Method: Created WearVox benchmark with 3,842 multi-channel egocentric audio recordings collected via AI glasses across five diverse tasks: Search-Grounded QA, Closed-Book QA, Side-Talk Rejection, Tool Calling, and Speech Translation. Recordings span indoor/outdoor environments with varied acoustic conditions, accompanied by rich metadata for nuanced analysis.

Result: Benchmarked leading SLLMs showing 29-59% accuracy on WearVox, with substantial performance degradation on noisy outdoor audio. Case study with new SLLMs demonstrated multi-channel audio inputs significantly enhance robustness to environmental noise and improve discrimination between device-directed and background speech.

Conclusion: WearVox establishes a comprehensive testbed for wearable voice AI research, highlighting the critical importance of spatial audio cues for context-aware voice assistants and revealing significant gaps in current SLLM performance for real-world wearable scenarios.

Abstract: Wearable devices such as AI glasses are transforming voice assistants into always-available, hands-free collaborators that integrate seamlessly with daily life, but they also introduce challenges like egocentric audio affected by motion and noise, rapid micro-interactions, and the need to distinguish device-directed speech from background conversations. Existing benchmarks largely overlook these complexities, focusing instead on clean or generic conversational audio. To bridge this gap, we present WearVox, the first benchmark designed to rigorously evaluate voice assistants in realistic wearable scenarios. WearVox comprises 3,842 multi-channel, egocentric audio recordings collected via AI glasses across five diverse tasks including Search-Grounded QA, Closed-Book QA, Side-Talk Rejection, Tool Calling, and Speech Translation, spanning a wide range of indoor and outdoor environments and acoustic conditions. Each recording is accompanied by rich metadata, enabling nuanced analysis of model performance under real-world constraints. We benchmark leading proprietary and open-source speech Large Language Models (SLLMs) and find that most real-time SLLMs achieve accuracies on WearVox ranging from 29% to 59%, with substantial performance degradation on noisy outdoor audio, underscoring the difficulty and realism of the benchmark. Additionally, we conduct a case study with two new SLLMs that perform inference with single-channel and multi-channel audio, demonstrating that multi-channel audio inputs significantly enhance model robustness to environmental noise and improve discrimination between device-directed and background speech. Our results highlight the critical importance of spatial audio cues for context-aware voice assistants and establish WearVox as a comprehensive testbed for advancing wearable voice AI research.

Method: EvoRoute is a self-evolving model routing paradigm that uses an expanding knowledge base of prior experience to dynamically select Pareto-optimal LLM backbones at each step, balancing accuracy, efficiency, and resource use while refining its selection policy through environment feedback.

Result: Experiments on GAIA and BrowseComp+ benchmarks show EvoRoute sustains or enhances system performance while reducing execution cost by up to 80% and latency by over 70% when integrated into off-the-shelf agentic systems.

Conclusion: EvoRoute successfully addresses the Agent System Trilemma by providing a dynamic, self-evolving routing solution that optimizes the trade-off between performance, cost, and latency in complex agentic AI systems.

Abstract: Complex agentic AI systems, powered by a coordinated ensemble of Large Language Models (LLMs), tool and memory modules, have demonstrated remarkable capabilities on intricate, multi-turn tasks. However, this success is shadowed by prohibitive economic costs and severe latency, exposing a critical, yet underexplored, trade-off. We formalize this challenge as the \textbf{Agent System Trilemma}: the inherent tension among achieving state-of-the-art performance, minimizing monetary cost, and ensuring rapid task completion. To dismantle this trilemma, we introduce EvoRoute, a self-evolving model routing paradigm that transcends static, pre-defined model assignments. Leveraging an ever-expanding knowledge base of prior experience, EvoRoute dynamically selects Pareto-optimal LLM backbones at each step, balancing accuracy, efficiency, and resource use, while continually refining its own selection policy through environment feedback. Experiments on challenging agentic benchmarks such as GAIA and BrowseComp+ demonstrate that EvoRoute, when integrated into off-the-shelf agentic systems, not only sustains or enhances system performance but also reduces execution cost by up to $80%$ and latency by over $70%$.

Method: PCEval introduces a fully automatic evaluation framework that assesses LLMs in generating circuits and producing compatible code across varying project complexity levels. It uses simulation environments to test 13 leading models without requiring human assessment, enabling reproducible and automatically validated empirical assessment.

Result: LLMs perform well in code generation and logical circuit design but struggle significantly with physical breadboard layout creation, particularly in managing proper pin connections and avoiding circuit errors.

Conclusion: PCEval advances understanding of AI assistance in hardware-dependent computing environments and establishes a foundation for developing more effective tools to support physical computing education, revealing important limitations in LLMs’ hardware reasoning capabilities.

Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities across various domains, including software development, education, and technical assistance. Among these, software development is one of the key areas where LLMs are increasingly adopted. However, when hardware constraints are considered-for instance, in physical computing, where software must interact with and control physical hardware -their effectiveness has not been fully explored. To address this gap, we introduce \textsc{PCEval} (Physical Computing Evaluation), the first benchmark in physical computing that enables a fully automatic evaluation of the capabilities of LLM in both the logical and physical aspects of the projects, without requiring human assessment. Our evaluation framework assesses LLMs in generating circuits and producing compatible code across varying levels of project complexity. Through comprehensive testing of 13 leading models, \textsc{PCEval} provides the first reproducible and automatically validated empirical assessment of LLMs’ ability to reason about fundamental hardware implementation constraints within a simulation environment. Our findings reveal that while LLMs perform well in code generation and logical circuit design, they struggle significantly with physical breadboard layout creation, particularly in managing proper pin connections and avoiding circuit errors. \textsc{PCEval} advances our understanding of AI assistance in hardware-dependent computing environments and establishes a foundation for developing more effective tools to support physical computing education.

Method: Semi-supervised iterated learning model combining supervised and unsupervised learning within autoencoder architecture, applied to language learning task with 128 seven-segment display glyphs.

Result: Agents successfully learned and transmitted a language that is expressive (distinct codes for all 128 glyphs), compositional (signal components consistently map to meaning components), and stable (language doesn’t change across generations).

Conclusion: The iterated learning model with bottleneck enables emergence of structured language even for complex meaning spaces, demonstrating the model’s scalability and applicability to more realistic language learning scenarios.

Abstract: The iterated learning model simulates the transmission of language from generation to generation in order to explore how the constraints imposed by language transmission facilitate the emergence of language structure. Despite each modelled language learner starting from a blank slate, the presence of a bottleneck limiting the number of utterances to which the learner is exposed can lead to the emergence of language that lacks ambiguity, is governed by grammatical rules, and is consistent over successive generations, that is, one that is expressive, compositional and stable. The recent introduction of a more computationally tractable and ecologically valid semi supervised iterated learning model, combining supervised and unsupervised learning within an autoencoder architecture, has enabled exploration of language transmission dynamics for much larger meaning-signal spaces. Here, for the first time, the model has been successfully applied to a language learning task involving the communication of much more complex meanings: seven-segment display images. Agents in this model are able to learn and transmit a language that is expressive: distinct codes are employed for all 128 glyphs; compositional: signal components consistently map to meaning components, and stable: the language does not change from generation to generation.

Method: Builds on RECIPE-NLG framework and introduces composite objectives including a novel topological loss that represents ingredient lists as point clouds in embedding space to minimize divergence between predicted and gold ingredients. Also explores Dice loss and mixed loss combinations.

Result: The topological loss significantly improves ingredient- and action-level metrics. Dice loss excels in time/temperature precision. Mixed loss yields competitive trade-offs with synergistic gains in quantity and time. Human preference analysis shows their model is preferred in 62% of cases.

Conclusion: Specialized loss functions beyond standard cross-entropy are crucial for high-quality recipe generation, with topological loss for ingredient accuracy, Dice loss for time/temperature precision, and mixed losses providing balanced improvements across recipe-specific requirements.

Abstract: Cooking recipes are complex procedures that require not only a fluent and factual text, but also accurate timing, temperature, and procedural coherence, as well as the correct composition of ingredients. Standard training procedures are primarily based on cross-entropy and focus solely on fluency. Building on RECIPE-NLG, we investigate the use of several composite objectives and present a new topological loss that represents ingredient lists as point clouds in embedding space, minimizing the divergence between predicted and gold ingredients. Using both standard NLG metrics and recipe-specific metrics, we find that our loss significantly improves ingredient- and action-level metrics. Meanwhile, the Dice loss excels in time/temperature precision, and the mixed loss yields competitive trade-offs with synergistic gains in quantity and time. A human preference analysis supports our finding, showing our model is preferred in 62% of the cases.

Method: ModeX constructs a similarity graph over candidate generations and recursively applies spectral clustering to identify the modal output representing dominant semantic consensus. ModeX-Lite adds early pruning for efficiency improvements.

Result: The approaches consistently outperform standard single- and multi-path baselines across open-ended tasks including text summarization, code generation, and mathematical reasoning, providing computationally efficient solutions.

Conclusion: ModeX offers an evaluator-free framework for robust open-ended text generation by generalizing majority voting to semantic consensus identification through graph-based clustering techniques.

Abstract: Selecting a single high-quality output from multiple stochastic generations remains a fundamental challenge for large language models (LLMs), particularly in open-ended tasks where no canonical answer exists. While Best-of-N and self-consistency methods show that aggregating multiple generations can improve performance, existing approaches typically rely on external evaluators, reward models, or exact string-match voting, limiting their applicability and efficiency. We propose Mode Extraction (ModeX), an evaluator-free Best-of-N selection framework that generalizes majority voting to open-ended text generation by identifying the modal output representing the dominant semantic consensus among generated texts. ModeX constructs a similarity graph over candidate generations and recursively applies spectral clustering to select a representative centroid, without requiring additional inference or auxiliary models. We further instantiate this selection principle as ModeX-Lite, an improved version of ModeX with early pruning for efficiency. Across open-ended tasks – including text summarization, code generation, and mathematical reasoning – our approaches consistently outperform standard single- and multi-path baselines, providing a computationally efficient solution for robust open-ended text generation. Code is released in https://github.com/deeplearning-wisc/ModeX.

Method: LoRA-Drop applies temporal compute schedules to a fixed subset of intermediate layers: on most decoding steps, selected layers reuse previous-token hidden state with low-rank LoRA correction, while periodic refresh steps execute the full model to prevent drift. No routing network needed, compatible with standard KV caching, and reduces KV-cache footprint by skipping KV updates in droppable layers.

Result: Achieves up to 2.6× faster decoding and 45-55% KV-cache reduction while staying within 0.5 percentage points of baseline accuracy across LLaMA2-7B, LLaMA3-8B, Qwen2.5-7B, and Qwen2.5-14B. Consistent performance on reasoning, code generation, and long-context/multilingual benchmarks.

Conclusion: LoRA-Drop provides a simple plug-and-play inference framework for adaptive-capacity LLMs, identifying a safe zone of scheduling configurations that preserves quality while delivering substantial efficiency gains without complex routing mechanisms.

Abstract: Autoregressive large language models (LLMs) are bottlenecked by sequential decoding, where each new token typically requires executing all transformer layers. Existing dynamic-depth and layer-skipping methods reduce this cost, but often rely on auxiliary routing mechanisms or incur accuracy degradation when bypassed layers are left uncompensated. We present \textbf{LoRA-Drop}, a plug-and-play inference framework that accelerates decoding by applying a \emph{temporal compute schedule} to a fixed subset of intermediate layers: on most decoding steps, selected layers reuse the previous-token hidden state and apply a low-rank LoRA correction, while periodic \emph{refresh} steps execute the full model to prevent drift. LoRA-Drop requires no routing network, is compatible with standard KV caching, and can reduce KV-cache footprint by skipping KV updates in droppable layers during LoRA steps and refreshing periodically. Across \textbf{LLaMA2-7B}, \textbf{LLaMA3-8B}, \textbf{Qwen2.5-7B}, and \textbf{Qwen2.5-14B}, LoRA-Drop achieves up to \textbf{2.6$\times$ faster decoding} and \textbf{45–55% KV-cache reduction} while staying within \textbf{0.5 percentage points (pp)} of baseline accuracy. Evaluations on reasoning (GSM8K, MATH, BBH), code generation (HumanEval, MBPP), and long-context/multilingual benchmarks (LongBench, XNLI, XCOPA) identify a consistent \emph{safe zone} of scheduling configurations that preserves quality while delivering substantial efficiency gains, providing a simple path toward adaptive-capacity inference in LLMs. Codes are available at https://github.com/hosseinbv/LoRA-Drop.git.

Method: Conducted neuron-level interpretability study across three open-source LALMs (Qwen2.5-Omni, Kimi-Audio, Audio Flamingo 3). Used multiple neuron selectors (frequency-, entropy-, magnitude-, contrast-based) on emotion recognition benchmarks. Employed inference-time interventions including ablation and gain-based amplification to test causal relationships.

Result: Found consistent emotion-specific signature: ablating emotion-sensitive neurons disproportionately degrades recognition of that specific emotion while preserving other classes; gain-based amplification steers predictions toward target emotion. Effects scale systematically with intervention strength. Emotion-sensitive neurons show non-uniform layer-wise clustering with partial cross-dataset transfer.

Conclusion: Provides first causal, neuron-level account of emotion decisions in LALMs. Demonstrates that targeted neuron interventions offer actionable handles for controllable affective behaviors in audio-language models.

Abstract: Emotion is a central dimension of spoken communication, yet, we still lack a mechanistic account of how modern large audio-language models (LALMs) encode it internally. We present the first neuron-level interpretability study of emotion-sensitive neurons (ESNs) in LALMs and provide causal evidence that such units exist in Qwen2.5-Omni, Kimi-Audio, and Audio Flamingo 3. Across these three widely used open-source models, we compare frequency-, entropy-, magnitude-, and contrast-based neuron selectors on multiple emotion recognition benchmarks. Using inference-time interventions, we reveal a consistent emotion-specific signature: ablating neurons selected for a given emotion disproportionately degrades recognition of that emotion while largely preserving other classes, whereas gain-based amplification steers predictions toward the target emotion. These effects arise with modest identification data and scale systematically with intervention strength. We further observe that ESNs exhibit non-uniform layer-wise clustering with partial cross-dataset transfer. Taken together, our results offer a causal, neuron-level account of emotion decisions in LALMs and highlight targeted neuron interventions as an actionable handle for controllable affective behaviors.

Method: Probabilistic Certainty and Consistency (PCC) framework estimates factual confidence by jointly modeling an LLM’s probabilistic certainty and reasoning consistency. It uses confidence signals to implement adaptive verification: answer directly when confident, trigger targeted retrieval when uncertain/inconsistent, and escalate to deep search when ambiguity is high.

Result: PCC achieves better uncertainty quantification than verbalized confidence and consistently outperforms strong LLM-based fact-checking baselines across three challenging benchmarks. The framework generalizes well across various LLMs.

Conclusion: Confidence-guided routing mechanisms that adaptively decide when to retrieve external evidence based on internal confidence signals can significantly improve both efficiency and reliability of LLM fact-checking systems.

Abstract: Large language models (LLMs) are increasingly used in applications requiring factual accuracy, yet their outputs often contain hallucinated responses. While fact-checking can mitigate these errors, existing methods typically retrieve external evidence indiscriminately, overlooking the model’s internal knowledge and potentially introducing irrelevant noise. Moreover, current systems lack targeted mechanisms to resolve specific uncertainties in the model’s reasoning. Inspired by how humans fact-check, we argue that LLMs should adaptively decide whether to rely on internal knowledge or initiate retrieval based on their confidence in a given claim. We introduce Probabilistic Certainty and Consistency (PCC), a framework that estimates factual confidence by jointly modeling an LLM’s probabilistic certainty and reasoning consistency. These confidence signals enable an adaptive verification strategy: the model answers directly when confident, triggers targeted retrieval when uncertain or inconsistent, and escalates to deep search when ambiguity is high. Our confidence-guided routing mechanism ensures that retrieval is invoked only when necessary, improving both efficiency and reliability. Extensive experiments across three challenging benchmarks show that PCC achieves better uncertainty quantification than verbalized confidence and consistently outperforms strong LLM-based fact-checking baselines. Furthermore, we demonstrate that PCC generalizes well across various LLMs.

Method: DataParasite decomposes tabular curation tasks into independent entity-level searches defined through lightweight configuration files, executed via a shared Python script. It uses large language models for agentic search and structured extraction, and can be repurposed for new tasks using natural-language instructions.

Result: The pipeline achieves high accuracy across multiple canonical computational social science tasks (faculty hiring histories, elite death events, political career trajectories) while reducing data-collection costs by an order of magnitude relative to manual curation.

Conclusion: DataParasite provides a practical foundation for scalable, transparent, and reusable data curation by lowering technical and labor barriers to online data assembly in computational social science and beyond.

Abstract: Many questions in computational social science rely on datasets assembled from heterogeneous online sources, a process that is often labor-intensive, costly, and difficult to reproduce. Recent advances in large language models enable agentic search and structured extraction from the web, but existing systems are frequently opaque, inflexible, or poorly suited to scientific data curation. Here we introduce DataParasite, an open-source, modular pipeline for scalable online data collection. DataParasite decomposes tabular curation tasks into independent, entity-level searches defined through lightweight configuration files and executed through a shared, task-agnostic python script. Crucially, the same pipeline can be repurposed to new tasks, including those without predefined entity lists, using only natural-language instructions. We evaluate the pipeline on multiple canonical tasks in computational social science, including faculty hiring histories, elite death events, and political career trajectories. Across tasks, DataParasite achieves high accuracy while reducing data-collection costs by an order of magnitude relative to manual curation. By lowering the technical and labor barriers to online data assembly, DataParasite provides a practical foundation for scalable, transparent, and reusable data curation in computational social science and beyond.

Method: TELEVAL evaluates SLMs on 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: Current SLMs perform well on semantic/knowledge tasks but struggle to produce natural, interactionally appropriate responses, showing a gap in conversational quality.

Conclusion: There’s a need for more interaction-faithful evaluation of spoken language models, as current benchmarks don’t adequately capture real conversational quality.

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: Fine-tune Qwen-3 LLMs with Low-Rank Adaptation (LoRA) to generate multiple-choice responses conditioned on discrete ability descriptors. Reconstruct probability of correct responses as function of ability to create synthetic Item Characteristic Curves (ICCs) for IRT parameter estimation.

