reinforcement learning paradigm that shifts from pointwise scalar scoring to intra-group relative ranking. ArenaRL introduces a process-aware pairwise evaluation mechanism, employing multi-level rubrics to assign fine-grained relative scores to trajectories. Additionally, we construct an intra-group adversarial arena and devise a tournament-based ranking scheme to obtain stable advantage signals. Empirical results confirm that the built seeded single-elimination scheme achieves nearly equivalent advantage estimation accuracy to full pairwise comparisons with O(N^2) complexity, while operating with only O(N) complexity, striking an optimal balance between efficiency and precision. Furthermore, to address the lack of full-cycle benchmarks for open-ended agents, we build Open-Travel and Open-DeepResearch, two high-quality benchmarks featuring a comprehensive pipeline covering SFT, RL training, and multi-dimensional evaluation. Extensive experiments show that ArenaRL substantially outperforms standard RL baselines, enabling LLM agents to generate more robust solutions for complex real-world tasks.

Method: Combines multi-step H-Entropy Search with pathwise sampling and neural policy optimization. Uses neural policies (including large language models) to navigate structured combinatorial action spaces like protein sequences, amortizing lookahead planning while incorporating domain-specific constraints.

Result: Outperforms strong myopic and nonmyopic baselines across nine synthetic benchmarks (2-8 dimensions) and two real-world tasks: geospatial optimization using NASA night-light imagery and protein sequence design with constrained token-level edits.

Conclusion: LookaHES provides a general, scalable, cost-aware solution for robust long-horizon optimization in complex decision spaces, making it a useful tool for researchers in machine learning, statistics, and applied domains.

Abstract: Bayesian optimization (BO) is a common framework for optimizing black-box functions, yet most existing methods assume static query costs and rely on myopic acquisition strategies. We introduce LookaHES, a nonmyopic BO framework designed for dynamic, history-dependent cost environments, where evaluation costs vary with prior actions, such as travel distance in spatial tasks or edit distance in sequence design. LookaHES combines a multi-step variant of $H$-Entropy Search with pathwise sampling and neural policy optimization, enabling long-horizon planning beyond twenty steps without the exponential complexity of existing nonmyopic methods. The key innovation is the integration of neural policies, including large language models, to effectively navigate structured, combinatorial action spaces such as protein sequences. These policies amortize lookahead planning and can be integrated with domain-specific constraints during rollout. Empirically, LookaHES outperforms strong myopic and nonmyopic baselines across nine synthetic benchmarks from two to eight dimensions and two real-world tasks: geospatial optimization using NASA night-light imagery and protein sequence design with constrained token-level edits. In short, LookaHES provides a general, scalable, and cost-aware solution for robust long-horizon optimization in complex decision spaces, which makes it a useful tool for researchers in machine learning, statistics, and applied domains. Our implementation is available at https://github.com/sangttruong/nonmyopia.

Method: Uses Programmable Wireless Environments (PWE) to control RF wave propagation as an inspector entity. Creates an RF sensing pipeline with wavefront encoding to retrieve workpiece images from a database. Establishes correlations between RF wavefronts and industrial assets, then uses a Generative Adversarial Network (GAN) to generate visual representations matching database entries.

Result: Achieves 99.5% Structural Similarity Index Measure (SSIM) matching score in visual outputs, demonstrating high accuracy in generating visual representations from RF wavefront data.

Conclusion: The proposed RF-based inspection method shows promising results for next-generation quality control in industrial settings, enabling inspection in challenging environments where optical cameras are ineffective.

Abstract: Contemporary industrial Non-Destructive Inspection (NDI) methods require sensing capabilities that operate in occluded, hazardous, or access restricted environments. Yet, the current visual inspection based on optical cameras offers limited quality of service to that respect. In that sense, novel methods for workpiece inspection, suitable, for smart manufacturing are needed. Programmable Wireless Environments (PWE) could help towards that direction, by redefining the wireless Radio Frequency (RF) wave propagation as a controllable inspector entity. In this work, we propose a novel approach to Non-Destructive Inspection, leveraging an RF sensing pipeline based on RF wavefront encoding for retrieving workpiece-image entries from a designated database. This approach combines PWE-enabled RF wave manipulation with machine learning (ML) tools trained to produce visual outputs for quality inspection. Specifically, we establish correlation relationships between RF wavefronts and target industrial assets, hence yielding a dataset which links wavefronts to their corresponding images in a structured manner. Subsequently, a Generative Adversarial Network (GAN) derives visual representations closely matching the database entries. Our results indicate that the proposed method achieves an SSIM 99.5% matching score in visual outputs, paving the way for next-generation quality control workflows in industry.