Result: The method competes with or outperforms baseline approaches on Grade 6 ELA items and BEA 2024 Shared Task dataset. Particularly effective at modeling item discrimination.

Conclusion: LLM-based simulation offers a cost-effective alternative to traditional field testing for IRT calibration, especially for estimating item discrimination parameters.

Abstract: Traditional methods for determining assessment item parameters, such as difficulty and discrimination, rely heavily on expensive field testing to collect student performance data for Item Response Theory (IRT) calibration. This study introduces a novel approach that implicitly models these psychometric properties by fine-tuning Large Language Models (LLMs) to simulate student responses across a spectrum of latent abilities. Leveraging the Qwen-3 dense model series and Low-Rank Adaptation (LoRA), we train models to generate responses to multiple choice questions conditioned on discrete ability descriptors. We reconstruct the probability of a correct response as a function of student ability, effectively generating synthetic Item Characteristic Curves (ICCs) to estimate IRT parameters. Evaluation on a dataset of Grade 6 English Language Arts (ELA) items and the BEA 2024 Shared Task dataset demonstrates that this method competes with or outperforms baseline approaches. This simulation-based technique seems particularly effective at modeling item discrimination.

Method: A three-stage framework: (1) Concept Graph Induction - extracts technical entities and relationships into a directed graph via expert-like reasoning; (2) Paragraph and Section Planning - reorganizes the graph into coherent clusters aligned with canonical patent sections; (3) Graph-Conditioned Generation - produces legally compliant paragraphs using section-specific subgraphs and tailored prompts.

Result: Experiments demonstrate that FlowPlan-G2P significantly improves logical coherence and legal compliance over end-to-end LLM baselines.

Conclusion: The framework establishes a new paradigm for paper-to-patent generation and advances structured text generation for specialized domains by mirroring expert cognitive workflows.

Abstract: Over 3.5 million patents are filed annually, with drafting patent descriptions requiring deep technical and legal expertise. Transforming scientific papers into patent descriptions is particularly challenging due to their differing rhetorical styles and stringent legal requirements. Unlike black-box text-to-text approaches that struggle to model structural reasoning and legal constraints, we propose FlowPlan-G2P, a novel framework that mirrors the cognitive workflow of expert drafters by reformulating this task into three stages: (1) Concept Graph Induction, extracting technical entities and relationships into a directed graph via expert-like reasoning; (2) Paragraph and Section Planning, reorganizing the graph into coherent clusters aligned with canonical patent sections; and (3) Graph-Conditioned Generation, producing legally compliant paragraphs using section-specific subgraphs and tailored prompts. Experiments demonstrate that FlowPlan-G2P significantly improves logical coherence and legal compliance over end-to-end LLM baselines. Our framework establishes a new paradigm for paper-to-patent generation and advances structured text generation for specialized domains.

Method: Study LLM interactions under social balance framework, examining effects of (i) interaction type, (ii) update mechanism, and (iii) population size across different LLM models.

Result: Social balance depends on all three factors; researchers characterized frequency of balance achievement, justifications for social dynamics, and diversity/stability of interactions.

Conclusion: Findings provide insights for deploying agentic systems by understanding how social balance emerges in LLM-based multi-agent environments.

Abstract: Large Language Models (LLMs) can be deployed in situations where they process positive/negative interactions with other agents. We study how this is done under the sociological framework of social balance, which explains the emergence of one faction or multiple antagonistic ones among agents. Across different LLM models, we find that balance depends on the (i) type of interaction, (ii) update mechanism, and (iii) population size. Across (i)-(iii), we characterize the frequency at which social balance is achieved, the justifications for the social dynamics, and the diversity and stability of interactions. Finally, we explain how our findings inform the deployment of agentic systems.

Method: Four-step pipeline: (1) identify medical concepts using MeSH thesaurus, (2) filter PubMed literature with permissive licenses (CC0), (3) extract (subject, relation, object) triplets using OpenIE, (4) enrich triplets with NER for biomedical relevance.

Result: T5 models fine-tuned on this dataset showed significantly improved performance and semantic coherence in comparative assessments using ROUGE and BERTScore metrics.

Conclusion: OpenIE-derived knowledge bases provide scalable, low-cost solutions for enhancing biomedical NLP, demonstrating potential for domain-specific LLM fine-tuning in oncology and other medical fields.

Abstract: The integration of Large Language Models (LLMs) into biomedical research offers new opportunities for domainspecific reasoning and knowledge representation. However, their performance depends heavily on the semantic quality of training data. In oncology, where precision and interpretability are vital, scalable methods for constructing structured knowledge bases are essential for effective fine-tuning. This study presents a pipeline for developing a lung cancer knowledge base using Open Information Extraction (OpenIE). The process includes: (1) identifying medical concepts with the MeSH thesaurus; (2) filtering open-access PubMed literature with permissive licenses (CC0); (3) extracting (subject, relation, object) triplets using OpenIE method; and (4) enriching triplet sets with Named Entity Recognition (NER) to ensure biomedical relevance. The resulting triplet sets provide a domain-specific, large-scale, and noise-aware resource for fine-tuning LLMs. We evaluated T5 models finetuned on this dataset through Supervised Semantic Fine-Tuning. Comparative assessments with ROUGE and BERTScore show significantly improved performance and semantic coherence, demonstrating the potential of OpenIE-derived resources as scalable, low-cost solutions for enhancing biomedical NLP.

Method: The paper proposes: (1) new comprehensive evidence-extraction metrics, and (2) a redact-and-retry framework with constrained filtering that improves over direct prompting approaches for document inconsistency detection.

Result: The paper reports promising experimental results that support the effectiveness of their proposed approach over direct prompting methods.

Conclusion: The introduced framework and metrics substantially improve LLM-based document inconsistency detection, particularly for the evidence extraction task of identifying inconsistent sentences within documents.

Abstract: Large language models (LLMs) are becoming useful in many domains due to their impressive abilities that arise from large training datasets and large model sizes. However, research on LLM-based approaches to document inconsistency detection is relatively limited. There are two key aspects of document inconsistency detection: (i) classification of whether there exists any inconsistency, and (ii) providing evidence of the inconsistent sentences. We focus on the latter, and introduce new comprehensive evidence-extraction metrics and a redact-and-retry framework with constrained filtering that substantially improves LLM-based document inconsistency detection over direct prompting. We back our claims with promising experimental results.

Method: Trained multiple encoder-based models (BERT, RoBERTa, DistilBERT, DeBERTa) with 512-token limit, built ensemble models integrating embeddings from multiple encoders, and compared with feature-based supervised ML models (Gradient-Boosted Decision Trees, XGBoost, LightGBM). Used dataset of 17,307 essays with 80%/10%/10% train/validation/test split.

Result: Ensemble-of-embeddings model combining multiple pre-trained language model representations with gradient-boosting classifier significantly outperformed individual language models for scoring long essays, as measured by Quadratic Weighted Kappa.

Conclusion: For automated scoring of long essays, ensemble approaches that combine multiple language model representations with gradient-boosting classifiers are more effective than individual encoder-only models with token limitations.

Abstract: Long context may impose challenges for encoder-only language models in text processing, specifically for automated scoring of essays. This study trained several commonly used encoder-based language models for automated scoring of long essays. The performance of these trained models was evaluated and compared with the ensemble models built upon the base language models with a token limit of 512?. The experimented models include BERT-based models (BERT, RoBERTa, DistilBERT, and DeBERTa), ensemble models integrating embeddings from multiple encoder models, and ensemble models of feature-based supervised machine learning models, including Gradient-Boosted Decision Trees, eXtreme Gradient Boosting, and Light Gradient Boosting Machine. We trained, validated, and tested each model on a dataset of 17,307 essays, with an 80%/10%/10% split, and evaluated model performance using Quadratic Weighted Kappa. This study revealed that an ensemble-of-embeddings model that combines multiple pre-trained language model representations with gradient-boosting classifier as the ensemble model significantly outperforms individual language models at scoring long essays.

Method: Used LangChain/LangGraph to compare one-shot baseline against plan-execute-replan agent with task-specific tools (DBpedia SPARQL/lookup/schema, Wikipedia retrieval, web search). Evaluated GPT-4o and GPT-4o-mini on 60 examples each from Event-QA and CMV with 3 splits of 20, measuring accuracy, mean latency, and token cost estimates.

Result: On Event-QA: Tool-augmented configuration improved accuracy (47.5% → 67.5% for GPT-4o) but increased latency dramatically (~8s → ~317s per example). On CMV: One-shot prompting was strongest (GPT-4o-mini achieved 75% at ~6s), with planning+search increasing latency substantially without consistent gains. Complex multi-tool orchestration exposed failure modes where smaller models degraded.

Conclusion: Findings highlight the need for task-specific, cost-aware choices of both model size and agent/tooling complexity, as accuracy gains from tool augmentation come with massive latency overheads that may not be justified for all tasks.

Abstract: Modern large language models (LLMs) increasingly rely on inference-time planning and external tools to improve reasoning. We benchmark this behavior on two real-world settings: event-centric question answering over graph-structured knowledge (Event-QA) and persuasive response generation in Reddit ChangeMyView (CMV). Using LangChain and LangGraph, we compare a one-shot baseline against a plan-execute-replan agent equipped with task-specific tools (DBpedia SPARQL/lookup/schema exploration, Wikipedia-focused retrieval, and topical web search). We evaluate on 60 examples each from Event-QA and CMV (3 splits of 20), and report both mean end-to-end latency and per-example token cost estimates. We evaluate GPT-4o and GPT-4o-mini under identical workflows and report accuracy and end-to-end latency. On Event-QA, the best tool-augmented configuration improves accuracy (e.g., 47.5% $\rightarrow$ 67.5% for GPT-4o) while increasing latency by orders of magnitude ($\sim$8s $\rightarrow$ $\sim$317s per example). On CMV, one-shot prompting is strongest (e.g., GPT-4o-mini achieves 75% at $\sim$6s), and planning+search increases latency substantially without consistent gains. However, complex multi-tool orchestration exposes failure modes where the smaller model degrades. Overall, the findings highlight the need for task-specific, cost-aware choices of both model size and agent/tooling complexity.

Method: FactArena integrates three components: (1) LLM-driven fact-checking process with standardized claim decomposition, tool-augmented evidence retrieval, and justification-based verdict prediction; (2) Arena-styled judgment mechanism with consolidated reference guidelines for unbiased pairwise comparisons; (3) Arena-driven claim-evolution module that adaptively generates more challenging claims to probe factual robustness.

Result: Across 16 state-of-the-art LLMs from seven model families, FactArena produces stable and interpretable rankings. Analyses reveal significant discrepancies between static claim-verification accuracy and end-to-end fact-checking competence.

Conclusion: FactArena offers a scalable, trustworthy paradigm for diagnosing LLMs’ factual reasoning, guiding future model development, and advancing reliable deployment of LLMs in safety-critical fact-checking applications. Holistic evaluation is necessary beyond static claim verification.

Abstract: Large Language Models (LLMs) are increasingly deployed in real-world fact-checking systems, yet existing evaluations focus predominantly on claim verification and overlook the broader fact-checking workflow, including claim extraction and evidence retrieval. This narrow focus prevents current benchmarks from revealing systematic reasoning failures, factual blind spots, and robustness limitations of modern LLMs. To bridge this gap, we present FactArena, a fully automated arena-style evaluation framework that conducts comprehensive, stage-wise benchmarking of LLMs across the complete fact-checking pipeline. FactArena integrates three key components: (i) an LLM-driven fact-checking process that standardizes claim decomposition, evidence retrieval via tool-augmented interactions, and justification-based verdict prediction; (ii) an arena-styled judgment mechanism guided by consolidated reference guidelines to ensure unbiased and consistent pairwise comparisons across heterogeneous judge agents; and (iii) an arena-driven claim-evolution module that adaptively generates more challenging and semantically controlled claims to probe LLMs’ factual robustness beyond fixed seed data. Across 16 state-of-the-art LLMs spanning seven model families, FactArena produces stable and interpretable rankings. Our analyses further reveal significant discrepancies between static claim-verification accuracy and end-to-end fact-checking competence, highlighting the necessity of holistic evaluation. The proposed framework offers a scalable and trustworthy paradigm for diagnosing LLMs’ factual reasoning, guiding future model development, and advancing the reliable deployment of LLMs in safety-critical fact-checking applications.

Method: Proposes Self-Verifying Agents with dual missions: complete tasks AND prove accomplishment with curated snapshot evidence guided by 3C Principles (Completeness, Conciseness, Creativity). Agents perform in-situ self-verification on minimal decisive snapshots.

Result: Experiments on mobile tasks show performance gains up to 26.08% for 8B models and 16.66% for 30B models, enabling scalable training of LLM-driven agents with competitive performance against larger models.

Conclusion: SmartSnap paradigm enables scalable RL agent training by shifting to proactive self-verification, where agents provide curated evidence rather than requiring exhaustive trajectory analysis, improving efficiency and reliability.

Abstract: Agentic reinforcement learning (RL) holds great promise for the development of autonomous agents under complex GUI tasks, but its scalability remains severely hampered by the verification of task completion. Existing task verification is treated as a passive, post-hoc process: a verifier (i.e., rule-based scoring script, reward or critic model, and LLM-as-a-Judge) analyzes the agent’s entire interaction trajectory to determine if the agent succeeds. Such processing of verbose context that contains irrelevant, noisy history poses challenges to the verification protocols and therefore leads to prohibitive cost and low reliability. To overcome this bottleneck, we propose SmartSnap, a paradigm shift from this passive, post-hoc verification to proactive, in-situ self-verification by the agent itself. We introduce the Self-Verifying Agent, a new type of agent designed with dual missions: to not only complete a task but also to prove its accomplishment with curated snapshot evidences. Guided by our proposed 3C Principles (Completeness, Conciseness, and Creativity), the agent leverages its accessibility to the online environment to perform self-verification on a minimal, decisive set of snapshots. Such evidences are provided as the sole materials for a general LLM-as-a-Judge verifier to determine their validity and relevance. Experiments on mobile tasks across model families and scales demonstrate that our SmartSnap paradigm allows training LLM-driven agents in a scalable manner, bringing performance gains up to 26.08% and 16.66% respectively to 8B and 30B models. The synergizing between solution finding and evidence seeking facilitates the cultivation of efficient, self-verifying agents with competitive performance against DeepSeek V3.1 and Qwen3-235B-A22B. Code is available at: https://github.com/TencentYoutuResearch/SmartSnap

Method: LATS reformulates jailbreaking as breadth-first tree search over multi-turn dialogues. Each node incrementally injects missing content words from the attack goal into benign prompts using lexical anchor injection, operating without attacker LLMs.

Result: Achieves 97-100% attack success rate on latest GPT, Claude, and Llama models with average of only ~6.4 queries, compared to 20+ queries required by other methods. Demonstrates superior query efficiency on AdvBench and HarmBench.

Conclusion: Conversational structure is a potent and under-protected attack surface. LATS shows superior query efficiency in an era where high ASR is readily achievable, offering a resource-light alternative to existing jailbreak methods.

Abstract: Most jailbreak methods achieve high attack success rates (ASR) but require attacker LLMs to craft adversarial queries and/or demand high query budgets. These resource limitations make jailbreaking expensive, and the queries generated by attacker LLMs often consist of non-interpretable random prefixes. This paper introduces Lexical Anchor Tree Search (), addressing these limitations through an attacker-LLM-free method that operates purely via lexical anchor injection. LATS reformulates jailbreaking as a breadth-first tree search over multi-turn dialogues, where each node incrementally injects missing content words from the attack goal into benign prompts. Evaluations on AdvBench and HarmBench demonstrate that LATS achieves 97-100% ASR on latest GPT, Claude, and Llama models with an average of only ~6.4 queries, compared to 20+ queries required by other methods. These results highlight conversational structure as a potent and under-protected attack surface, while demonstrating superior query efficiency in an era where high ASR is readily achievable. Our code will be released to support reproducibility.

Method: Two-phase procedure: (1) initial probe with Best-of-N jailbreak when needed, (2) iterative continuation prompts to extract books; evaluated on Claude 3.7 Sonnet, GPT-4.1, Gemini 2.5 Pro, and Grok 3 using nv-recall (block-based longest common substring approximation).

Result: Varying extraction success: Gemini 2.5 Pro and Grok 3 extracted text without jailbreaks (76.8% and 70.3% nv-recall for Harry Potter), Claude 3.7 Sonnet extracted near-verbatim books with jailbreak (95.8% nv-recall), GPT-4.1 required 20X more attempts and eventually refused (4.0% nv-recall).

Conclusion: Extraction of copyrighted training data remains a risk for production LLMs even with model- and system-level safeguards, highlighting ongoing copyright concerns.

Abstract: Many unresolved legal questions over LLMs and copyright center on memorization: whether specific training data have been encoded in the model’s weights during training, and whether those memorized data can be extracted in the model’s outputs. While many believe that LLMs do not memorize much of their training data, recent work shows that substantial amounts of copyrighted text can be extracted from open-weight models. However, it remains an open question if similar extraction is feasible for production LLMs, given the safety measures these systems implement. We investigate this question using a two-phase procedure: (1) an initial probe to test for extraction feasibility, which sometimes uses a Best-of-N (BoN) jailbreak, followed by (2) iterative continuation prompts to attempt to extract the book. We evaluate our procedure on four production LLMs – Claude 3.7 Sonnet, GPT-4.1, Gemini 2.5 Pro, and Grok 3 – and we measure extraction success with a score computed from a block-based approximation of longest common substring (nv-recall). With different per-LLM experimental configurations, we were able to extract varying amounts of text. For the Phase 1 probe, it was unnecessary to jailbreak Gemini 2.5 Pro and Grok 3 to extract text (e.g, nv-recall of 76.8% and 70.3%, respectively, for Harry Potter and the Sorcerer’s Stone), while it was necessary for Claude 3.7 Sonnet and GPT-4.1. In some cases, jailbroken Claude 3.7 Sonnet outputs entire books near-verbatim (e.g., nv-recall=95.8%). GPT-4.1 requires significantly more BoN attempts (e.g., 20X), and eventually refuses to continue (e.g., nv-recall=4.0%). Taken together, our work highlights that, even with model- and system-level safeguards, extraction of (in-copyright) training data remains a risk for production LLMs.

Method: Structured pruning framework that eliminates entire architectural components (channels). Uses hybrid multi-domain calibration set and iterative calibration strategy to identify redundant channels while maintaining standard hardware compatibility.

Result: Achieves significant compression with minimal performance degradation across various models and diverse downstream tasks.

Conclusion: Structured pruning with hybrid calibration provides effective solution for LLM deployment challenges, offering hardware-compatible compression without substantial performance loss.

Abstract: Large Language Models (LLMs) have achieved remarkable success across a wide spectrum of natural language processing tasks. However, their ever-growing scale introduces significant barriers to real-world deployment, including substantial computational overhead, memory footprint, and inference latency. While model pruning presents a viable solution to these challenges, existing unstructured pruning techniques often yield irregular sparsity patterns that necessitate specialized hardware or software support. In this work, we explore structured pruning, which eliminates entire architectural components and maintains compatibility with standard hardware accelerators. We introduce a novel structured pruning framework that leverages a hybrid multi-domain calibration set and an iterative calibration strategy to effectively identify and remove redundant channels. Extensive experiments on various models across diverse downstream tasks show that our approach achieves significant compression with minimal performance degradation.

Method: Evaluated three transformer models (BERT-base-multilingual-cased, RoBERTa-base, XLM-RoBERTa-base with first 8 layers frozen) across five languages (English, Korean, Japanese, Chinese, French) using standard metrics (accuracy, precision, recall, F1-score) and integrated LIME framework for interpretability.

Result: Performance comparison of transformer models on multilingual sentiment analysis and hate speech detection tasks across five languages, with enhanced interpretability through LIME explanations showing word-level contributions to predictions.

Conclusion: Combining state-of-the-art transformer architectures with explainability techniques improves both effectiveness and transparency of multilingual sentiment analysis and hate speech detection systems for content moderation.

Abstract: Sentiment analysis focuses on identifying the emotional polarity expressed in textual data, typically categorized as positive, negative, or neutral. Hate speech detection, on the other hand, aims to recognize content that incites violence, discrimination, or hostility toward individuals or groups based on attributes such as race, gender, sexual orientation, or religion. Both tasks play a critical role in online content moderation by enabling the detection and mitigation of harmful or offensive material, thereby contributing to safer digital environments. In this study, we examine the performance of three transformer-based models: BERT-base-multilingual-cased, RoBERTa-base, and XLM-RoBERTa-base with the first eight layers frozen, for multilingual sentiment analysis and hate speech detection. The evaluation is conducted across five languages: English, Korean, Japanese, Chinese, and French. The models are compared using standard performance metrics, including accuracy, precision, recall, and F1-score. To enhance model interpretability and provide deeper insight into prediction behavior, we integrate the Local Interpretable Model-agnostic Explanations (LIME) framework, which highlights the contribution of individual words to the models decisions. By combining state-of-the-art transformer architectures with explainability techniques, this work aims to improve both the effectiveness and transparency of multilingual sentiment analysis and hate speech detection systems.

Method: Systematic experimentation across model scales (ELECTRA-small to ELECTRA-base), comprehensive multi-level error analysis using five categorization schemes, evaluation of adversarial fine-tuning ratios, data augmentation experiments, and implementation of three targeted mitigation strategies including Entity-Aware contrastive learning.

Result: Identified 80% clean + 20% adversarial data as optimal fine-tuning ratio; found scaling eliminates robustness-accuracy trade-off; Entity-Aware contrastive learning achieved best performance: 89.89% AddSent EM and 90.73% SQuAD EM, representing 94.9% closure of adversarial gap.

Conclusion: Targeted mitigation strategies, particularly Entity-Aware contrastive learning guided by NER, can achieve near-parity between clean and adversarial performance, demonstrating that comprehensive linguistic error analysis combined with targeted interventions effectively addresses adversarial vulnerabilities in QA systems.

Abstract: Question answering (QA) systems achieve impressive performance on standard benchmarks like SQuAD, but remain vulnerable to adversarial examples. This project investigates the adversarial robustness of transformer models on the AddSent adversarial dataset through systematic experimentation across model scales and targeted mitigation strategies. We perform comprehensive multi-level error analysis using five complementary categorization schemes, identifying negation confusion and entity substitution as the primary failure modes. Through systematic evaluation of adversarial fine-tuning ratios, we identify 80% clean + 20% adversarial data as optimal. Data augmentation experiments reveal a capacity bottleneck in small models. Scaling from ELECTRA-small (14M parameters) to ELECTRA-base (110M parameters) eliminates the robustness-accuracy trade-off, achieving substantial improvements on both clean and adversarial data. We implement three targeted mitigation strategies, with Entity-Aware contrastive learning achieving best performance: 89.89% AddSent Exact Match (EM) and 90.73% SQuAD EM, representing 94.9% closure of the adversarial gap. To our knowledge, this is the first work integrating comprehensive linguistic error analysis with Named Entity Recognition (NER)-guided contrastive learning for adversarial QA, demonstrating that targeted mitigation can achieve near-parity between clean and adversarial performance.

Method: Builds on knowledge distillation chain-style model, introduces code module to guide knowledge-graph exploration, incorporates code as part of chain-of-thought prompt to provide structured external knowledge, and uses this to constrain LLM reasoning process.

Result: Experimental evaluation on GPT-4 and LLaMA-3.3 shows significant improvements: HIT@1, HIT@3, and HIT@5 improve by 15.64%, 13.38%, and 13.28% respectively. Method achieves HIT@1, HIT@3, and HIT@5 scores exceeding 95% across multiple evaluation settings.

Conclusion: The proposed code-enhanced approach substantially reduces hallucination behavior while improving accuracy and verifiability of large language models, demonstrating that structured code modules effectively mitigate prompt-induced hallucinations.

Abstract: To address hallucination issues in large language models (LLMs), this paper proposes a method for mitigating prompt-induced hallucinations. Building on a knowledge distillation chain-style model, we introduce a code module to guide knowledge-graph exploration and incorporate code as part of the chain-of-thought prompt, forming an external knowledge input that provides more accurate and structured information to the model. Based on this design, we develop an improved knowledge distillation chain-style model and leverage it to analyze and constrain the reasoning process of LLMs, thereby improving inference accuracy. We empirically evaluate the proposed approach using GPT-4 and LLaMA-3.3 on multiple public datasets. Experimental results demonstrate that incorporating code modules significantly enhances the model’s ability to capture contextual information and effectively mitigates prompt-induced hallucinations. Specifically, HIT@1, HIT@3, and HIT@5 improve by 15.64%, 13.38%, and 13.28%, respectively. Moreover, the proposed method achieves HIT@1, HIT@3, and HIT@5 scores exceeding 95% across several evaluation settings. These results indicate that the proposed approach substantially reduces hallucination behavior while improving the accuracy and verifiability of large language models.

Method: Developed a likelihood function quantifying alignment between language processing units and working memory capacity; used computational simulations of symbol sequences and validation analyses of natural language sentences; compared hierarchical vs. linear processing.

Result: Hierarchical processing far surpasses linear processing in constraining memory load within human working memory limits as sequence length increases; shows converging pattern with children’s working memory development.

Conclusion: Hierarchical language structure optimally solves the challenge of limited working memory capacity, explaining the universal hierarchical nature of human language as an efficiency optimization.

Abstract: Language is a uniquely human trait, conveying information efficiently by organizing word sequences in sentences into hierarchical structures. A central question persists: Why is human language hierarchical? In this study, we show that hierarchization optimally solves the challenge of our limited working memory capacity. We established a likelihood function that quantifies how well the average number of units according to the language processing mechanisms aligns with human working memory capacity (WMC) in a direct fashion. The maximum likelihood estimate (MLE) of this function, tehta_MLE, turns out to be the mean of units. Through computational simulations of symbol sequences and validation analyses of natural language sentences, we uncover that compared to linear processing, hierarchical processing far surpasses it in constraining the tehta_MLE values under the human WMC limit, along with the increase of sequence/sentence length successfully. It also shows a converging pattern related to children’s WMC development. These results suggest that constructing hierarchical structures optimizes the processing efficiency of sequential language input while staying within memory constraints, genuinely explaining the universal hierarchical nature of human language.

Method: Synapse models memory as a dynamic graph where relevance emerges from spreading activation rather than pre-computed links, integrating lateral inhibition and temporal decay. It uses Triple Hybrid Retrieval that fuses geometric embeddings with activation-based graph traversal.

Result: Comprehensive evaluations on the LoCoMo benchmark show Synapse significantly outperforms state-of-the-art methods in complex temporal and multi-hop reasoning tasks.

Conclusion: Synapse offers a robust solution to the “Contextual Tunneling” problem and provides a unified memory architecture that transcends static vector similarity for LLMs.

Abstract: While Large Language Models (LLMs) excel at generalized reasoning, standard retrieval-augmented approaches fail to address the disconnected nature of long-term agentic memory. To bridge this gap, we introduce Synapse (Synergistic Associative Processing Semantic Encoding), a unified memory architecture that transcends static vector similarity. Drawing from cognitive science, Synapse models memory as a dynamic graph where relevance emerges from spreading activation rather than pre-computed links. By integrating lateral inhibition and temporal decay, the system dynamically highlights relevant sub-graphs while filtering interference. We implement a Triple Hybrid Retrieval strategy that fuses geometric embeddings with activation-based graph traversal. Comprehensive evaluations on the LoCoMo benchmark show that Synapse significantly outperforms state-of-the-art methods in complex temporal and multi-hop reasoning tasks, offering a robust solution to the “Contextual Tunneling” problem. Our code and data will be made publicly available upon acceptance.

Method: WBC (Window-Based Comparison) uses a sliding window approach with sign-based aggregation. It slides windows of varying sizes across text sequences, with each window casting a binary vote on membership based on loss comparisons between target and reference models. Votes are ensembled across geometrically spaced window sizes to capture memorization patterns from token-level artifacts to phrase-level structures.

Result: Extensive experiments across eleven datasets show WBC substantially outperforms established baselines, achieving higher AUC scores and 2-3 times improvements in detection rates at low false positive thresholds.

Conclusion: Aggregating localized evidence is fundamentally more effective than global averaging for membership inference attacks, exposing critical privacy vulnerabilities in fine-tuned LLMs.

Abstract: Most membership inference attacks (MIAs) against Large Language Models (LLMs) rely on global signals, like average loss, to identify training data. This approach, however, dilutes the subtle, localized signals of memorization, reducing attack effectiveness. We challenge this global-averaging paradigm, positing that membership signals are more pronounced within localized contexts. We introduce WBC (Window-Based Comparison), which exploits this insight through a sliding window approach with sign-based aggregation. Our method slides windows of varying sizes across text sequences, with each window casting a binary vote on membership based on loss comparisons between target and reference models. By ensembling votes across geometrically spaced window sizes, we capture memorization patterns from token-level artifacts to phrase-level structures. Extensive experiments across eleven datasets demonstrate that WBC substantially outperforms established baselines, achieving higher AUC scores and 2-3 times improvements in detection rates at low false positive thresholds. Our findings reveal that aggregating localized evidence is fundamentally more effective than global averaging, exposing critical privacy vulnerabilities in fine-tuned LLMs.

Method: Proposed EComStage benchmark evaluates LLMs across three reasoning stages: Perception (understanding user intent), Planning (formulating action plan), and Action (executing decisions). It includes seven representative tasks spanning diverse e-commerce scenarios, with all samples human-annotated and quality-checked. The benchmark covers both customer-oriented and merchant-oriented scenarios.

Result: Evaluated over 30 LLMs ranging from 1B to over 200B parameters, including open-source models and closed-source APIs. Results revealed stage/orientation-specific strengths and weaknesses, providing fine-grained insights into model performance across different reasoning stages and scenario types.

Conclusion: EComStage provides comprehensive, actionable insights for designing and optimizing LLM-based agents in real-world e-commerce settings by evaluating the complete reasoning process across both customer and merchant scenarios, addressing limitations of existing benchmarks.

Abstract: Large Language Model (LLM)-based agents are increasingly deployed in e-commerce applications to assist customer services in tasks such as product inquiries, recommendations, and order management. Existing benchmarks primarily evaluate whether these agents successfully complete the final task, overlooking the intermediate reasoning stages that are crucial for effective decision-making. To address this gap, we propose EComStage, a unified benchmark for evaluating agent-capable LLMs across the comprehensive stage-wise reasoning process: Perception (understanding user intent), Planning (formulating an action plan), and Action (executing the decision). EComStage evaluates LLMs through seven separate representative tasks spanning diverse e-commerce scenarios, with all samples human-annotated and quality-checked. Unlike prior benchmarks that focus only on customer-oriented interactions, EComStage also evaluates merchant-oriented scenarios, including promotion management, content review, and operational support relevant to real-world applications. We evaluate a wide range of over 30 LLMs, spanning from 1B to over 200B parameters, including open-source models and closed-source APIs, revealing stage/orientation-specific strengths and weaknesses. Our results provide fine-grained, actionable insights for designing and optimizing LLM-based agents in real-world e-commerce settings.

Method: Uses Mixture-of-Experts architecture with hybrid attention (Sliding Window + global attention in 5:1 ratio), pre-trained on 27T tokens with Multi-Token Prediction, extended to 256k context. Introduces Multi-Teacher On-Policy Distillation where domain-specialized teachers provide dense token-level rewards.

Result: Rivals DeepSeek-V3.2 and Kimi-K2 despite using only 1/2 and 1/3 of their total parameters. Achieves up to 3.6 acceptance length and 2.6x decoding speedup using MTP as draft model for speculative decoding.

Conclusion: MiMo-V2-Flash demonstrates that efficient architecture design combined with innovative training techniques can produce models that match or exceed larger competitors while offering practical inference advantages, with open-sourcing promoting community collaboration.

Abstract: We present MiMo-V2-Flash, a Mixture-of-Experts (MoE) model with 309B total parameters and 15B active parameters, designed for fast, strong reasoning and agentic capabilities. MiMo-V2-Flash adopts a hybrid attention architecture that interleaves Sliding Window Attention (SWA) with global attention, with a 128-token sliding window under a 5:1 hybrid ratio. The model is pre-trained on 27 trillion tokens with Multi-Token Prediction (MTP), employing a native 32k context length and subsequently extended to 256k. To efficiently scale post-training compute, MiMo-V2-Flash introduces a novel Multi-Teacher On-Policy Distillation (MOPD) paradigm. In this framework, domain-specialized teachers (e.g., trained via large-scale reinforcement learning) provide dense and token-level reward, enabling the student model to perfectly master teacher expertise. MiMo-V2-Flash rivals top-tier open-weight models such as DeepSeek-V3.2 and Kimi-K2, despite using only 1/2 and 1/3 of their total parameters, respectively. During inference, by repurposing MTP as a draft model for speculative decoding, MiMo-V2-Flash achieves up to 3.6 acceptance length and 2.6x decoding speedup with three MTP layers. We open-source both the model weights and the three-layer MTP weights to foster open research and community collaboration.

Method: PHSA uses punctuation tokens as semantic boundary anchors with: (1) dual-branch aggregation fusing global semantic representations with punctuation-enhanced boundary features, and (2) extreme-sparsity-adaptive training/inference strategy for stable behavior at low token activation ratios.

Result: PHSA consistently outperforms dense attention and state-of-the-art sparse attention baselines (including InfLLM v2). For 0.6B-parameter model with 32k-token sequences, reduces information loss by 10.8% at 97.3% sparsity ratio.

Conclusion: PHSA provides an effective sparse attention framework that leverages punctuation as semantic anchors to preserve critical information while maintaining computational efficiency for long-context modeling in LLMs.

Abstract: Attention serves as the fundamental mechanism for long-context modeling in large language models (LLMs), yet dense attention becomes structurally prohibitive for long sequences due to its quadratic complexity. Consequently, sparse attention has received increasing attention as a scalable alternative. However, existing sparse attention methods rely on coarse-grained semantic representations during block selection, which blur intra-block semantic boundaries and lead to the loss of critical information. To address this issue, we propose \textbf{P}unctuation-aware \textbf{H}ybrid \textbf{S}parse \textbf{A}ttention \textbf{(PHSA)}, a natively trainable sparse attention framework that leverages punctuation tokens as semantic boundary anchors. Specifically, (1) we design a dual-branch aggregation mechanism that fuses global semantic representations with punctuation-enhanced boundary features, preserving the core semantic structure while introducing almost no additional computational overhead; (2) we introduce an extreme-sparsity-adaptive training and inference strategy that stabilizes model behavior under very low token activation ratios; Extensive experiments on general benchmarks and long-context evaluations demonstrate that PHSA consistently outperforms dense attention and state-of-the-art sparse attention baselines, including InfLLM v2. Specifically, for the 0.6B-parameter model with 32k-token input sequences, PHSA can reduce the information loss by 10.8% at a sparsity ratio of 97.3%.

Method: Iterative vector updating: constructs sparse vectors based on representative texts and refines them through LLM guidance. Works on any embedder, uses relatively small LLMs, and is model-agnostic.

Result: Achieves comparable or superior results to state-of-the-art contrastive learning methods without assuming prior knowledge of clusters or labels. Scales to large datasets and reduces computational costs.

Conclusion: The method’s low-resource requirements, adaptability, and scalability make it more aligned with real-world scenarios than existing clustering methods for customer chatbot applications.

Abstract: In this paper, we propose a training-free and label-free method for short text clustering that can be used on top of any existing embedder. In the context of customer-facing chatbots, companies are dealing with large amounts of user utterances that need to be clustered according to their intent. In these commercial settings, no labeled data is typically available, and the number of clusters is not known. Our method is based on iterative vector updating: it constructs sparse vectors based on representative texts, and then iteratively refines them through LLM guidance. Our method achieves comparable or superior results to state-of-the-art methods that use contrastive learning, but without 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 that our method scales to large datasets, reducing the computational cost of the LLM. These low-resource, adaptable settings and the scalability of our method make it more aligned with real-world scenarios than existing clustering methods.

Method: Direct-questioning paradigm evaluating models (GPT-5.1, DeepSeek-V3.2, Qwen3-Max, Gemini2.5Pro) on Chinese cultural knowledge including history, literature, and poetry.