Method: Proposes a model with two parallel modules: 1) wavelet-based convolutional kernels applied to overlapping patches of varying lengths to extract local-temporal dependencies under multi-frequency, and 2) dynamic cross-variable dependency capture under multi-frequency to model evolving inter-variable relationships across time-frequency domain.

Result: Outperforms state-of-the-art models on four representative electricity markets from Australia with varying renewable penetration levels. Ablation study validates complementary benefits of both modules. Built-in interpretability enables understanding of predictive behavior.

Conclusion: The proposed integrated approach effectively addresses key challenges in CIF forecasting by capturing complex temporal and cross-variable dependencies under multi-frequency patterns, while providing interpretable insights into model behavior.

Abstract: Accurate forecasting of the grid carbon intensity factor (CIF) is critical for enabling demand-side management and reducing emissions in modern electricity systems. Leveraging multiple interrelated time series, CIF prediction is typically formulated as a multivariate time series forecasting problem. Despite advances in deep learning-based methods, it remains challenging to capture the fine-grained local-temporal dependencies, dynamic higher-order cross-variable dependencies, and complex multi-frequency patterns for CIF forecasting. To address these issues, we propose a novel model that integrates two parallel modules: 1) one enhances the extraction of local-temporal dependencies under multi-frequency by applying multiple wavelet-based convolutional kernels to overlapping patches of varying lengths; 2) the other captures dynamic cross-variable dependencies under multi-frequency to model how inter-variable relationships evolve across the time-frequency domain. Evaluations on four representative electricity markets from Australia, featuring varying levels of renewable penetration, demonstrate that the proposed method outperforms the state-of-the-art models. An ablation study further validates the complementary benefits of the two proposed modules. Designed with built-in interpretability, the proposed model also enables better understanding of its predictive behavior, as shown in a case study where it adaptively shifts attention to relevant variables and time intervals during a disruptive event.

Method: Multi-Frequency-Reconstruction-based Diffusion (MFRD) model with four key steps: 1) Combine original data with decomposed multi-frequency modes for new data representation, 2) Diffusion forward process adds noise to reduce original data noise, 3) Reverse process uses LSTM-Transformer hybrid denoising network, 4) Inference generates predictions from trained network.

Result: Experimental results on AEMO and ISO-NE datasets show MFRD consistently outperforms compared models, demonstrating effectiveness for short-term electricity load forecasting.

Conclusion: The proposed MFRD model successfully applies diffusion models to STELF by addressing data characteristics through multi-frequency reconstruction and hybrid denoising, achieving superior forecasting performance compared to existing methods.

Abstract: Diffusion models have emerged as a powerful method in various applications. However, their application to Short-Term Electricity Load Forecasting (STELF) – a typical scenario in energy systems – remains largely unexplored. Considering the nonlinear and fluctuating characteristics of the load data, effectively utilizing the powerful modeling capabilities of diffusion models to enhance STELF accuracy remains a challenge. This paper proposes a novel diffusion model with multi-frequency reconstruction for STELF, referred to as the Multi-Frequency-Reconstruction-based Diffusion (MFRD) model. The MFRD model achieves accurate load forecasting through four key steps: (1) The original data is combined with the decomposed multi-frequency modes to form a new data representation; (2) The diffusion model adds noise to the new data, effectively reducing and weakening the noise in the original data; (3) The reverse process adopts a denoising network that combines Long Short-Term Memory (LSTM) and Transformer to enhance noise removal; and (4) The inference process generates the final predictions based on the trained denoising network. To validate the effectiveness of the MFRD model, we conducted experiments on two data platforms: Australian Energy Market Operator (AEMO) and Independent System Operator of New England (ISO-NE). The experimental results show that our model consistently outperforms the compared models.

Method: Mosaic shifts from local, static memory management to global, dynamic paradigm with three key components: 1) mask-only logits kernel to eliminate redundancy, 2) lazy chunking optimizer using online heuristic search to adaptively mitigate dynamic peaks, and 3) global memory manager using virtual addressing to resolve fragmentation.