Result: Chinese models generally outperform US models on Chinese cultural tasks. Among US models, Gemini 2.5Pro and GPT-5.1 achieve relatively higher accuracy.

Conclusion: Performance differences likely stem from variations in training data distribution, localization strategies, and emphasis on Chinese cultural content during model development.

Abstract: Cultural backgrounds shape individuals’ perspectives and approaches to problem-solving. Since the emergence of GPT-1 in 2018, large language models (LLMs) have undergone rapid development. To date, the world’s ten leading LLM developers are primarily based in China and the United States. To examine whether LLMs released by Chinese and U.S. developers exhibit cultural differences in Chinese-language settings, we evaluate their performance on questions about Chinese culture. This study adopts a direct-questioning paradigm to evaluate models such as GPT-5.1, DeepSeek-V3.2, Qwen3-Max, and Gemini2.5Pro. We assess their understanding of traditional Chinese culture, including history, literature, poetry, and related domains. Comparative analyses between LLMs developed in China and the U.S. indicate that Chinese models generally outperform their U.S. counterparts on these tasks. Among U.S.-developed models, Gemini 2.5Pro and GPT-5.1 achieve relatively higher accuracy. The observed performance differences may potentially arise from variations in training data distribution, localization strategies, and the degree of emphasis on Chinese cultural content during model development.

Method: TiMem organizes conversations through a Temporal Memory Tree (TMT) with three core properties: 1) temporal-hierarchical organization through TMT, 2) semantic-guided consolidation enabling memory integration across hierarchical levels without fine-tuning, and 3) complexity-aware memory recall that balances precision and efficiency across varying query complexities.

Result: TiMem achieves state-of-the-art accuracy: 75.30% on LoCoMo and 76.88% on LongMemEval-S, outperforming all baselines while reducing recalled memory length by 52.20% on LoCoMo. Manifold analysis shows clear persona separation on LoCoMo and reduced dispersion on LongMemEval-S.

Conclusion: TiMem treats temporal continuity as a first-class organizing principle for long-horizon memory in conversational agents, providing systematic memory consolidation from raw observations to abstracted persona representations through its temporal-hierarchical framework.

Abstract: Long-horizon conversational agents have to manage ever-growing interaction histories that quickly exceed the finite context windows of large language models (LLMs). Existing memory frameworks provide limited support for temporally structured information across hierarchical levels, often leading to fragmented memories and unstable long-horizon personalization. We present TiMem, a temporal–hierarchical memory framework that organizes conversations through a Temporal Memory Tree (TMT), enabling systematic memory consolidation from raw conversational observations to progressively abstracted persona representations. TiMem is characterized by three core properties: (1) temporal–hierarchical organization through TMT; (2) semantic-guided consolidation that enables memory integration across hierarchical levels without fine-tuning; and (3) complexity-aware memory recall that balances precision and efficiency across queries of varying complexity. Under a consistent evaluation setup, TiMem achieves state-of-the-art accuracy on both benchmarks, reaching 75.30% on LoCoMo and 76.88% on LongMemEval-S. It outperforms all evaluated baselines while reducing the recalled memory length by 52.20% on LoCoMo. Manifold analysis indicates clear persona separation on LoCoMo and reduced dispersion on LongMemEval-S. Overall, TiMem treats temporal continuity as a first-class organizing principle for long-horizon memory in conversational agents.

Method: Use inverse socio-demographic prompting (ISDP) where LLMs discriminate and predict demographic proxies from actual and simulated user behavior. Test on Goodreads-CSI dataset with users from India, Mexico, and USA, using four LLMs: Aya-23, Gemma-2, GPT-4o, and LLaMA-3.1.

Result: Models perform better with actual behaviors than simulated ones (contrary to SDP findings). Performance with both behavior types diminishes and becomes nearly equal at individual level, indicating limits to personalization.

Conclusion: ISDP provides a more reliable method than SDP for assessing LLM cultural competency, revealing that LLMs can better predict demographics from actual user behavior but have limited ability for individual-level personalization.

Abstract: Socio-demographic prompting (SDP) - prompting Large Language Models (LLMs) using demographic proxies to generate culturally aligned outputs - often shows LLM responses as stereotypical and biased. While effective in assessing LLMs’ cultural competency, SDP is prone to confounding factors such as prompt sensitivity, decoding parameters, and the inherent difficulty of generation over discrimination tasks due to larger output spaces. These factors complicate interpretation, making it difficult to determine if the poor performance is due to bias or the task design. To address this, we use inverse socio-demographic prompting (ISDP), where we prompt LLMs to discriminate and predict the demographic proxy from actual and simulated user behavior from different users. We use the Goodreads-CSI dataset (Saha et al., 2025), which captures difficulty in understanding English book reviews for users from India, Mexico, and the USA, and test four LLMs: Aya-23, Gemma-2, GPT-4o, and LLaMA-3.1 with ISDP. Results show that models perform better with actual behaviors than simulated ones, contrary to what SDP suggests. However, performance with both behavior types diminishes and becomes nearly equal at the individual level, indicating limits to personalization.

Method: Introduce homotokens (alternative valid subword segmentations) as data augmentation. Use lightweight architecture with auxiliary causal encoder and block-causal cross-attention to condition canonical next-token prediction on sampled homotoken variants.

Result: In data-constrained pretraining: consistently delays overfitting under repeated data exposure and improves generalization across diverse evaluation datasets. In multilingual fine-tuning: effectiveness depends on tokenizer quality - strongest gains when canonical tokens are highly compressed, diminishes when tokenizer already over-fragments input.

Conclusion: Homotokens provide a simple, modular mechanism for inducing tokenization invariance in language models, offering benefits for generalization and robustness.

Abstract: Subword tokenization introduces a computational layer in language models where many distinct token sequences decode to the same surface form and preserve meaning, yet induce different internal computations. Despite this non-uniqueness, language models are typically trained using a single canonical longest-prefix tokenization. We formalize homotokens-alternative valid subword segmentations of the same lexical item-as a strictly meaning-preserving form of data augmentation. We introduce a lightweight training architecture that conditions canonical next-token prediction on sampled homotoken variants via an auxiliary causal encoder and block-causal cross-attention, without modifying the training objective or token interface. In data-constrained pretraining, homotoken augmentation consistently delays overfitting under repeated data exposure and improves generalization across diverse evaluation datasets. In multilingual fine-tuning, we find that the effectiveness of homotokens depends on tokenizer quality: gains are strongest when canonical tokens are highly compressed and diminish when the tokenizer already over-fragments the input. Overall, homotokens provide a simple and modular mechanism for inducing tokenization invariance in language models.

Method: Human-Model Collaborative Construction pipeline: frontier LLMs draft challenging questions, reference answers, design rationales, and solution processes, then experts validate correctness and refine problematic cases. Supports fine-grained analysis with task-specific metrics and multi-dimensional taxonomy.

Result: Evaluation of 46 long-context LLMs reveals: (1) long-context optimization contributes more to comprehension than parameter scaling; (2) effective context length is shorter than claimed, with cross-lingual misalignment; (3) “thinking” paradigm primarily helps models trained with native reasoning, while mixed-thinking offers Pareto trade-off.

Conclusion: LongBench Pro provides a robust, realistic testbed for advancing long-context understanding, balancing quality with scalability through human-model collaboration.

Abstract: The rapid expansion of context length in large language models (LLMs) has outpaced existing evaluation benchmarks. Current long-context benchmarks often trade off scalability and realism: synthetic tasks underrepresent real-world complexity, while fully manual annotation is costly to scale to extreme lengths and diverse scenarios. We present LongBench Pro, a more realistic and comprehensive bilingual benchmark of 1,500 naturally occurring long-context samples in English and Chinese spanning 11 primary tasks and 25 secondary tasks, with input lengths from 8k to 256k tokens. LongBench Pro supports fine-grained analysis with task-specific metrics and a multi-dimensional taxonomy of context requirement (full vs. partial dependency), length (six levels), and difficulty (four levels calibrated by model performance). To balance quality with scalability, we propose a Human-Model Collaborative Construction pipeline: frontier LLMs draft challenging questions and reference answers, along with design rationales and solution processes, to reduce the cost of expert verification. Experts then rigorously validate correctness and refine problematic cases. Evaluating 46 widely used long-context LLMs on LongBench Pro yields three findings: (1) long-context optimization contributes more to long-context comprehension than parameter scaling; (2) effective context length is typically shorter than the claimed context length, with pronounced cross-lingual misalignment; and (3) the “thinking” paradigm helps primarily models trained with native reasoning, while mixed-thinking designs offer a promising Pareto trade-off. In summary, LongBench Pro provides a robust testbed for advancing long-context understanding.

Method: The researchers explore different methods to use LLMs as data compressors without additional model training, testing various configurations to optimize compression performance across different data types.

Result: Achieved new state-of-the-art adjusted compression rate of around 18% on enwik9 dataset. Also demonstrated LLMs can compress non-English data, code data, and byte stream sequences effectively when properly configured.

Conclusion: LLMs excel at compressing text-dominant domains and can achieve competitive compression rates on non-natural text sequences with appropriate configuration, making them promising for practical data compression applications.

Abstract: In this report, we investigate the potential use of large language models (LLM’s) in the task of data compression. Previous works have demonstrated promising results in applying LLM’s towards compressing not only text, but also a wide range of multi-modal data. Despite the favorable performance achieved, there still remains several practical questions that pose a challenge towards replacing existing data compression algorithms with LLM’s. In this work, we explore different methods to achieve a lower adjusted compression rate using LLM’s as data compressors. In comparison to previous works, we were able to achieve a new state-of-the-art (SOTA) adjusted compression rate of around $18%$ on the enwik9 dataset without additional model training. Furthermore, we explore the use of LLM’s in compressing non-English data, code data, byte stream sequences. We show that while LLM’s excel in compressing data in text-dominant domains, their ability in compressing non-natural text sequences still remain competitive if configured in the right way.

Method: Uses dependency co-occurrence patterns of words instead of neural embeddings for semantic change detection.

Result: The dependency-based method is effective and even outperforms several distributional semantic models on LSC benchmarks.

Conclusion: Dependency co-occurrence patterns provide a plausible, interpretable alternative to opaque neural methods for semantic change detection.

Abstract: Most modern computational approaches to lexical semantic change detection (LSC) rely on embedding-based distributional word representations with neural networks. Despite the strong performance on LSC benchmarks, they are often opaque. We investigate an alternative method which relies purely on dependency co-occurrence patterns of words. We demonstrate that it is effective for semantic change detection and even outperforms a number of distributional semantic models. We provide an in-depth quantitative and qualitative analysis of the predictions, showing that they are plausible and interpretable.

Method: Discover that script is linearly encoded in activation space, then modify activations at inference time using script vectors to control output script, even for unconventional language-script pairings.

Result: Achieves competitive script control performance across all Whisper model sizes, enabling post-hoc control over speech recognition output script.

Conclusion: Script control is feasible through linear manipulation of activation vectors, providing deterministic script output for multilingual speech models without retraining.

Abstract: Multilingual speech foundation models such as Whisper are trained on web-scale data, where data for each language consists of a myriad of regional varieties. However, different regional varieties often employ different scripts to write the same language, rendering speech recognition output also subject to non-determinism in the output script. To mitigate this problem, we show that script is linearly encoded in the activation space of multilingual speech models, and that modifying activations at inference time enables direct control over output script. We find the addition of such script vectors to activations at test time can induce script change even in unconventional language-script pairings (e.g. Italian in Cyrillic and Japanese in Latin script). We apply this approach to inducing post-hoc control over the script of speech recognition output, where we observe competitive performance across all model sizes of Whisper.

Method: Proposes a unified lifecycle-based taxonomy organizing LLM research into six stages: Data Preparation, Model Preparation, Training, Alignment, Inference, and Evaluation. Provides systematic review of foundational theories and internal mechanisms within this framework.

Result: Creates a structured roadmap for LLM research, analyzing core theoretical issues like mathematical justification for data mixtures, representational limits of architectures, optimization dynamics of alignment algorithms, and identifying frontier challenges.

Conclusion: The survey provides a comprehensive framework to connect empirical observations with rigorous scientific inquiry, aiming to transition LLM development from engineering heuristics toward a principled scientific discipline by addressing theoretical fragmentation.

Abstract: The rapid emergence of Large Language Models (LLMs) has precipitated a profound paradigm shift in Artificial Intelligence, delivering monumental engineering successes that increasingly impact modern society. However, a critical paradox persists within the current field: despite the empirical efficacy, our theoretical understanding of LLMs remains disproportionately nascent, forcing these systems to be treated largely as ``black boxes’’. To address this theoretical fragmentation, this survey proposes a unified lifecycle-based taxonomy that organizes the research landscape into six distinct stages: Data Preparation, Model Preparation, Training, Alignment, Inference, and Evaluation. Within this framework, we provide a systematic review of the foundational theories and internal mechanisms driving LLM performance. Specifically, we analyze core theoretical issues such as the mathematical justification for data mixtures, the representational limits of various architectures, and the optimization dynamics of alignment algorithms. Moving beyond current best practices, we identify critical frontier challenges, including the theoretical limits of synthetic data self-improvement, the mathematical bounds of safety guarantees, and the mechanistic origins of emergent intelligence. By connecting empirical observations with rigorous scientific inquiry, this work provides a structured roadmap for transitioning LLM development from engineering heuristics toward a principled scientific discipline.

Method: RAL2M repositions LLMs as query-response matching judges within a retrieval-based system. It uses query-adaptive latent ensemble strategy that models heterogeneous model competence and interdependencies among LLMs to derive calibrated consensus decisions.

Result: Extensive experiments on large-scale benchmarks demonstrate the method effectively leverages “wisdom of the crowd” and significantly outperforms strong baselines.

Conclusion: The framework provides a robust alternative to purely generative approaches by eliminating generation hallucination. Future work should explore further exploitation of latent representations.

Abstract: Hallucination is a major concern in LLM-driven service systems, necessitating explicit knowledge grounding for compliance-guaranteed responses. In this paper, we introduce Retrieval-Augmented Learning-to-Match (RAL2M), a novel framework that eliminates generation hallucination by repositioning LLMs as query-response matching judges within a retrieval-based system, providing a robust alternative to purely generative approaches. To further mitigate judgment hallucination, we propose a query-adaptive latent ensemble strategy that explicitly models heterogeneous model competence and interdependencies among LLMs, deriving a calibrated consensus decision. Extensive experiments on large-scale benchmarks demonstrate that the proposed method effectively leverages the “wisdom of the crowd” and significantly outperforms strong baselines. Finally, we discuss best practices and promising directions for further exploiting latent representations in future work.

Method: Proposed a controlled Knowledge Graph-based synthetic framework generating text from symmetric/inverse triples, trained GPT-style autoregressive models from scratch, and evaluated memorization, logical inference, and in-context generalization to unseen entities.

Result: Found a sharp phase transition where relational semantics emerge with sufficient logic-bearing supervision, even in shallow models; successful generalization aligns with stable intermediate-layer signals; order-matched forward/reverse tests and diffusion baseline show reversal failures are primarily driven by autoregressive order bias.

Conclusion: Autoregressive LLMs can learn relational semantics when given adequate logic-bearing supervision, and their reversal failures are mainly due to the inherent left-to-right order bias of autoregressive modeling rather than deficiencies in relational knowledge.

Abstract: Autoregressive LLMs perform well on relational tasks that require linking entities via relational words (e.g., father/son, friend), but it is unclear whether they learn the logical semantics of such relations (e.g., symmetry and inversion logic) and, if so, whether reversal-type failures arise from missing relational semantics or left-to-right order bias. We propose a controlled Knowledge Graph-based synthetic framework that generates text from symmetric/inverse triples, train GPT-style autoregressive models from scratch, and evaluate memorization, logical inference, and in-context generalization to unseen entities to address these questions. We find a sharp phase transition in which relational semantics emerge with sufficient logic-bearing supervision, even in shallow (2-3 layer) models, and that successful generalization aligns with stable intermediate-layer signals. Finally, order-matched forward/reverse tests and a diffusion baseline indicate that reversal failures are primarily driven by autoregressive order bias rather than deficient inversion semantics.

Method: Pearmut provides a feature-rich platform supporting standard evaluation protocols (DA, ESA, MQM), document-level context, absolute/contrastive evaluation, attention checks, ESAAI pre-annotations, and both static/active learning assignment strategies.

Result: The platform removes common entry barriers for human evaluation, enables reliable multilingual evaluation, and supports prototyping new evaluation protocols.

Conclusion: Pearmut makes human evaluation practical and routine for model development and diagnosis rather than an occasional effort, bridging the gap between automatic metrics and gold-standard human assessment.

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: 1) Identify three structural biases: exposure bias, gold availability prior (both English-centric), and cultural priors. 2) Propose DeLP metric to factor out these confounds. 3) Develop DELTA framework that strategically leverages monolingual alignment between query and document languages for optimized cross-lingual retrieval and generation.

Result: DeLP reveals that English preference is largely due to evidence distribution, not inherent model bias. Retrievers fundamentally favor monolingual alignment. DELTA consistently outperforms English pivoting and mRAG baselines across diverse languages.

Conclusion: The perceived English preference in mRAG is an artifact of evaluation biases. Monolingual alignment is key to effective cross-lingual retrieval, and DELTA provides a lightweight, efficient solution that outperforms existing approaches.

Abstract: Multilingual Retrieval-Augmented Generation (mRAG) systems often exhibit a perceived preference for high-resource languages, particularly English, resulting in the widespread adoption of English pivoting. While prior studies attribute this advantage to the superior English-centric capabilities of Large Language Models (LLMs), we find that such measurements are significantly distorted by structural priors inherent in evaluation benchmarks. Specifically, we identify exposure bias and a gold availability prior-both driven by the disproportionate concentration of resources in English-as well as cultural priors rooted in topic locality, as factors that hinder accurate assessment of genuine language preference. To address these biases, we propose DeLP (Debiased Language Preference), a calibrated metric designed to explicitly factor out these structural confounds. Our analysis using DeLP reveals that the previously reported English preference is largely a byproduct of evidence distribution rather than an inherent model bias. Instead, we find that retrievers fundamentally favor monolingual alignment between the query and the document language. Building on this insight, we introduce DELTA (DEbiased Language preference-guided Text Augmentation), a lightweight and efficient mRAG framework that strategically leverages monolingual alignment to optimize cross-lingual retrieval and generation. Experimental results demonstrate that DELTA consistently outperforms English pivoting and mRAG baselines across diverse languages.

Method: Proposes ensemble method aggregating results from ten distinct changepoint detection algorithms for superior performance and robustness. Adds LLM-powered explanation pipeline that automatically generates contextual narratives linking changepoints to real-world events. For private data, uses Retrieval-Augmented Generation (RAG) for explanations grounded in user-provided documents.

Result: Achieves superior performance and robustness compared to individual methods. The open source Python framework demonstrates practical utility in diverse domains including finance, political science, and environmental science.

Conclusion: The framework transforms raw statistical output into actionable insights for analysts and decision-makers by combining ensemble detection with LLM-powered interpretability.

Abstract: This paper introduces a novel changepoint detection framework that combines ensemble statistical methods with Large Language Models (LLMs) to enhance both detection accuracy and the interpretability of regime changes in time series data. Two critical limitations in the field are addressed. First, individual detection methods exhibit complementary strengths and weaknesses depending on data characteristics, making method selection non-trivial and prone to suboptimal results. Second, automated, contextual explanations for detected changes are largely absent. The proposed ensemble method aggregates results from ten distinct changepoint detection algorithms, achieving superior performance and robustness compared to individual methods. Additionally, an LLM-powered explanation pipeline automatically generates contextual narratives, linking detected changepoints to potential real-world historical events. For private or domain-specific data, a Retrieval-Augmented Generation (RAG) solution enables explanations grounded in user-provided documents. The open source Python framework demonstrates practical utility in diverse domains, including finance, political science, and environmental science, transforming raw statistical output into actionable insights for analysts and decision-makers.

Method: Combines advanced table and non-table detection techniques with probability-based post-processing heuristics. First applies detection algorithms to improve input data quality, then employs probabilistic error correction on OCR output.

Result: Experimental results show substantial improvement in recognition accuracy, increasing from 72.86% to 79.26%.

Conclusion: The work contributes valuable resources for Bahnar language preservation and provides a framework applicable to other minority language digitization efforts.

Abstract: Bahnar, a minority language spoken across Vietnam, Cambodia, and Laos, faces significant preservation challenges due to limited research and data availability. This study addresses the critical need for accurate digitization of Bahnar language documents through optical character recognition (OCR) technology. Digitizing scanned paper documents poses significant challenges, as degraded image quality from broken or blurred areas introduces considerable OCR errors that compromise information retrieval systems. We propose a comprehensive approach combining advanced table and non-table detection techniques with probability-based post-processing heuristics to enhance recognition accuracy. Our method first applies detection algorithms to improve input data quality, then employs probabilistic error correction on OCR output. Experimental results indicate a substantial improvement, with recognition accuracy increasing from 72.86% to 79.26%. This work contributes valuable resources for Bahnar language preservation and provides a framework applicable to other minority language digitization efforts.

Method: ReASC reframes adaptive sampling from response counting to evidence sufficiency using response-level confidence. It has two stages: 1) single-sample decision stage for confidently answerable instances, and 2) reliability-aware accumulation stage that aggregates responses using both frequency and confidence.

Result: Across five models and four datasets, ReASC achieves the best accuracy-cost trade-off compared to existing baselines, with improved inference efficiency across model scales from 3B to 27B parameters. Reduces inference cost by up to 70% relative to self-consistency while preserving accuracy on GSM8K using Gemma-3-4B-it.

Conclusion: ReASC provides a more efficient adaptive self-consistency approach by leveraging response confidence for principled information aggregation, significantly reducing computational costs while maintaining reasoning reliability.

Abstract: Self-Consistency improves reasoning reliability through multi-sample aggregation, but incurs substantial inference cost. Adaptive self-consistency methods mitigate this issue by adjusting the sampling budget; however, they rely on count-based stopping rules that treat all responses equally, often leading to unnecessary sampling. We propose Reliability-Aware Adaptive Self-Consistency (ReASC), which addresses this limitation by reframing adaptive sampling from response counting to evidence sufficiency, leveraging response-level confidence for principled information aggregation. ReASC operates in two stages: a single-sample decision stage that resolves instances confidently answerable from a single response, and a reliability-aware accumulation stage that aggregates responses by jointly leveraging their frequency and confidence. Across five models and four datasets, ReASC consistently achieves the best accuracy-cost trade-off compared to existing baselines, yielding improved inference efficiency across model scales from 3B to 27B parameters. As a concrete example, ReASC reduces inference cost by up to 70% relative to self-consistency while preserving accuracy on GSM8K using Gemma-3-4B-it.

Method: Combines supervised fine-tuning (via rejection sampling or reasoning trace reformatting) with reinforcement learning using adaptive length penalty. Uses lightweight reward function that penalizes tokens after first correct answer but encourages beneficial self-verification.

Result: Reduces response length by 28% for 8B models and 40% for 32B models with only minor performance drops (1.6 and 2.5 points). Achieves superior trade-off with 76.6 AUC_OAA score - 5 points above base model and 2.5 points above second-best approach.

Conclusion: The method effectively addresses overthinking in LLMs, achieving better accuracy-length trade-off than complex state-of-the-art methods through conceptually simple multi-stage training with adaptive length penalty.

Abstract: The reasoning capabilities of large language models (LLMs) have improved substantially through increased test-time computation, typically in the form of intermediate tokens known as chain-of-thought (CoT). However, CoT often becomes unnecessarily long, increasing computation cost without actual accuracy gains or sometimes even degrading performance, a phenomenon known as ``overthinking’’. We propose a multi-stage efficient reasoning method that combines supervised fine-tuning – via rejection sampling or reasoning trace reformatting – with reinforcement learning using an adaptive length penalty. We introduce a lightweight reward function that penalizes tokens generated after the first correct answer but encouraging self-verification only when beneficial. We conduct a holistic evaluation across seven diverse reasoning tasks, analyzing the accuracy-response length trade-off. Our approach reduces response length by an average of 28% for 8B models and 40% for 32B models, while incurring only minor performance drops of 1.6 and 2.5 points, respectively. Despite its conceptual simplicity, it achieves a superior trade-off compared to more complex state-of-the-art efficient reasoning methods, scoring 76.6, in terms of the area under the Overthinking-Adjusted Accuracy curve ($\text{AUC}_{\text{OAA}}$) – 5 points above the base model and 2.5 points above the second-best approach.

Method: A Sparse Autoencoder-based framework with contrastive feature retrieval pipeline using controlled semantic oppositions, combining statistical activation analysis and generation-based validation to distill monosemantic functional features from sparse activation spaces.

Result: The method enables precise bidirectional steering of model behavior (demonstrated with Big Five personality traits) with superior stability and performance compared to existing methods like CAA. Identifies “Functional Faithfulness” effect where feature intervention induces coherent shifts across multiple linguistic dimensions.

Conclusion: LLMs internalize deeply integrated representations of high-order concepts, and the framework provides a novel, robust mechanistic path for regulating complex AI behaviors.

Abstract: Recent work in Mechanistic Interpretability (MI) has enabled the identification and intervention of internal features in Large Language Models (LLMs). However, a persistent challenge lies in linking such internal features to the reliable control of complex, behavior-level semantic attributes in language generation. In this paper, we propose a Sparse Autoencoder-based framework for retrieving and steering semantically interpretable internal features associated with high-level linguistic behaviors. Our method employs a contrastive feature retrieval pipeline based on controlled semantic oppositions, combing statistical activation analysis and generation-based validation to distill monosemantic functional features from sparse activation spaces. Using the Big Five personality traits as a case study, we demonstrate that our method enables precise, bidirectional steering of model behavior while maintaining superior stability and performance compared to existing activation steering methods like Contrastive Activation Addition (CAA). We further identify an empirical effect, which we term Functional Faithfulness, whereby intervening on a specific internal feature induces coherent and predictable shifts across multiple linguistic dimensions aligned with the target semantic attribute. Our findings suggest that LLMs internalize deeply integrated representations of high-order concepts, and provide a novel, robust mechanistic path for the regulation of complex AI behaviors.

Method: Proposes P-Check framework with a plug-and-play checklist generator that synthesizes dynamic evaluation criteria. Introduces Preference-Contrastive Criterion Weighting to assign saliency scores to criteria based on their discriminative power for personalized judgment.

Result: Extensive experiments show P-Check improves reward accuracy, enhances downstream personalized generation, and remains robust in out-of-distribution (OOD) scenarios.

Conclusion: P-Check effectively addresses limitations of static user context modeling by generating dynamic, personalized evaluation checklists, leading to better alignment with individual preferences and improved performance across various scenarios.

Abstract: Recent approaches in personalized reward modeling have primarily focused on leveraging user interaction history to align model judgments with individual preferences. However, existing approaches largely treat user context as a static or implicit conditioning signal, failing to capture the dynamic and multi-faceted nature of human judgment. In this paper, we propose P-Check, a novel personalized reward modeling framework, designed to train a plug-and-play checklist generator that synthesizes dynamic evaluation criteria for guiding the reward prediction. To better align these checklists with personalized nuances, we introduce Preference-Contrastive Criterion Weighting, a training strategy that assigns saliency scores to criteria based on their discriminative power for personalized judgment. We conduct extensive experiments and demonstrate that P-Check not only improves reward accuracy but also enhances downstream personalized generation, and remains robust in OOD scenarios.

Method: Proposes a System-2 cognitive process-inspired test-time strategy that decomposes large counting tasks into smaller, independent sub-problems. Uses observational and causal mediation analyses to understand the underlying mechanism, identifying key components: latent counts stored in final item representations, transferred via dedicated attention heads, and aggregated in final stages.

Result: The strategy enables LLMs to surpass architectural limitations and achieve high accuracy on large-scale counting tasks. Mechanistic analysis reveals how latent counts are computed, stored, transferred, and aggregated within the model architecture.

Conclusion: This work provides mechanistic insight into System-2 counting in LLMs and presents a generalizable approach for improving and understanding their reasoning behavior, offering a solution to transformer architectural limitations in counting tasks.

Abstract: Large language models (LLMs), despite strong performance on complex mathematical problems, exhibit systematic limitations in counting tasks. This issue arises from architectural limits of transformers, where counting is performed across layers, leading to degraded precision for larger counting problems due to depth constraints. To address this limitation, we propose a simple test-time strategy inspired by System-2 cognitive processes that decomposes large counting tasks into smaller, independent sub-problems that the model can reliably solve. We evaluate this approach using observational and causal mediation analyses to understand the underlying mechanism of this System-2-like strategy. Our mechanistic analysis identifies key components: latent counts are computed and stored in the final item representations of each part, transferred to intermediate steps via dedicated attention heads, and aggregated in the final stage to produce the total count. Experimental results demonstrate that this strategy enables LLMs to surpass architectural limitations and achieve high accuracy on large-scale counting tasks. This work provides mechanistic insight into System-2 counting in LLMs and presents a generalizable approach for improving and understanding their reasoning behavior.

Method: Stable-RAG runs the generator under multiple retrieval orders, clusters hidden states to identify dominant reasoning patterns, decodes from cluster-center representations, and uses these results to align hallucinated outputs toward correct answers for consistent predictions.

Result: Experiments on three QA datasets show Stable-RAG significantly improves answer accuracy, reasoning consistency, and robust generalization across datasets, retrievers, and input lengths compared to baselines.

Conclusion: Permutation sensitivity is a critical but overlooked issue in RAG systems, and Stable-RAG provides an effective solution that enhances both accuracy and consistency by leveraging permutation sensitivity estimation to stabilize model outputs.

Abstract: Retrieval-Augmented Generation (RAG) has become a key paradigm for reducing factual hallucinations in large language models (LLMs), yet little is known about how the order of retrieved documents affects model behavior. We empirically show that under Top-5 retrieval with the gold document included, LLM answers vary substantially across permutations of the retrieved set, even when the gold document is fixed in the first position. This reveals a previously underexplored sensitivity to retrieval permutations. Although robust RAG methods primarily focus on enhancing LLM robustness to low-quality retrieval and mitigating positional bias to distribute attention fairly over long contexts, neither approach directly addresses permutation sensitivity. In this paper, we propose Stable-RAG, which exploits permutation sensitivity estimation to mitigate permutation-induced hallucinations. Stable-RAG runs the generator under multiple retrieval orders, clusters hidden states, and decodes from a cluster-center representation that captures the dominant reasoning pattern. It then uses these reasoning results to align hallucinated outputs toward the correct answer, encouraging the model to produce consistent and accurate predictions across document permutations. Experiments on three QA datasets show that Stable-RAG significantly improves answer accuracy, reasoning consistency and robust generalization across datasets, retrievers, and input lengths compared with baselines.

Method: Used truncation-based strategy across 11 languages, examining how correct answers emerge with partial reasoning traces. Performed representational analyses to understand internal mechanisms and cross-language consistency.

Result: Clear evidence of multilingual latent reasoning, but uneven: strong in resource-rich languages, weaker in low-resource ones, and less observable on harder benchmarks. Internal prediction evolution is highly consistent across languages and aligns with English patterns.

Conclusion: LRMs exhibit multilingual latent reasoning that follows an English-centered pathway, suggesting that despite surface-level language differences, the internal reasoning mechanisms are fundamentally aligned with English patterns.

Abstract: Large reasoning models (LRMs) achieve strong performance on mathematical reasoning tasks, often attributed to their capability to generate explicit chain-of-thought (CoT) explanations. However, recent work shows that LRMs often arrive at the correct answer before completing these textual reasoning steps, indicating the presence of latent reasoning – internal, non-verbal computation encoded in hidden states. While this phenomenon has been explored in English, its multilingual behavior remains largely unknown. In this paper, we conduct a systematic investigation of multilingual latent reasoning in LRMs across 11 languages. Using a truncation-based strategy, we examine how the correct answer emerges as the model is given only partial reasoning traces, allowing us to measure stepwise latent prediction formation. Our results reveal clear evidence of multilingual latent reasoning, though unevenly: strong in resource-rich languages, weaker in low-resource ones, and broadly less observable on harder benchmarks. To understand whether these differences reflect distinct internal mechanisms, we further perform representational analyses. Despite surface-level disparities, we find that the internal evolution of predictions is highly consistent across languages and broadly aligns with English – a pattern suggesting an English-centered latent reasoning pathway.

Method: SentGraph constructs a hierarchical sentence graph offline using Rhetorical Structure Theory to distinguish nucleus and satellite sentences, organizing them into topic-level subgraphs with cross-document entity bridges. During online retrieval, it performs graph-guided evidence selection and path expansion to retrieve fine-grained sentence-level evidence.

Result: Extensive experiments on four multi-hop question answering benchmarks demonstrate the effectiveness of SentGraph, validating the importance of explicitly modeling sentence-level logical dependencies for multi-hop reasoning.

Conclusion: SentGraph successfully addresses multi-hop QA challenges by modeling fine-grained logical relationships between sentences, outperforming traditional chunk-based RAG approaches through its graph-based evidence retrieval framework.

Abstract: Traditional Retrieval-Augmented Generation (RAG) effectively supports single-hop question answering with large language models but faces significant limitations in multi-hop question answering tasks, which require combining evidence from multiple documents. Existing chunk-based retrieval often provides irrelevant and logically incoherent context, leading to incomplete evidence chains and incorrect reasoning during answer generation. To address these challenges, we propose SentGraph, a sentence-level graph-based RAG framework that explicitly models fine-grained logical relationships between sentences for multi-hop question answering. Specifically, we construct a hierarchical sentence graph offline by first adapting Rhetorical Structure Theory to distinguish nucleus and satellite sentences, and then organizing them into topic-level subgraphs with cross-document entity bridges. During online retrieval, SentGraph performs graph-guided evidence selection and path expansion to retrieve fine-grained sentence-level evidence. Extensive experiments on four multi-hop question answering benchmarks demonstrate the effectiveness of SentGraph, validating the importance of explicitly modeling sentence-level logical dependencies for multi-hop reasoning.

Method: MMFormalizer uses adaptive grounding with real-world entities, recursively constructs formal propositions from perceptually grounded primitives through recursive grounding and axiom composition, with adaptive recursive termination ensuring visual evidence support.

Result: Evaluation on PhyX-AF benchmark (115 samples) shows frontier models like GPT-5 and Gemini-3-Pro achieve highest compile and semantic accuracy, with GPT-5 excelling in physical reasoning, while geometry remains most challenging domain.

Conclusion: MMFormalizer provides a scalable framework for unified multimodal autoformalization, bridging perception and formal reasoning, and is the first multimodal autoformalization method capable of handling classical mechanics, relativity, quantum mechanics, and thermodynamics.

Abstract: Autoformalization, which translates natural language mathematics into formal statements to enable machine reasoning, faces fundamental challenges in the wild due to the multimodal nature of the physical world, where physics requires inferring hidden constraints (e.g., mass or energy) from visual elements. To address this, we propose MMFormalizer, which extends autoformalization beyond text by integrating adaptive grounding with entities from real-world mathematical and physical domains. MMFormalizer recursively constructs formal propositions from perceptually grounded primitives through recursive grounding and axiom composition, with adaptive recursive termination ensuring that every abstraction is supported by visual evidence and anchored in dimensional or axiomatic grounding. We evaluate MMFormalizer on a new benchmark, PhyX-AF, comprising 115 curated samples from MathVerse, PhyX, Synthetic Geometry, and Analytic Geometry, covering diverse multimodal autoformalization tasks. Results show that frontier models such as GPT-5 and Gemini-3-Pro achieve the highest compile and semantic accuracy, with GPT-5 excelling in physical reasoning, while geometry remains the most challenging domain. Overall, MMFormalizer provides a scalable framework for unified multimodal autoformalization, bridging perception and formal reasoning. To the best of our knowledge, this is the first multimodal autoformalization method capable of handling classical mechanics (derived from the Hamiltonian), as well as relativity, quantum mechanics, and thermodynamics. More details are available on our project page: MMFormalizer.github.io

Method: Introduces Dementia-R1, an RL-based framework with Cold-Start RL strategy that pre-trains the model to predict verifiable clinical indices extracted from patient histories, enhancing disease progression reasoning before determining final clinical status.