Result: Mosaic achieves average 2.71× reduction in memory peak-to-average ratio, increases maximum inference sequence length support by 15.89-32.98× on identical hardware, reduces latency by 4.12%-23.26%, all without compromising accuracy.

Conclusion: Mosaic effectively addresses the memory inefficiencies in dLLM inference through global, dynamic memory management, enabling practical deployment of diffusion-based LLMs for long-context generation tasks.

Abstract: Diffusion-based large language models (dLLMs) have emerged as a promising paradigm, utilizing simultaneous denoising to enable global planning and iterative refinement. While these capabilities are particularly advantageous for long-context generation, deploying such models faces a prohibitive memory capacity barrier stemming from severe system inefficiencies. We identify that existing inference systems are ill-suited for this paradigm: unlike autoregressive models constrained by the cumulative KV-cache, dLLMs are bottlenecked by transient activations recomputed at every step. Furthermore, general-purpose memory reuse mechanisms lack the global visibility to adapt to dLLMs’ dynamic memory peaks, which toggle between logits and FFNs. To address these mismatches, we propose Mosaic, a memory-efficient inference system that shifts from local, static management to a global, dynamic paradigm. Mosaic integrates a mask-only logits kernel to eliminate redundancy, a lazy chunking optimizer driven by an online heuristic search to adaptively mitigate dynamic peaks, and a global memory manager to resolve fragmentation via virtual addressing. Extensive evaluations demonstrate that Mosaic achieves an average 2.71$\times$ reduction in the memory peak-to-average ratio and increases the maximum inference sequence length supportable on identical hardware by 15.89-32.98$\times$. This scalability is achieved without compromising accuracy and speed, and in fact reducing latency by 4.12%-23.26%.

Method: Proposes HELVAE using Hellinger distance-based opinion pooling derived from Hölder pooling with α=0.5, avoiding sub-sampling and enabling efficient multimodal inference.

Result: HELVAE learns more expressive latent representations with additional modalities and achieves better trade-offs between generative coherence and quality than state-of-the-art multimodal VAEs.

Conclusion: Probabilistic opinion pooling with Hellinger distance provides an effective alternative to traditional PoE/MoE approaches for multimodal VAEs, improving both representation learning and generative performance.

Abstract: Multimodal variational autoencoders (VAEs) are widely used for weakly supervised generative learning with multiple modalities. Predominant methods aggregate unimodal inference distributions using either a product of experts (PoE), a mixture of experts (MoE), or their combinations to approximate the joint posterior. In this work, we revisit multimodal inference through the lens of probabilistic opinion pooling, an optimization-based approach. We start from Hölder pooling with $α=0.5$, which corresponds to the unique symmetric member of the $α\text{-divergence}$ family, and derive a moment-matching approximation, termed Hellinger. We then leverage such an approximation to propose HELVAE, a multimodal VAE that avoids sub-sampling, yielding an efficient yet effective model that: (i) learns more expressive latent representations as additional modalities are observed; and (ii) empirically achieves better trade-offs between generative coherence and quality, outperforming state-of-the-art multimodal VAE models.

Method: The authors introduce soft symmetry-respecting inductive biases that create approximate degeneracies (pseudo-Goldstone modes) in the loss landscape. They quantify functional complexity using first-principles Hessian analysis and compressibility metrics, applied to HEP classification tasks.

Result: The soft symmetry bias creates pseudo-Goldstone modes in the loss landscape. Solutions with lower functional complexity (measured via Hessian analysis and compressibility) produce abstractions that are more generalizable, robust to input perturbations, and efficiently distillable.

Conclusion: Inductive biases that respect soft symmetries can shape loss geometry to favor lower-complexity solutions, which in turn yield more generalizable, robust, and efficiently distillable abstractions for HEP applications.

Abstract: Learning robust and generalisable abstractions from high-dimensional input data is a central challenge in machine learning and its applications to high-energy physics (HEP). Solutions of lower functional complexity are known to produce abstractions that generalise more effectively and are more robust to input perturbations. In complex hypothesis spaces, inductive biases make such solutions learnable by shaping the loss geometry during optimisation. In a HEP classification task, we show that a soft symmetry respecting inductive bias creates approximate degeneracies in the loss, which we identify as pseudo-Goldstone modes. We quantify functional complexity using metrics derived from first principles Hessian analysis and via compressibility. Our results demonstrate that solutions of lower complexity give rise to abstractions that are more generalisable, robust, and efficiently distillable.

Method: The authors develop a geometric framework analyzing learning dynamics in quotient spaces after factoring out model symmetries. They show stochastic differential equations gain geometric corrections favoring orbits with small local volume. The approach is constructive: given symmetries, one can derive implicit bias; conversely, one can inverse-design biases by computing specific redundant parameterizations.

Result: The framework unifies several known implicit bias results under a single geometric perspective. It provides practical methodology for deriving implicit biases in new settings and yields testable predictions confirmed by numerical simulations on toy models with synthetic data.

Conclusion: The paper establishes a general geometric mechanism for implicit bias arising from interaction between model symmetries and stochastic optimization. The framework enables both analysis of existing biases and inverse-design of new ones, with applications demonstrated for sparsity and spectral biases.

Abstract: A central problem in machine learning theory is to characterize how learning dynamics select particular solutions among the many compatible with the training objective, a phenomenon, called implicit bias, which remains only partially characterized. In the present work, we identify a general mechanism, in terms of an explicit geometric correction of the learning dynamics, for the emergence of implicit biases, arising from the interaction between continuous symmetries in the model’s parametrization and stochasticity in the optimization process. Our viewpoint is constructive in two complementary directions: given model symmetries, one can derive the implicit bias they induce; conversely, one can inverse-design a wide class of different implicit biases by computing specific redundant parameterizations. More precisely, we show that, when the dynamics is expressed in the quotient space obtained by factoring out the symmetry group of the parameterization, the resulting stochastic differential equation gains a closed form geometric correction in the stationary distribution of the optimizer dynamics favoring orbits with small local volume. We compute the resulting symmetry induced bias for a range of architectures, showing how several well known results fit into a single unified framework. The approach also provides a practical methodology for deriving implicit biases in new settings, and it yields concrete, testable predictions that we confirm by numerical simulations on toy models trained on synthetic data, leaving more complex scenarios for future work. Finally, we test the implicit bias inverse-design procedure in notable cases, including biases toward sparsity in linear features or in spectral properties of the model parameters.

Method: Uses DS-specific structured prompts with input fields to guide agentic system. Implements separate LLM agents for planning and coding that generate interleaved plan/code blocks. Keeps data local using function calls, only injecting aggregate statistics into prompts. Features iterative code generation and smart history rendering for fault tolerance.

Result: Demonstrated viability using canonical Kaggle challenges, showing the system can effectively automate data science workflows while managing context and computational constraints.

Conclusion: CEDAR successfully addresses key challenges in LLM-based data science automation through context engineering, agentic design, local data processing, and fault-tolerant workflows, proving the feasibility of automated data science agents.

Abstract: We demonstrate CEDAR, an application for automating data science (DS) tasks with an agentic setup. Solving DS problems with LLMs is an underexplored area that has immense market value. The challenges are manifold: task complexities, data sizes, computational limitations, and context restrictions. We show that these can be alleviated via effective context engineering. We first impose structure into the initial prompt with DS-specific input fields, that serve as instructions for the agentic system. The solution is then materialized as an enumerated sequence of interleaved plan and code blocks generated by separate LLM agents, providing a readable structure to the context at any step of the workflow. Function calls for generating these intermediate texts, and for corresponding Python code, ensure that data stays local, and only aggregate statistics and associated instructions are injected into LLM prompts. Fault tolerance and context management are introduced via iterative code generation and smart history rendering. The viability of our agentic data scientist is demonstrated using canonical Kaggle challenges.

Method: KASER uses reinforcement learning with a hybrid reward function that evaluates three aspects: 1) code similarity to ground-truth, 2) error matching, and 3) code prediction diversity to align errors with student knowledge.

Result: On two real-world datasets, KASER outperforms baselines at both per-student-problem pair level (code and error prediction) and per-problem level (error coverage and simulated code diversity).

Conclusion: KASER effectively addresses LLM limitations in simulating diverse student coding errors by knowledge-aligned reinforcement learning, improving error prediction and coverage in educational contexts.