Result: Achieves 77.03% F1 score on real-world unstructured clinical datasets. On ADNI benchmark, their 7B model rivals GPT-4o in capturing fluctuating cognitive trajectories.

Conclusion: Dementia-R1 effectively addresses longitudinal dementia prognosis challenges by combining RL with clinical index pre-training, demonstrating strong performance and competitive results against larger models like GPT-4o.

Abstract: While Large Language Models (LLMs) have shown strong performance on clinical text understanding, they struggle with longitudinal prediction tasks such as dementia prognosis, which require reasoning over complex, non-monotonic symptom trajectories across multiple visits. Standard supervised training lacks explicit annotations for symptom evolution, while direct Reinforcement Learning (RL) is hindered by sparse binary rewards. To address this challenge, we introduce Dementia-R1, an RL-based framework for longitudinal dementia prognosis from unstructured clinical notes. Our approach adopts a Cold-Start RL strategy that pre-trains the model to predict verifiable clinical indices extracted from patient histories, enhancing the capability to reason about disease progression before determining the final clinical status. Extensive experiments demonstrate that Dementia-R1 achieves an F1 score of 77.03% on real-world unstructured clinical datasets. Notably, on the ADNI benchmark, our 7B model rivals GPT-4o, effectively capturing fluctuating cognitive trajectories. Code is available at https://anonymous.4open.science/r/dementiar1-CDB5

Method: Uses multi-agent system to synthesize realistic patient cases without real EHR data; Patient Agent with atomic medical facts and hallucination detection; structured rubric-generation pipeline using EBM guidelines and reject sampling for “must-ask” items.

Result: Comprehensive evaluation shows current models face substantial challenges across multiple assessment dimensions, indicating that improving medical dialogue requires advances in dialogue management architectures beyond just base-model tuning.

Conclusion: MedDialogRubrics provides a privacy-preserving, rigorous benchmark for assessing medical LLMs’ diagnostic capabilities, revealing that architectural improvements in dialogue management are needed rather than just incremental model tuning.

Abstract: Medical conversational AI (AI) plays a pivotal role in the development of safer and more effective medical dialogue systems. However, existing benchmarks and evaluation frameworks for assessing the information-gathering and diagnostic reasoning abilities of medical large language models (LLMs) have not been rigorously evaluated. To address these gaps, we present MedDialogRubrics, a novel benchmark comprising 5,200 synthetically constructed patient cases and over 60,000 fine-grained evaluation rubrics generated by LLMs and subsequently refined by clinical experts, specifically designed to assess the multi-turn diagnostic capabilities of LLM. Our framework employs a multi-agent system to synthesize realistic patient records and chief complaints from underlying disease knowledge without accessing real-world electronic health records, thereby mitigating privacy and data-governance concerns. We design a robust Patient Agent that is limited to a set of atomic medical facts and augmented with a dynamic guidance mechanism that continuously detects and corrects hallucinations throughout the dialogue, ensuring internal coherence and clinical plausibility of the simulated cases. Furthermore, we propose a structured LLM-based and expert-annotated rubric-generation pipeline that retrieves Evidence-Based Medicine (EBM) guidelines and utilizes the reject sampling to derive a prioritized set of rubric items (“must-ask” items) for each case. We perform a comprehensive evaluation of state-of-the-art models and demonstrate that, across multiple assessment dimensions, current models face substantial challenges. Our results indicate that improving medical dialogue will require advances in dialogue management architectures, not just incremental tuning of the base-model.

Method: Created LittiChoQA dataset with automatically generated QA pairs from naturally authored literary texts. Evaluated multilingual LLMs on non-factoid abstractive QA under full-context and context-shortened settings (paragraph selection and vector retrieval).

Result: Full-context fine-tuning yields highest token-level and semantic-level scores. Krutrim-2 achieves best performance with 76.1 semantic score with full context. Context shortening improves throughput but reduces performance (74.9 with paragraph selection, 71.4 with vector retrieval).

Conclusion: Clear trade-off between performance and efficiency in long-context QA for Indic languages. Full-context yields best results but context shortening improves throughput. Krutrim-2 shows strongest performance among evaluated models.

Abstract: Long-context question answering (QA) over literary texts poses significant challenges for modern large language models, particularly in low-resource languages. We address the scarcity of long-context QA resources for Indic languages by introducing LittiChoQA, the largest literary QA dataset to date covering many languages spoken in the Gangetic plains of India. The dataset comprises over 270K automatically generated question-answer pairs with a balanced distribution of factoid and non-factoid questions, generated from naturally authored literary texts collected from the open web. We evaluate multiple multilingual LLMs on non-factoid, abstractive QA, under both full-context and context-shortened settings. Results demonstrate a clear trade-off between performance and efficiency: full-context fine-tuning yields the highest token-level and semantic-level scores, while context shortening substantially improves throughput. Among the evaluated models, Krutrim-2 achieves the strongest performance, obtaining a semantic score of 76.1 with full context. While, in shortened context settings it scores 74.9 with answer paragraph selection and 71.4 with vector-based retrieval. Qualitative evaluations further corroborate these findings.

Method: F-DPO extends DPO by: (1) applying label-flipping to correct misordered preference pairs, (2) adding factuality-aware margins to emphasize clear correctness differences, and (3) constructing factuality-aware preference data with binary factuality indicators and synthetic hallucinated variants.

Result: Across seven LLMs (1B-14B), F-DPO consistently improves factuality and reduces hallucinations. Qwen3-8B saw 5x reduction in hallucination rates (0.424 to 0.084) and 50% improvement in factuality scores (5.26 to 7.90). On TruthfulQA, Qwen2.5-14B achieved +17% MC1 accuracy (0.500 to 0.585) and +49% MC2 accuracy (0.357 to 0.531).

Conclusion: F-DPO effectively improves LLM factuality and reduces hallucinations using simple binary factuality labels, requiring no auxiliary reward models, token-level annotations, or multi-stage training, while generalizing well to out-of-distribution benchmarks.

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: Built models on diverse Transformer architectures (GPT, Mistral, Llama2, Mixtral, Magistral), either pretrained from scratch or continually pretrained on 25B-88.45B tokens using Norwegian-extended tokenizers and advanced post-training strategies.

Result: Instruction-tuned variants (Mistral-7B-Instruct and Mixtral-8x7B-Instruct) demonstrate strong assistant-style capabilities, showing potential for practical deployment in interactive and domain-specific applications.

Conclusion: The NorwAI large language models are openly available to Nordic organizations, companies, and students for research and experimental use, with detailed documentation provided on architectures, training, and evaluations.

Abstract: Norwegian, spoken by approximately five million people, remains underrepresented in many of the most significant breakthroughs in Natural Language Processing (NLP). To address this gap, the NorLLM team at NorwAI has developed a family of models specifically tailored to Norwegian and other Scandinavian languages, building on diverse Transformer-based architectures such as GPT, Mistral, Llama2, Mixtral and Magistral. These models are either pretrained from scratch or continually pretrained on 25B - 88.45B tokens, using a Norwegian-extended tokenizer and advanced post-training strategies to optimize performance, enhance robustness, and improve adaptability across various real-world tasks. Notably, instruction-tuned variants (e.g., Mistral-7B-Instruct and Mixtral-8x7B-Instruct) showcase strong assistant-style capabilities, underscoring their potential for practical deployment in interactive and domain-specific applications. The NorwAI large language models are openly available to Nordic organizations, companies and students for both research and experimental use. This report provides detailed documentation of the model architectures, training data, tokenizer design, fine-tuning strategies, deployment, and evaluations.

Method: Two approaches: 1) BaseCal-ReEval: feeds PoLLM responses into base LLM to get average probabilities as confidence; 2) BaseCal-Proj: trains lightweight projection to map PoLLM hidden states to base LLM states, then uses base LLM’s output layer for confidence scoring. Both are unsupervised and plug-and-play.

Result: Experiments across five datasets and three LLM families show BaseCal reduces Expected Calibration Error (ECE) by an average of 42.90% compared to the best unsupervised baselines.

Conclusion: BaseCal effectively calibrates PoLLM confidence using base LLMs as reference, addressing overconfidence issues without requiring human labels or LLM modifications, making it a practical solution for improving trust in LLM outputs.

Abstract: Reliable confidence is essential for trusting the outputs of LLMs, yet widely deployed post-trained LLMs (PoLLMs) typically compromise this trust with severe overconfidence. In contrast, we observe that their corresponding base LLMs often remain well-calibrated. This naturally motivates us to calibrate PoLLM confidence using the base LLM as a reference. This work proposes two ways to achieve this. A straightforward solution, BaseCal-ReEval, evaluates PoLLM’s responses by feeding them into the base LLM to get average probabilities as confidence. While effective, this approach introduces additional inference overhead. To address this, we propose BaseCal-Proj, which trains a lightweight projection to map the final-layer hidden states of PoLLMs back to those of their base LLMs. These projected states are then processed by the base LLM’s output layer to derive base-calibrated confidence for PoLLM’s responses. Notably, BaseCal is an unsupervised, plug-and-play solution that operates without human labels or LLM modifications. Experiments across five datasets and three LLM families demonstrate the effectiveness of BaseCal, reducing Expected Calibration Error (ECE) by an average of 42.90% compared to the best unsupervised baselines.

Method: Proposes an early-stopping algorithm that detects the threshold where information loss exceeds information gain during sparse decoding, allowing termination before excessive token generation.

Result: Early-stopping reduces token consumption by up to 90% with marginal accuracy degradation (<2%) across reasoning-intensive benchmarks, effectively mitigating the “Less is Less” problem.

Conclusion: Sparse attention can paradoxically increase overall complexity, but the proposed early-stopping algorithm successfully addresses this by balancing information loss against gain, significantly improving inference efficiency.

Abstract: Large language models (LLMs) demonstrate strong capabilities across a wide range of complex tasks and are increasingly deployed at scale, placing significant demands on inference efficiency. Prior work typically decomposes inference into prefill and decode stages, with the decode stage dominating total latency. To reduce time and memory complexity in the decode stage, a line of work introduces sparse-attention algorithms. In this paper, we show, both empirically and theoretically, that sparse attention can paradoxically increase end-to-end complexity: information loss often induces significantly longer sequences, a phenomenon we term ``Less is Less’’ (Lil). To mitigate the Lil problem, we propose an early-stopping algorithm that detects the threshold where information loss exceeds information gain during sparse decoding. Our early-stopping algorithm reduces token consumption by up to 90% with a marginal accuracy degradation of less than 2% across reasoning-intensive benchmarks.

Method: Represent each dialogue turn as a node encoded with sentence transformer. Create two edge types: shared-entity edges (connect turns referring to same entities) and temporal edges (connect contiguous turns). Use message passing to update node embeddings, then combine them with attention pooling into single vector for classification of hallucination presence and type.

Result: Method offers slightly improved performance over existing methods. Attention mechanism provides interpretability by justifying decision making process.

Conclusion: Graph-based approach with temporal modeling and entity-aware connections effectively detects dialogue-level hallucinations while providing interpretable results through attention mechanisms.

Abstract: Hallucinations can be produced by conversational AI systems, particularly in multi-turn conversations where context changes and contradictions may eventually surface. By representing the entire conversation as a temporal graph, we present a novel graph-based method for detecting dialogue-level hallucinations. Our framework models each dialogue as a node, encoding it using a sentence transformer. We explore two different ways of connectivity: i) shared-entity edges, which connect turns that refer to the same entities; ii) temporal edges, which connect contiguous turns in the conversation. Message-passing is used to update the node embeddings, allowing flow of information between related nodes. The context-aware node embeddings are then combined using attention pooling into a single vector, which is then passed on to a classifier to determine the presence and type of hallucinations. We demonstrate that our method offers slightly improved performance over existing methods. Further, we show the attention mechanism can be used to justify the decision making process. The code and model weights are made available at: https://github.com/sambuaneesh/anlp-project.

Method: 1) Extends layer-wise relevance propagation from token to semantic level to construct internal reasoning graph based on attribution vectors for faithful semantic dependency representation. 2) Designs general framework using small pre-trained language model to leverage LLM reasoning dependencies for training and detection, with dynamic threshold adjustment for correct sample pass rate.

Result: Experimental results show the method achieves better overall performance compared to state-of-the-art baselines on RAGTruth and Dolly-15k datasets.

Conclusion: The semantic-level internal reasoning graph approach effectively detects faithfulness hallucinations in RAG systems by better capturing LLMs’ internal reasoning processes and dependencies.

Abstract: The Retrieval-augmented generation (RAG) system based on Large language model (LLM) has made significant progress. It can effectively reduce factuality hallucinations, but faithfulness hallucinations still exist. Previous methods for detecting faithfulness hallucinations either neglect to capture the models’ internal reasoning processes or handle those features coarsely, making it difficult for discriminators to learn. This paper proposes a semantic-level internal reasoning graph-based method for detecting faithfulness hallucination. Specifically, we first extend the layer-wise relevance propagation algorithm from the token level to the semantic level, constructing an internal reasoning graph based on attribution vectors. This provides a more faithful semantic-level representation of dependency. Furthermore, we design a general framework based on a small pre-trained language model to utilize the dependencies in LLM’s reasoning for training and hallucination detection, which can dynamically adjust the pass rate of correct samples through a threshold. Experimental results demonstrate that our method achieves better overall performance compared to state-of-the-art baselines on RAGTruth and Dolly-15k.

Method: Propose greedy pruning - a likelihood-preserving deletion procedure that iteratively removes reasoning tokens whose removal minimally degrades model likelihood under a specified objective, yielding length-controlled reasoning chains.

Result: Students trained on pruned chains outperform frontier-model-supervised compression baseline at matched reasoning lengths. Analysis reveals systematic pruning patterns and shows attention scores can predict greedy pruning ranks.

Conclusion: Models encode a nontrivial functional importance structure over reasoning tokens, and greedy pruning provides an effective diagnostic tool for understanding and compressing LLM reasoning chains while maintaining performance.

Abstract: Large language models solve complex tasks by generating long reasoning chains, achieving higher accuracy at the cost of increased computational cost and reduced ability to isolate functionally relevant reasoning. Prior work on compact reasoning shortens such chains through probabilistic sampling, heuristics, or supervision from frontier models, but offers limited insight into whether models internally encode token-level functional importance for answer generation. We address this gap diagnostically and propose greedy pruning, a likelihood-preserving deletion procedure that iteratively removes reasoning tokens whose removal minimally degrades model likelihood under a specified objective, yielding length-controlled reasoning chains. We evaluate pruned reasoning in a distillation framework and show that students trained on pruned chains outperform a frontier-model-supervised compression baseline at matched reasoning lengths. Finally, our analysis reveals systematic pruning patterns and shows that attention scores can predict greedy pruning ranks, further suggesting that models encode a nontrivial functional importance structure over reasoning tokens.

Method: Two pragmatic inference methods that facilitate LLMs to: 1) diagnose morally benign and hazardous input, and 2) correct moral errors. The methods offer a unified perspective by designing pragmatic inference procedures grounded in inferential loads rather than modeling diverse surface forms.

Result: Empirical evidence demonstrates that the pragmatic methods can enhance moral sensitivity in LLMs and achieve strong performance on representative morality-relevant benchmarks.

Conclusion: The proposed pragmatic inference methods represent a step toward enhancing moral sensitivity in LLMs, addressing the challenge of enabling moral sensitivity through principled inference procedures rather than surface-level modeling.

Abstract: Moral sensitivity is fundamental to human moral competence, as it guides individuals in regulating everyday behavior. Although many approaches seek to align large language models (LLMs) with human moral values, how to enable them morally sensitive has been extremely challenging. In this paper, we take a step toward answering the question: how can we enhance moral sensitivity in LLMs? Specifically, we propose two pragmatic inference methods that faciliate LLMs to diagnose morally benign and hazardous input and correct moral errors, whereby enhancing LLMs’ moral sensitivity. A central strength of our pragmatic inference methods is their unified perspective: instead of modeling moral discourses across semantically diverse and complex surface forms, they offer a principled perspective for designing pragmatic inference procedures grounded in their inferential loads. Empirical evidence demonstrates that our pragmatic methods can enhance moral sensitivity in LLMs and achieves strong performance on representative morality-relevant benchmarks.

Method: Grad-ELLM aggregates channel importance from gradients of output logits with respect to attention layers and spatial importance from attention maps to generate attribution heatmaps at each generation step without architectural modifications.

Result: Grad-ELLM consistently achieves superior faithfulness compared to other attribution methods across sentiment classification, question answering, and open-generation tasks using different models.

Conclusion: Grad-ELLM provides a faithful gradient-based attribution method specifically designed for decoder-only transformer LLMs, with improved evaluation metrics for fair comparison of attribution methods.

Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks, yet their black-box nature raises concerns about transparency and faithfulness. Input attribution methods aim to highlight each input token’s contributions to the model’s output, but existing approaches are typically model-agnostic, and do not focus on transformer-specific architectures, leading to limited faithfulness. To address this, we propose Grad-ELLM, a gradient-based attribution method for decoder-only transformer-based LLMs. By aggregating channel importance from gradients of the output logit with respect to attention layers and spatial importance from attention maps, Grad-ELLM generates heatmaps at each generation step without requiring architectural modifications. Additionally, we introduce two faithfulneses metrics $π$-Soft-NC and $π$-Soft-NS, which are modifications of Soft-NC/NS that provide fairer comparisons by controlling the amount of information kept when perturbing the text. We evaluate Grad-ELLM on sentiment classification, question answering, and open-generation tasks using different models. Experiment results show that Grad-ELLM consistently achieves superior faithfulness than other attribution methods.

Method: Researchers modeled humor preference heterogeneity in Oogiri (Japanese creative response game) by clustering users based on voting logs and estimating cluster-specific weights over interpretable preference factors using Bradley-Terry-Luce models. They elicited preference judgments from LLMs by prompting them to select the funnier response.

Result: User clusters exhibited distinct preference patterns, and LLM results resembled those of particular clusters. The study demonstrated that through persona prompting, LLM preferences can be directed toward specific user clusters.

Conclusion: LLMs can capture diverse humor preferences and can be guided to mimic specific user clusters through persona prompting, providing a framework for more nuanced humor evaluation that accounts for preference heterogeneity.