Abstract: Open-ended tasks, such as coding problems that are common in computer science education, provide detailed insights into student knowledge. However, training large language models (LLMs) to simulate and predict possible student errors in their responses to these problems can be challenging: they often suffer from mode collapse and fail to fully capture the diversity in syntax, style, and solution approach in student responses. In this work, we present KASER (Knowledge-Aligned Student Error Simulator), a novel approach that aligns errors with student knowledge. We propose a training method based on reinforcement learning using a hybrid reward that reflects three aspects of student code prediction: i) code similarity to the ground-truth, ii) error matching, and iii) code prediction diversity. On two real-world datasets, we perform two levels of evaluation and show that: At the per-student-problem pair level, our method outperforms baselines on code and error prediction; at the per-problem level, our method outperforms baselines on error coverage and simulated code diversity.

Method: Proposes PromptMIA, where a malicious server inserts adversarially crafted prompts and monitors their updates during collaborative training to infer membership. Formalizes the threat as a security game and provides theoretical analysis establishing a lower bound on attack advantage.

Result: PromptMIA consistently achieves high advantage across diverse benchmark datasets. Standard membership inference defenses developed for gradient/output-based attacks show limited effectiveness against PromptMIA, highlighting non-trivial challenges.

Conclusion: Federated prompt-tuning creates new privacy vulnerabilities requiring specifically tailored defense strategies, as current defenses are insufficient against prompt-based attacks like PromptMIA.

Abstract: Membership inference attack (MIA) poses a significant privacy threat in federated learning (FL) as it allows adversaries to determine whether a client’s private dataset contains a specific data sample. While defenses against membership inference attacks in standard FL have been well studied, the recent shift toward federated fine-tuning has introduced new, largely unexplored attack surfaces. To highlight this vulnerability in the emerging FL paradigm, we demonstrate that federated prompt-tuning, which adapts pre-trained models with small input prefixes to improve efficiency, also exposes a new vector for privacy attacks. We propose PromptMIA, a membership inference attack tailored to federated prompt-tuning, in which a malicious server can insert adversarially crafted prompts and monitors their updates during collaborative training to accurately determine whether a target data point is in a client’s private dataset. We formalize this threat as a security game and empirically show that PromptMIA consistently attains high advantage in this game across diverse benchmark datasets. Our theoretical analysis further establishes a lower bound on the attack’s advantage which explains and supports the consistently high advantage observed in our empirical results. We also investigate the effectiveness of standard membership inference defenses originally developed for gradient or output based attacks and analyze their interaction with the distinct threat landscape posed by PromptMIA. The results highlight non-trivial challenges for current defenses and offer insights into their limitations, underscoring the need for defense strategies that are specifically tailored to prompt-tuning in federated settings.

Method: Used repeated-measures experimental design with constant GPU instance, identical TinyLlama model architecture (1.1B parameters), optimizer settings, and epoch counts. Trained at three token counts (500K, 1M, 2M) while measuring power consumption and execution duration.

Result: Conventional performance metrics showed inconsistent/diminishing returns, but including power consumption revealed strictly monotonic decline in training efficiency as token count increased. Repeated-measures ANOVA showed strong effect of token count on parameter efficiency, with all pairwise comparisons significant after Bonferroni correction.

Conclusion: Increasing training token counts may be energetically inefficient even with marginal performance improvements, emphasizing the need for efficiency-aware evaluation in LLM training.

Abstract: Research in machine learning has questioned whether increases in training token counts reliably produce proportional performance gains in large language models. Building on prior work introducing an energy-aware parameter efficiency metric, this study empirically examines the effects of increasing training token counts under fixed hardware and training conditions. The significance of this work lies in the explicit integration of power consumption and execution duration, as reflected by the power sampling frequency, into token-scale analysis. This addresses a gap in prior studies emphasizing performance outcomes while underrepresenting computational and energy costs. Using a repeated-measures experimental design on a constant GPU instance with an identical model architecture, optimizer settings, and epoch counts, a 1.1-billion-parameter TinyLlama model was trained at three token counts (500K, 1M, and 2M). While conventional performance metrics exhibited inconsistent or diminishing returns across token scales, the inclusion of power consumption and execution duration revealed a strictly monotonic decline in training efficiency as token count increased. Repeated-measures ANOVA demonstrated a strong effect of token count on parameter efficiency, with all pairwise comparisons remaining significant following Bonferroni correction. These findings indicate that increases in training token counts may be energetically inefficient even when marginal performance improvements are observed, underscoring the importance of efficiency-aware evaluation in large language model training.

Method: Reinforcement Learning-guided Dynamic Multi-Graph Fusion (RL-DMF) with dynamic multi-graph fusion module and RL-based intelligent feature selection and ranking (RL-IFSR) to mask irrelevant features.

Result: 95% accuracy (RMSE=293.9) for 1-hour traffic flow prediction on unseen hurricane Milton (2024), and 90% accuracy (RMSE=426.4) for 6-hour forecasts, outperforming state-of-the-art models.

Conclusion: RL-DMF provides a generalized, interpretable model for real-time evacuation traffic forecasting with significant implications for evacuation traffic management during hurricanes.

Abstract: Real-time traffic prediction is critical for managing transportation systems during hurricane evacuations. Although data-driven graph-learning models have demonstrated strong capabilities in capturing the complex spatiotemporal dynamics of evacuation traffic at a network level, they mostly consider a single dimension (e.g., travel-time or distance) to construct the underlying graph. Furthermore, these models often lack interpretability, offering little insight into which input variables contribute most to their predictive performance. To overcome these limitations, we develop a novel Reinforcement Learning-guided Dynamic Multi-Graph Fusion (RL-DMF) framework for evacuation traffic prediction. We construct multiple dynamic graphs at each time step to represent heterogeneous spatiotemporal relationships between traffic detectors. A dynamic multi-graph fusion (DMF) module is employed to adaptively learn and combine information from these graphs. To enhance model interpretability, we introduce RL-based intelligent feature selection and ranking (RL-IFSR) method that learns to mask irrelevant features during model training. The model is evaluated using a real-world dataset of 12 hurricanes affecting Florida from 2016 to 2024. For an unseen hurricane (Milton, 2024), the model achieves a 95% accuracy (RMSE = 293.9) for predicting the next 1-hour traffic flow. Moreover, the model can forecast traffic flow for up to next 6 hours with 90% accuracy (RMSE = 426.4). The RL-DMF framework outperforms several state-of-the-art traffic prediction models. Furthermore, ablation experiments confirm the effectiveness of dynamic multi-graph fusion and RL-IFSR approaches for improving model performance. This research provides a generalized and interpretable model for real-time evacuation traffic forecasting, with significant implications for evacuation traffic management.

Method: Train models on single A40 GPU (48GB) for under 24 hours using Reinforcement Learning with Verifiable Rewards (RLVR) and Low-Rank Adaptation (LoRA), testing different adapter ranks (r=8 vs r=256) on various model initializations.

Result: High-rank adapters (r=256) unlock significant plasticity in standard instruction-tuned models, achieving 40.0% Pass@1 on AIME 24 (11.1% absolute improvement) and 70.0% Pass@16. However, heavily math-aligned models suffered performance collapse due to destructive interference from noisy RL updates.

Conclusion: Micro-budget training can induce strong reasoning in small models, but success depends critically on adapter capacity and model initialization. Standard instruction-tuned models benefit from plasticity, while specialized models near task optima are vulnerable to destructive interference from low-budget RL updates.

Abstract: Recent advances in mathematical reasoning typically rely on massive scale, yet the question remains: can strong reasoning capabilities be induced in small language models ($\leq1.5\text{B}$) under extreme constraints? We investigate this by training models on a single A40 GPU (48GB) for under 24 hours using Reinforcement Learning with Verifiable Rewards (RLVR) and Low-Rank Adaptation (LoRA). We find that the success of this ``micro-budget" regime depends critically on the interplay between adapter capacity and model initialization. While low-rank adapters ($r=8$) consistently fail to capture the complex optimization dynamics of reasoning, high-rank adapters ($r=256$) unlock significant plasticity in standard instruction-tuned models. Our best result achieved an impressive 40.0% Pass@1 on AIME 24 (an 11.1% absolute improvement over baseline) and pushed Pass@16 to 70.0%, demonstrating robust exploration capabilities. However, this plasticity is not universal: while instruction-tuned models utilized the budget to elongate their chain-of-thought and maximize reward, heavily math-aligned models suffered performance collapse, suggesting that noisy, low-budget RL updates can act as destructive interference for models already residing near a task-specific optimum.

Method: ExCIR (Explainability through Correlation Impact Ratio) uses a theoretically grounded, simple metric that captures dependencies from correlated features through a lightweight single-pass formulation. It’s extended with an information theoretic foundation unifying correlation ratio with Canonical Correlation Analysis under mutual information bounds.