Abstract: Humor preferences vary widely across individuals and cultures, complicating the evaluation of humor using large language models (LLMs). In this study, we model heterogeneity in humor preferences in Oogiri, a Japanese creative response game, by clustering users with voting logs and estimating cluster-specific weights over interpretable preference factors using Bradley-Terry-Luce models. We elicit preference judgments from LLMs by prompting them to select the funnier response and found that user clusters exhibit distinct preference patterns and that the LLM results can resemble those of particular clusters. Finally, we demonstrate that, by persona prompting, LLM preferences can be directed toward a specific cluster. The scripts for data collection and analysis will be released to support reproducibility.

Method: ToxiGAN uses a two-step directional training strategy with LLM-generated neutral texts as semantic ballast. It dynamically selects neutral exemplars for balanced guidance and explicitly optimizes toxic samples to diverge from these exemplars, reinforcing class-specific contrastive signals.

Result: Experiments on four hate speech benchmarks show ToxiGAN achieves the strongest average performance in both macro-F1 and hate-F1, consistently outperforming traditional and LLM-based augmentation methods.

Conclusion: ToxiGAN effectively addresses GAN-based augmentation issues like mode collapse and semantic drift through semantic ballast and directional training, enhancing classifier robustness for toxicity detection.

Abstract: Augmenting toxic language data in a controllable and class-specific manner is crucial for improving robustness in toxicity classification, yet remains challenging due to limited supervision and distributional skew. We propose ToxiGAN, a class-aware text augmentation framework that combines adversarial generation with semantic guidance from large language models (LLMs). To address common issues in GAN-based augmentation such as mode collapse and semantic drift, ToxiGAN introduces a two-step directional training strategy and leverages LLM-generated neutral texts as semantic ballast. Unlike prior work that treats LLMs as static generators, our approach dynamically selects neutral exemplars to provide balanced guidance. Toxic samples are explicitly optimized to diverge from these exemplars, reinforcing class-specific contrastive signals. Experiments on four hate speech benchmarks show that ToxiGAN achieves the strongest average performance in both macro-F1 and hate-F1, consistently outperforming traditional and LLM-based augmentation methods. Ablation and sensitivity analyses further confirm the benefits of semantic ballast and directional training in enhancing classifier robustness.

Method: Used a controlled LLM-to-LLM simulation framework across multi-turn scam scenarios. Evaluated eight state-of-the-art models in English and Chinese, analyzing dialogue outcomes and qualitatively annotating attacker strategies, defensive responses, and failure modes.

Result: Scam interactions follow recurrent escalation patterns, while defenses employ verification and delay mechanisms. Interactional failures frequently stem from safety guardrail activation and role instability. Eight models were evaluated across both languages.

Conclusion: Multi-turn interactional safety is a critical, distinct dimension of LLM behavior that requires specialized evaluation beyond single-turn safety assessments.

Abstract: As LLMs gain persuasive agentic capabilities through extended dialogues, they introduce novel risks in multi-turn conversational scams that single-turn safety evaluations fail to capture. We systematically study these risks using a controlled LLM-to-LLM simulation framework across multi-turn scam scenarios. Evaluating eight state-of-the-art models in English and Chinese, we analyze dialogue outcomes and qualitatively annotate attacker strategies, defensive responses, and failure modes. Results reveal that scam interactions follow recurrent escalation patterns, while defenses employ verification and delay mechanisms. Furthermore, interactional failures frequently stem from safety guardrail activation and role instability. Our findings highlight multi-turn interactional safety as a critical, distinct dimension of LLM behavior.

Method: Augment curated parallel datasets with synthetic sentence pairs from multilingual translation models, fine-tune mBART, apply language-specific preprocessing (orthographic normalization, noise-aware filtering), and evaluate with chrF++ metric.

Result: Synthetic data augmentation consistently improves chrF++ scores for Guarani-Spanish and Quechua-Spanish translation, but generic preprocessing has limitations for highly agglutinative languages like Aymara.

Conclusion: Synthetic data generation is effective for low-resource indigenous language NMT, but language-specific approaches are needed for highly agglutinative languages.

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: Conducted systematic dataset audit of widely used VLA corpora, quantifying instruction language along multiple dimensions: lexical variety, duplication and overlap, semantic similarity, and syntactic complexity.

Result: Analysis shows many datasets rely on highly repetitive, template-like commands with limited structural variation, yielding a narrow distribution of instruction forms rather than diverse language coverage.

Conclusion: Findings provide descriptive documentation of language signals in current VLA data to support better dataset reporting, principled dataset selection, and targeted curation/augmentation strategies that broaden language coverage.

Abstract: Language plays a critical role in Vision-Language-Action (VLA) models, yet the linguistic characteristics of the datasets used to train and evaluate these systems remain poorly documented. In this work, we present a systematic dataset audit of several widely used VLA corpora, aiming to characterize what kinds of instructions these datasets actually contain and how much linguistic variety they provide. We quantify instruction language along complementary dimensions-including lexical variety, duplication and overlap, semantic similarity, and syntactic complexity. Our analysis shows that many datasets rely on highly repetitive, template-like commands with limited structural variation, yielding a narrow distribution of instruction forms. We position these findings as descriptive documentation of the language signal available in current VLA training and evaluation data, intended to support more detailed dataset reporting, more principled dataset selection, and targeted curation or augmentation strategies that broaden language coverage.

Method: Developed a self-verification model trained on a newly constructed dataset that faithfully replicates the authentic exam format and evaluation scale. The approach maintains original question structure and scoring rules without alteration.

Result: The model exceeded the official passing score when evaluated on the actual exam scale, marking the first demonstration of an LLM passing the Japanese bar examination without altering original structure or scoring. Extensive comparisons showed alternative strategies (multi-agent inference and decomposition-based supervision) failed to achieve comparable performance.

Conclusion: Format-faithful supervision and consistency verification are crucial for high-stakes professional reasoning tasks. Carefully designed single-model approaches can outperform more complex systems, highlighting the importance of maintaining authentic exam structure and evaluation criteria.

Abstract: Despite rapid advances in large language models (LLMs), achieving reliable performance on highly professional and structured examinations remains a significant challenge. The Japanese bar examination is a particularly demanding benchmark, requiring not only advanced legal reasoning but also strict adherence to complex answer formats that involve joint evaluation of multiple propositions. While recent studies have reported improvements by decomposing such questions into simpler true–false judgments, these approaches have not been systematically evaluated under the original exam format and scoring scheme, leaving open the question of whether they truly capture exam-level competence. In this paper, we present a self-verification model trained on a newly constructed dataset that faithfully replicates the authentic format and evaluation scale of the exam. Our model is able to exceed the official passing score when evaluated on the actual exam scale, marking the first demonstration, to our knowledge, of an LLM passing the Japanese bar examination without altering its original question structure or scoring rules. We further conduct extensive comparisons with alternative strategies, including multi-agent inference and decomposition-based supervision, and find that these methods fail to achieve comparable performance. Our results highlight the importance of format-faithful supervision and consistency verification, and suggest that carefully designed single-model approaches can outperform more complex systems in high-stakes professional reasoning tasks. Our dataset and codes are publicly available.

Method: Systematic disentanglement experiments on distribution-based tasks using Cross-CoT experiments to isolate reasoning text effects from intrinsic model priors.

Result: Found a “decoupled mechanism”: CoT improves distributional alignment but final accuracy is 99% determined by CoT content, while distributional ranking is over 80% governed by model priors. CoT’s influence on accuracy grows monotonically during reasoning, but distributional structure is largely determined by LLM’s intrinsic priors.

Conclusion: Long CoT serves as a decisive decision-maker for top options but fails as a granular distribution calibrator for ambiguous tasks requiring probabilistic modeling.

Abstract: Reasoning-tuned LLMs utilizing long Chain-of-Thought (CoT) excel at single-answer tasks, yet their ability to model Human Label Variation–which requires capturing probabilistic ambiguity rather than resolving it–remains underexplored. We investigate this through systematic disentanglement experiments on distribution-based tasks, employing Cross-CoT experiments to isolate the effect of reasoning text from intrinsic model priors. We observe a distinct “decoupled mechanism”: while CoT improves distributional alignment, final accuracy is dictated by CoT content (99% variance contribution), whereas distributional ranking is governed by model priors (over 80%). Step-wise analysis further shows that while CoT’s influence on accuracy grows monotonically during the reasoning process, distributional structure is largely determined by LLM’s intrinsic priors. These findings suggest that long CoT serves as a decisive LLM decision-maker for the top option but fails to function as a granular distribution calibrator for ambiguous tasks.

Method: A two-stage RL framework: Stage 1 optimizes first-step planning using fine-grained rubrics from self-play experiences and human calibration; Stage 2 aligns execution with the initial plan through sparse rewards to ensure stable and efficient tool usage.

Result: Anchor-GRPO outperforms baseline GRPO and First-step GRPO across four benchmarks (BrowseComp, BrowseComp-Zh, GAIA, XBench-DeepSearch) for models from 3B to 30B parameters, improving task success and tool efficiency. WebAnchor-30B achieves 46.0% pass@1 on BrowseComp and 76.4% on GAIA, with strong scalability as model size and context length increase.

Conclusion: The proposed Anchor-GRPO framework effectively addresses the plan anchor problem in long-horizon web reasoning by decoupling planning and execution, leading to significant performance improvements and demonstrating strong scalability across different model sizes and contexts.

Abstract: Large Language Model(LLM)-based agents have shown strong capabilities in web information seeking, with reinforcement learning (RL) becoming a key optimization paradigm. However, planning remains a bottleneck, as existing methods struggle with long-horizon strategies. Our analysis reveals a critical phenomenon, plan anchor, where the first reasoning step disproportionately impacts downstream behavior in long-horizon web reasoning tasks. Current RL algorithms, fail to account for this by uniformly distributing rewards across the trajectory. To address this, we propose Anchor-GRPO, a two-stage RL framework that decouples planning and execution. In Stage 1, the agent optimizes its first-step planning using fine-grained rubrics derived from self-play experiences and human calibration. In Stage 2, execution is aligned with the initial plan through sparse rewards, ensuring stable and efficient tool usage. We evaluate Anchor-GRPO on four benchmarks: BrowseComp, BrowseComp-Zh, GAIA, and XBench-DeepSearch. Across models from 3B to 30B, Anchor-GRPO outperforms baseline GRPO and First-step GRPO, improving task success and tool efficiency. Notably, WebAnchor-30B achieves 46.0% pass@1 on BrowseComp and 76.4% on GAIA. Anchor-GRPO also demonstrates strong scalability, getting higher accuracy as model size and context length increase.

Method: Systematically evaluated five embedding similarity metrics across 816 transfer experiments spanning three NLP tasks, three African-centric multilingual models, and 12 languages from four language families.

Result: Cosine gap and retrieval-based metrics (P@1, CSLS) reliably predict transfer success (ρ=0.4-0.6), while CKA shows negligible predictive power (ρ≈0.1). Correlation signs reverse when pooling across models (Simpson’s Paradox), requiring per-model validation.

Conclusion: Embedding metrics achieve comparable predictive power to linguistic typology, providing concrete guidance for source language selection and highlighting the importance of model-specific analysis due to Simpson’s Paradox.

Abstract: Cross-lingual transfer is essential for building NLP systems for low-resource African languages, but practitioners lack reliable methods for selecting source languages. We systematically evaluate five embedding similarity metrics across 816 transfer experiments spanning three NLP tasks, three African-centric multilingual models, and 12 languages from four language families. We find that cosine gap and retrieval-based metrics (P@1, CSLS) reliably predict transfer success ($ρ= 0.4-0.6$), while CKA shows negligible predictive power ($ρ\approx 0.1$). Critically, correlation signs reverse when pooling across models (Simpson’s Paradox), so practitioners must validate per-model. Embedding metrics achieve comparable predictive power to URIEL linguistic typology. Our results provide concrete guidance for source language selection and highlight the importance of model-specific analysis.

Method: Dynamic In-Context Planner (DIP) leverages the diffusion generation paradigm to efficiently adjust context dynamically during generation, selecting and inserting in-context examples on-the-fly rather than providing all examples upfront.

Result: DIP maintains generation quality while achieving up to 12.9× inference speedup over standard inference and 1.17× over KV cache-enhanced inference.

Conclusion: The diffusion generation paradigm enables efficient dynamic context adjustment, and DIP effectively addresses computational bottlenecks in DLMs while preserving performance.

Abstract: Diffusion language models (DLMs) have shown strong potential for general natural language tasks with in-context examples. However, due to the bidirectional attention mechanism, DLMs incur substantial computational cost as context length increases. This work addresses this issue with a key discovery: unlike the sequential generation in autoregressive language models (ARLMs), the diffusion generation paradigm in DLMs allows \textit{efficient dynamic adjustment of the context} during generation. Building on this insight, we propose \textbf{D}ynamic \textbf{I}n-Context \textbf{P}lanner (DIP), a context-optimization method that dynamically selects and inserts in-context examples during generation, rather than providing all examples in the prompt upfront. Results show DIP maintains generation quality while achieving up to 12.9$\times$ inference speedup over standard inference and 1.17$\times$ over KV cache-enhanced inference.

Method: PALU uses prefix-aware localized unlearning with local entropy maximization across temporal and vocabulary dimensions. It suppresses only sensitive prefixes to sever causal generation links and flattens only top-k logits to maximize uncertainty in critical subspaces.

Result: Extensive experiments show PALU achieves superior forgetting efficacy and utility preservation compared to state-of-the-art baselines, avoiding redundant optimization across full vocabulary and parameter space.

Conclusion: Localized unlearning focusing on sensitive prefixes and top-k logits is sufficient for effective machine unlearning while minimizing collateral damage to general model performance.

Abstract: Machine unlearning aims to forget sensitive knowledge from Large Language Models (LLMs) while maintaining general utility. However, existing approaches typically treat all tokens in a response indiscriminately and enforce uncertainty over the entire vocabulary. This global treatment results in unnecessary utility degradation and extends optimization to content-agnostic regions. To address these limitations, we propose PALU (Prefix-Aware Localized Unlearning), a framework driven by a local entropy maximization objective across both temporal and vocabulary dimensions. PALU reveals that (i) suppressing the sensitive prefix alone is sufficient to sever the causal generation link, and (ii) flattening only the top-$k$ logits is adequate to maximize uncertainty in the critical subspace. These findings allow PALU to avoid redundant optimization across the full vocabulary and parameter space while minimizing collateral damage to general model performance. Extensive experiments validate that PALU achieves superior forgetting efficacy and utility preservation compared to state-of-the-art baselines.

Method: MemRL separates frozen LLM reasoning from plastic episodic memory. Uses Two-Phase Retrieval: first semantic filtering, then Q-value (utility) selection. Q-values are refined via environmental feedback through trial-and-error reinforcement learning.

Result: Significantly outperforms SOTA baselines on HLE, BigCodeBench, ALFWorld, and Lifelong Agent Bench. Effectively reconciles stability-plasticity dilemma, enabling continuous runtime improvement without weight updates.

Conclusion: MemRL provides a framework for LLM agents to achieve human-like self-evolution through non-parametric reinforcement learning on episodic memory, enabling continuous improvement while avoiding catastrophic forgetting.

Abstract: The hallmark of human intelligence is the ability to master new skills through Constructive Episodic Simulation-retrieving past experiences to synthesize solutions for novel tasks. While Large Language Models possess strong reasoning capabilities, they struggle to emulate this self-evolution: fine-tuning is computationally expensive and prone to catastrophic forgetting, while existing memory-based methods rely on passive semantic matching that often retrieves noise. To address these challenges, we propose MemRL, a framework that enables agents to self-evolve via non-parametric reinforcement learning on episodic memory. MemRL explicitly separates the stable reasoning of a frozen LLM from the plastic, evolving memory. Unlike traditional methods, MemRL employs a Two-Phase Retrieval mechanism that filters candidates by semantic relevance and then selects them based on learned Q-values (utility). These utilities are continuously refined via environmental feedback in an trial-and-error manner, allowing the agent to distinguish high-value strategies from similar noise. Extensive experiments on HLE, BigCodeBench, ALFWorld, and Lifelong Agent Bench demonstrate that MemRL significantly outperforms state-of-the-art baselines. Our analysis experiments confirm that MemRL effectively reconciles the stability-plasticity dilemma, enabling continuous runtime improvement without weight updates.

Method: 1) UltraLogic framework decouples logical core from natural language using Code-based Solving methodology. 2) Includes hundreds of unique task types with automated calibration across 10 difficulty levels. 3) Introduces Bipolar Float Reward (BFR) mechanism with graded penalties to distinguish perfect responses from flawed ones.

Result: Task diversity is the primary driver for reasoning enhancement. BFR combined with difficulty matching strategy significantly improves training efficiency and guides models toward global logical optima.

Conclusion: UltraLogic addresses critical bottlenecks in LLM reasoning by automating high-quality data production, providing calibrated difficulty levels, and solving reward sparsity through innovative BFR mechanism, enabling more effective training for complex reasoning tasks.

Abstract: While Large Language Models (LLMs) have demonstrated significant potential in natural language processing , complex general-purpose reasoning requiring multi-step logic, planning, and verification remains a critical bottleneck. Although Reinforcement Learning with Verifiable Rewards (RLVR) has succeeded in specific domains , the field lacks large-scale, high-quality, and difficulty-calibrated data for general reasoning. To address this, we propose UltraLogic, a framework that decouples the logical core of a problem from its natural language expression through a Code-based Solving methodology to automate high-quality data production. The framework comprises hundreds of unique task types and an automated calibration pipeline across ten difficulty levels. Furthermore, to mitigate binary reward sparsity and the Non-negative Reward Trap, we introduce the Bipolar Float Reward (BFR) mechanism, utilizing graded penalties to effectively distinguish perfect responses from those with logical flaws. Our experiments demonstrate that task diversity is the primary driver for reasoning enhancement , and that BFR, combined with a difficulty matching strategy, significantly improves training efficiency, guiding models toward global logical optima.

Method: Proposes X-MuTeST framework combining LLM semantic reasoning with attention techniques. Uses human-annotated token-level rationales for training. Computes explanations by comparing prediction probabilities of original text vs. n-grams (unigrams, bigrams, trigrams). Final explanations are union of LLM and X-MuTeST explanations.