Result: Experimental evaluations on diverse datasets (EEG, synthetic vehicular data, Digits, Cats-Dogs) show ExCIR is effective, stable across domains, achieves more interpretable feature explanations than existing methods, and remains computationally efficient.

Conclusion: ExCIR provides a reliable, scalable explainability method that properly handles correlated features, offering stable and consistent explanations under noise and sampling variations while enabling multi-output and class-conditioned explainability at scale.

Abstract: Complex AI systems make better predictions but often lack transparency, limiting trustworthiness, interpretability, and safe deployment. Common post hoc AI explainers, such as LIME, SHAP, HSIC, and SAGE, are model agnostic but are too restricted in one significant regard: they tend to misrank correlated features and require costly perturbations, which do not scale to high dimensional data. We introduce ExCIR (Explainability through Correlation Impact Ratio), a theoretically grounded, simple, and reliable metric for explaining the contribution of input features to model outputs, which remains stable and consistent under noise and sampling variations. We demonstrate that ExCIR captures dependencies arising from correlated features through a lightweight single pass formulation. Experimental evaluations on diverse datasets, including EEG, synthetic vehicular data, Digits, and Cats-Dogs, validate the effectiveness and stability of ExCIR across domains, achieving more interpretable feature explanations than existing methods while remaining computationally efficient. To this end, we further extend ExCIR with an information theoretic foundation that unifies the correlation ratio with Canonical Correlation Analysis under mutual information bounds, enabling multi output and class conditioned explainability at scale.

Method: Proposes a trustworthiness test using stochastic proof-by-contradiction: trains and tests models on carefully permuted spurious labels based on potential outcomes framework. A trustworthy model should fail under permutation; comparable accuracy across real and permuted labels indicates problems. Quantifies behavior through interpretable Fisher-style p-values.

Result: Evaluated on multiple bacterial diagnostics to separate tasks/models learning genuine causal relationships from those driven by dataset artifacts or statistical coincidences.

Conclusion: Establishes foundation to build rigor and trust between ML and life-science communities, moving ML models closer to clinical adoption.

Abstract: Machine learning (ML) models show strong promise for new biomedical prediction tasks, but concerns about trustworthiness have hindered their clinical adoption. In particular, it is often unclear whether a model relies on true clinical cues or on spurious hierarchical correlations in the data. This paper introduces a simple yet broadly applicable trustworthiness test grounded in stochastic proof-by-contradiction. Instead of just showing high test performance, our approach trains and tests on spurious labels carefully permuted based on a potential outcomes framework. A truly trustworthy model should fail under such label permutation; comparable accuracy across real and permuted labels indicates overfitting, shortcut learning, or data leakage. Our approach quantifies this behavior through interpretable Fisher-style p-values, which are well understood by domain experts across medical and life sciences. We evaluate our approach on multiple new bacterial diagnostics to separate tasks and models learning genuine causal relationships from those driven by dataset artifacts or statistical coincidences. Our work establishes a foundation to build rigor and trust between ML and life-science research communities, moving ML models one step closer to clinical adoption.

Method: Proposes a framework integrating heterogeneous graph deep learning models with traditional ML algorithms for comparison. Uses graph metapath structure and incorporates dynamic assessment features that progressively influence student success prediction. Applied to Open University Learning Analytics (OULA) dataset.

Result: Achieved 68.6% validation F1 score with only 7% of semester completed, reaching up to 89.5% near semester’s end. Outperformed top machine learning models by 4.7% in validation F1 score during critical early 7% of semester.

Conclusion: The approach demonstrates the value of dynamic features and heterogeneous graph representations in student success prediction, particularly for early intervention when only limited semester data is available.

Abstract: Early identification of student success is crucial for enabling timely interventions, reducing dropout rates, and promoting on time graduation. In educational settings, AI powered systems have become essential for predicting student performance due to their advanced analytical capabilities. However, effectively leveraging diverse student data to uncover latent and complex patterns remains a key challenge. While prior studies have explored this area, the potential of dynamic data features and multi category entities has been largely overlooked. To address this gap, we propose a framework that integrates heterogeneous graph deep learning models to enhance early and continuous student performance prediction, using traditional machine learning algorithms for comparison. Our approach employs a graph metapath structure and incorporates dynamic assessment features, which progressively influence the student success prediction task. Experiments on the Open University Learning Analytics (OULA) dataset demonstrate promising results, achieving a 68.6% validation F1 score with only 7% of the semester completed, and reaching up to 89.5% near the semester’s end. Our approach outperforms top machine learning models by 4.7% in validation F1 score during the critical early 7% of the semester, underscoring the value of dynamic features and heterogeneous graph representations in student success prediction.

Method: Synthesizes insights from machine learning, statistics, and optimization literature to organize causes into three categories: statistical sources (finite samples, data noise), structural sources (non-convex objectives, unobserved variables), and procedural sources (algorithm limitations, model class restrictions).

Result: Identifies key distinctions: statistical multiplicity diminishes with more data, structural multiplicity persists asymptotically and requires different data/assumptions to resolve, and procedural multiplicity reflects practitioner choices. Provides a framework for understanding when and why multiple good models exist.

Conclusion: The Rashomon effect has diverse causes with different implications - statistical issues can be addressed with more data, structural issues are fundamental and require different approaches, while procedural issues reflect design choices. Understanding these distinctions is crucial for addressing challenges and leveraging opportunities in inference, interpretability, fairness, and decision-making.

Abstract: The Rashomon effect – the existence of multiple, distinct models that achieve nearly equivalent predictive performance – has emerged as a fundamental phenomenon in modern machine learning and statistics. In this paper, we explore the causes underlying the Rashomon effect, organizing them into three categories: statistical sources arising from finite samples and noise in the data-generating process; structural sources arising from non-convexity of optimization objectives and unobserved variables that create fundamental non-identifiability; and procedural sources arising from limitations of optimization algorithms and deliberate restrictions to suboptimal model classes. We synthesize insights from machine learning, statistics, and optimization literature to provide a unified framework for understanding why the multiplicity of good models arises. A key distinction emerges: statistical multiplicity diminishes with more data, structural multiplicity persists asymptotically and cannot be resolved without different data or additional assumptions, and procedural multiplicity reflects choices made by practitioners. Beyond characterizing causes, we discuss both the challenges and opportunities presented by the Rashomon effect, including implications for inference, interpretability, fairness, and decision-making under uncertainty.

Method: DP-FedEPC combines elastic weight consolidation (EWC) to constrain updates on important parameters, prototype-based rehearsal to preserve class structure without storing raw images, and client-side differential privacy (DP-SGD) with calibrated Gaussian noise to clipped gradients within a standard FedAvg framework.

Result: The method provides formal privacy guarantees for individual radiographs while enabling continual learning across temporally evolving hospital data streams without catastrophic forgetting.

Conclusion: DP-FedEPC addresses the critical challenges of privacy-preserving federated continual learning for medical imaging by combining EWC, prototype rehearsal, and differential privacy, making it suitable for real-world clinical settings where data evolves over time.

Abstract: Deep learning models for radiology interpretation increasingly rely on multi-institutional data, yet privacy regulations and distribution shift across hospitals limit central data pooling. Federated learning (FL) allows hospitals to collaboratively train models without sharing raw images, but current FL algorithms typically assume a static data distribution. In practice, hospitals experience continual evolution in case mix, annotation protocols, and imaging devices, which leads to catastrophic forgetting when models are updated sequentially. Federated continual learning (FCL) aims to reconcile these challenges but existing methods either ignore the stringent privacy constraints of healthcare or rely on replay buffers and public surrogate datasets that are difficult to justify in clinical settings. We study FCL for chest radiography classification in a setting where hospitals are clients that receive temporally evolving streams of cases and labels. We introduce DP-FedEPC (Differentially Private Federated Elastic Prototype Consolidation), a method that combines elastic weight consolidation (EWC), prototype-based rehearsal, and client-side differential privacy within a standard FedAvg framework. EWC constrains updates along parameters deemed important for previous tasks, while a memory of latent prototypes preserves class structure without storing raw images. Differentially private stochastic gradient descent (DP-SGD) at each client adds calibrated Gaussian noise to clipped gradients, providing formal privacy guarantees for individual radiographs.

Method: Control Optimal Lie-Poisson Neural Networks (CO-LPNets) learn phase-space dynamics by composing Poisson maps obtained as flows from Hamiltonians that can be integrated explicitly. The method preserves Poisson brackets and Casimirs to machine precision.