Result: Leveraging human rationales improves both classification performance and explainability. Combining human rationales with explainability method to refine model attention yields further improvements. Dataset includes 6,004 Hindi, 4,492 Telugu, and 6,334 English samples with token-level rationale annotations.

Conclusion: The work advances hate speech detection across diverse linguistic contexts by focusing on under-resourced languages. The explainability-guided framework enhances both accuracy and interpretability, with publicly available data and code.

Abstract: Hate speech detection on social media faces challenges in both accuracy and explainability, especially for underexplored Indic languages. We propose a novel explainability-guided training framework, X-MuTeST (eXplainable Multilingual haTe Speech deTection), for hate speech detection that combines high-level semantic reasoning from large language models (LLMs) with traditional attention-enhancing techniques. We extend this research to Hindi and Telugu alongside English by providing benchmark human-annotated rationales for each word to justify the assigned class label. The X-MuTeST explainability method computes the difference between the prediction probabilities of the original text and those of unigrams, bigrams, and trigrams. Final explanations are computed as the union between LLM explanations and X-MuTeST explanations. We show that leveraging human rationales during training enhances both classification performance and explainability. Moreover, combining human rationales with our explainability method to refine the model attention yields further improvements. We evaluate explainability using Plausibility metrics such as Token-F1 and IOU-F1 and Faithfulness metrics such as Comprehensiveness and Sufficiency. By focusing on under-resourced languages, our work advances hate speech detection across diverse linguistic contexts. Our dataset includes token-level rationale annotations for 6,004 Hindi, 4,492 Telugu, and 6,334 English samples. Data and code are available on https://github.com/ziarehman30/X-MuTeST

Method: Developed MalruleLib framework that translates 67 learning-science sources into 101 executable malrules across 498 parameterized problem templates. Creates paired dual-path traces showing both correct reasoning and malrule-consistent student reasoning. Formalizes Malrule Reasoning Accuracy (MRA) as a core student-modeling problem.

Result: Across nine language models (4B-120B), accuracy drops from 66% on direct problem solving to 40% on cross-template misconception prediction. MalruleLib can generate over one million instances, enabling scalable supervision. Cross-template degradations of 10-21% observed, while providing student step traces improves prediction by 3-15%.

Conclusion: MalruleLib provides infrastructure for educational AI to model student procedures across contexts, enabling diagnosis and feedback that targets underlying misconceptions. Released as a tool for scalable evaluation and supervision in educational AI applications.

Abstract: Student mistakes in mathematics are often systematic: a learner applies a coherent but wrong procedure and repeats it across contexts. We introduce MalruleLib, a learning-science-grounded framework that translates documented misconceptions into executable procedures, drawing on 67 learning-science and mathematics education sources, and generates step-by-step traces of malrule-consistent student work. We formalize a core student-modeling problem as Malrule Reasoning Accuracy (MRA): infer a misconception from one worked mistake and predict the student’s next answer under cross-template rephrasing. Across nine language models (4B-120B), accuracy drops from 66% on direct problem solving to 40% on cross-template misconception prediction. MalruleLib encodes 101 malrules over 498 parameterized problem templates and produces paired dual-path traces for both correct reasoning and malrule-consistent student reasoning. Because malrules are executable and templates are parameterizable, MalruleLib can generate over one million instances, enabling scalable supervision and controlled evaluation. Using MalruleLib, we observe cross-template degradations of 10-21%, while providing student step traces improves prediction by 3-15%. We release MalruleLib as infrastructure for educational AI that models student procedures across contexts, enabling diagnosis and feedback that targets the underlying misconception.

Method: Created RXL-RADSet containing 1,600 synthetic radiology reports across 10 RADS frameworks using LLMs with scenario plans and simulated radiologist styles, followed by two-stage radiologist verification. Evaluated 41 quantized SLMs (0.135-32B parameters) and GPT-5.2 using fixed guided prompting, with primary endpoints of validity and accuracy.

Result: GPT-5.2 achieved 99.8% validity and 81.1% accuracy. SLMs achieved 96.8% validity and 61.1% accuracy overall, with top 20-32B models reaching ~99% validity and mid-to-high 70% accuracy. Performance scaled with model size and declined with RADS complexity. Guided prompting improved both validity and accuracy compared to zero-shot.

Conclusion: RXL-RADSet provides a valuable multi-RADS benchmark. Large SLMs (20-32B) can approach proprietary model performance for RADS assignment under guided prompting, but performance gaps remain for higher-complexity classification schemes.

Abstract: Background: Reporting and Data Systems (RADS) standardize radiology risk communication but automated RADS assignment from narrative reports is challenging because of guideline complexity, output-format constraints, and limited benchmarking across RADS frameworks and model sizes. Purpose: To create RXL-RADSet, a radiologist-verified synthetic multi-RADS benchmark, and compare validity and accuracy of open-weight small language models (SLMs) with a proprietary model for RADS assignment. Materials and Methods: RXL-RADSet contains 1,600 synthetic radiology reports across 10 RADS (BI-RADS, CAD-RADS, GB-RADS, LI-RADS, Lung-RADS, NI-RADS, O-RADS, PI-RADS, TI-RADS, VI-RADS) and multiple modalities. Reports were generated by LLMs using scenario plans and simulated radiologist styles and underwent two-stage radiologist verification. We evaluated 41 quantized SLMs (12 families, 0.135-32B parameters) and GPT-5.2 under a fixed guided prompt. Primary endpoints were validity and accuracy; a secondary analysis compared guided versus zero-shot prompting. Results: Under guided prompting GPT-5.2 achieved 99.8% validity and 81.1% accuracy (1,600 predictions). Pooled SLMs (65,600 predictions) achieved 96.8% validity and 61.1% accuracy; top SLMs in the 20-32B range reached ~99% validity and mid-to-high 70% accuracy. Performance scaled with model size (inflection between <1B and >=10B) and declined with RADS complexity primarily due to classification difficulty rather than invalid outputs. Guided prompting improved validity (99.2% vs 96.7%) and accuracy (78.5% vs 69.6%) compared with zero-shot. Conclusion: RXL-RADSet provides a radiologist-verified multi-RADS benchmark; large SLMs (20-32B) can approach proprietary-model performance under guided prompting, but gaps remain for higher-complexity schemes.

Method: Introduces ST-Bench benchmark with four core tasks via network SDE-based multi-agent data synthesis, then proposes STReasoner that integrates time series, graph structure, and text for explicit reasoning. Also introduces S-GRPO reinforcement learning algorithm that rewards performance gains specifically from spatial information.

Result: STReasoner achieves average accuracy gains between 17% and 135% at only 0.004X the cost of proprietary models and generalizes robustly to real-world data.

Conclusion: The proposed approach successfully addresses the gap in spatio-temporal reasoning by combining benchmark development, explicit reasoning models, and spatially-grounded reinforcement learning, enabling more interpretable and effective decision-making in complex systems.

Abstract: Spatio-temporal reasoning in time series involves the explicit synthesis of temporal dynamics, spatial dependencies, and textual context. This capability is vital for high-stakes decision-making in systems such as traffic networks, power grids, and disease propagation. However, the field remains underdeveloped because most existing works prioritize predictive accuracy over reasoning. To address the gap, we introduce ST-Bench, a benchmark consisting of four core tasks, including etiological reasoning, entity identification, correlation reasoning, and in-context forecasting, developed via a network SDE-based multi-agent data synthesis pipeline. We then propose STReasoner, which empowers LLM to integrate time series, graph structure, and text for explicit reasoning. To promote spatially grounded logic, we introduce S-GRPO, a reinforcement learning algorithm that rewards performance gains specifically attributable to spatial information. Experiments show that STReasoner achieves average accuracy gains between 17% and 135% at only 0.004X the cost of proprietary models and generalizes robustly to real-world data.

Method: Proposes Automated Semantic Rules Detection (ASRD) algorithm that extracts patterns from messages exchanged by agents trained on two different datasets in the Lewis Game. The algorithm relates extracted patterns to specific attributes of the input data.

Result: ASRD successfully helps interpret emergent communication by identifying semantic patterns and connecting them to data attributes, significantly simplifying subsequent analysis of agent communication.

Conclusion: The ASRD algorithm provides a valuable tool for improving interpretability in emergent communication research, bridging the gap between autonomous language development and human understanding of what agents are communicating.

Abstract: The field of emergent communication within multi-agent systems examines how autonomous agents can independently develop communication strategies, without explicit programming, and adapt them to varied environments. However, few studies have focused on the interpretability of emergent languages. The research exposed in this paper proposes an Automated Semantic Rules Detection (ASRD) algorithm, which extracts relevant patterns in messages exchanged by agents trained with two different datasets on the Lewis Game, which is often studied in the context of emergent communication. ASRD helps at the interpretation of the emergent communication by relating the extracted patterns to specific attributes of the input data, thereby considerably simplifying subsequent analysis.

Method: Collected 10M tweets for “tweet-tuning” base models and 6.25M posts from over 5,000 X users and their peers. Used negative log-likelihood as predictability measure. Tested four large language models (1.5B to 70B parameters) with different context strategies (user history vs. social circle posts) and both prompting and fine-tuning approaches.

Result: Predicting individual users’ posts remains surprisingly difficult. Models using users’ own history significantly outperform those using social circle context. Results consistent across model sizes and both prompting/fine-tuning approaches. Up to 20% of in-context learning involves @-mentions and hashtags usage. Main findings hold across demographic groups.

Conclusion: Personal history is crucial for predicting individual social media behavior, while social circle context provides limited predictive value. The difficulty in predicting individual posts suggests unique personal expression patterns that models struggle to capture, even with large-scale data and advanced architectures.

Abstract: Our paper studies the predictability of online speech – that is, how well language models learn to model the distribution of user generated content on X (previously Twitter). We define predictability as a measure of the model’s uncertainty, i.e. its negative log-likelihood. As the basis of our study, we collect 10M tweets for ``tweet-tuning’’ base models and a further 6.25M posts from more than five thousand X (previously Twitter) users and their peers. In our study involving more than 5000 subjects, we find that predicting posts of individual users remains surprisingly hard. Moreover, it matters greatly what context is used: models using the users’ own history significantly outperform models using posts from their social circle. We validate these results across four large language models ranging in size from 1.5 billion to 70 billion parameters. Moreover, our results replicate if instead of prompting the model with additional context, we finetune on it. We follow up with a detailed investigation on what is learned in-context and a demographic analysis. Up to 20% of what is learned in-context is the use of @-mentions and hashtags. Our main results hold across the demographic groups we studied.

Method: The researchers tested various LLMs on thematic fit tasks, comparing closed and open models, examining the effects of multi-step reasoning, generated sentences, and output form on performance.

Result: Closed models set new SOTA for thematic fit knowledge, while open models also capture relevant knowledge but score lower. Multi-step reasoning only helped closed models, generated sentences hurt closed models’ performance, and output form had minimal effect.

Conclusion: More foundational work is needed for a single LLM to perform best on all tasks under the same conditions, indicating current limitations in achieving consistent, optimal performance across different semantic tasks.

Abstract: We show closed models possess much thematic fit knowledge and set a new state of the art, while open models also seem to capture much relevant knowledge (in semantic filtering), but yield lower scores. Surprisingly, multi-step reasoning only helped closed models (with few exceptions); generated sentences hurt closed models’ performance; and output form had little to no effect. We analyze the reasons for these findings, and conclude that more foundational work is needed for a single LLM to perform the best on all tasks with the same experimental condition, let alone improve results further. Source code is available at: https://github.com/SafeyahShemali/LLM_Thematic_Fit_25

Method: Two-stage pipeline: 1) Infer authors’ implicit writing characteristics and organize them into Author Writing Sheets (validated as high-quality by humans), 2) Simulate author persona using tailored persona descriptions and personalized story rules. Dataset: Mythos with 3.6k stories from 112 authors across five diverse sources.

Result: Personalized stories using Author Writing Sheets outperform non-personalized baseline with 78% win-rate in capturing authors’ past style and 59% similarity to ground-truth author stories. Human evaluation confirms findings and shows Reddit stories are easier to personalize, and Creativity/Language Use aspects are easier to personalize than Plot.

Conclusion: Personalized story generation using structured author writing characteristics is effective for mimicking author styles, with significant improvements over non-personalized approaches across diverse story-writing settings.

Abstract: Personalization is critical for improving user experience in interactive writing and educational applications, yet remains understudied in story generation. We study the task of personalizing story generation, where our goal is to mimic an author’s writing style, given other stories written by them. We collect Mythos, a dataset of 3.6k stories from 112 authors, with an average of 16 stories per author, across five distinct sources reflecting diverse story-writing settings. We propose a two-stage pipeline for personalized story generation: first, we infer authors’ implicit writing characteristics and organize them into an Author Writing Sheet, which is validated by humans to be of high quality; second, we simulate the author’s persona using tailored persona descriptions and personalized story rules. We find that stories personalized using the Author Writing Sheet outperform a non-personalized baseline, achieving a 78% win-rate in capturing authors’ past style and 59% in similarity to ground-truth author stories. Human evaluation supports these findings and further highlights trends, such as Reddit stories being easier to personalize, and the Creativity and Language Use aspects of stories being easier to personalize than the Plot.

Method: Proposes ReFreeKV as the first threshold-free KV pruning method that automatically adjusts budget sizes while ensuring full-cache performance. The method lifts threshold constraints for robust KV pruning.

Result: Validated by intensive experiments on 13 datasets of diverse context lengths, task types, and model sizes. The method achieves robust performance without requiring input-specific threshold tuning.

Conclusion: ReFreeKV addresses the inherent limitation of threshold-dependent KV pruning methods, enabling more practical and robust memory reduction for LLM inference across diverse real-world scenarios.

Abstract: To reduce memory consumption during LLM inference, prior works have proposed numerous methods that focus on KV cache pruning based on various criteria. While these techniques often accomplish lossless memory reduction on many datasets, they often rely on an under-emphasized condition: a dataset/domain-specific budget size threshold needs to be pre-determined to achieve the optimal performance. However, such input-specific tuning may be considerably limited in real-world scenarios, as open-domain inputs span diverse domains, lengths and difficulty levels, without clear boundaries for pre-tuning. Thus, the dependence of an input-sensitive threshold can be an inherent limitation that may cause large degradation on arbitrary inputs. In this work, we propose a new objective that lifts the threshold constraints for robust KV pruning, calling for “threshold-free” methods that automatically adjust budget sizes while ensuring full-cache performance. We then propose a novel method ReFreeKV as the first solution fulfilling this objective, validated by intensive experiments on 13 datasets of diverse context lengths, task types, and model sizes.

Method: The study compares six inference-time methods to improve QA performance on misleading visualizations without compromising accuracy on non-misleading ones. Methods include table-based QA and redrawing the visualization, among others.

Result: MLLM QA accuracy on misleading visualizations drops to random baseline levels. Two methods proved effective: table-based QA and redrawing the visualization, with improvements of up to 19.6 percentage points.

Conclusion: MLLMs have critical vulnerability to misleading visualizations, but practical inference-time solutions exist. Table-based QA and visualization redrawing are promising approaches to enhance robustness while maintaining performance on legitimate charts.

Abstract: Visualizations play a pivotal role in daily communication in an increasingly data-driven world. Research on multimodal large language models (MLLMs) for automated chart understanding has accelerated massively, with steady improvements on standard benchmarks. However, for MLLMs to be reliable, they must be robust to misleading visualizations, i.e., charts that distort the underlying data, leading readers to draw inaccurate conclusions. Here, we uncover an important vulnerability: MLLM question-answering (QA) accuracy on misleading visualizations drops on average to the level of the random baseline. To address this, we provide the first comparison of six inference-time methods to improve QA performance on misleading visualizations, without compromising accuracy on non-misleading ones. We find that two methods, table-based QA and redrawing the visualization, are effective, with improvements of up to 19.6 percentage points. We make our code and data available.

Method: SR-RAG casts selective retrieval as knowledge source selection, treating the LLM itself as a first-class knowledge source. It learns to select appropriate sources, optionally verbalize parametric knowledge, and answer using selected sources within a single left-to-right generation pass. It also combines LLM-based uncertainty with an external policy datastore to improve decision calibration.

Result: Across four benchmarks and three 7B-class LLMs, SR-RAG outperforms a strong selective retrieval baseline by 8.5%/2.1%/4.7% while performing 26%/40%/21% fewer retrievals, achieving favorable accuracy-latency trade-offs without dataset-specific threshold tuning.

Conclusion: SR-RAG demonstrates that treating LLMs as first-class knowledge sources and learning multi-source selection within a single generation pass significantly improves selective retrieval efficiency and reliability compared to binary approaches.

Abstract: Selective retrieval aims to make retrieval-augmented generation (RAG) more efficient and reliable by skipping retrieval when an LLM’s parametric knowledge suffices. Despite promising results, existing methods are constrained by a binary design choice: either retrieve from a single external source or skip retrieval and let the LLM directly produce the final answer. We argue that this fallback underestimates the model’s knowledge and obscures the more general multi-source decision problem that arises in practical systems. We propose Self-Routing RAG (SR-RAG), which casts selective retrieval as knowledge source selection and treats the LLM itself as a first-class knowledge source. SR-RAG learns to select an appropriate knowledge source, optionally verbalize parametric knowledge, and answer using the selected source, all within a single left-to-right generation pass. SR-RAG further augments source selection by combining LLM-based uncertainty with a flexible external policy datastore to improve decision calibration. Across four benchmarks and three 7B-class LLMs, SR-RAG outperforms a strong selective retrieval baseline by 8.5%/2.1%/4.7% while performing 26%/40%/21% fewer retrievals, and it achieves favorable accuracy-latency trade-offs without dataset-specific threshold tuning.

Method: 1) Modality-aware routing that dynamically identifies the most appropriate modality-specific corpus and performs targeted retrieval within it. 2) Organizes each modality into multiple granularity levels for fine-tuned retrieval tailored to query complexity and scope. 3) Theoretical analysis justifies the effectiveness of modality-aware routing.

Result: Validated on 10 benchmarks of multiple modalities, showing superiority over various modality-specific and unified baselines.

Conclusion: UniversalRAG effectively addresses the limitations of single-modality RAG systems by enabling retrieval from heterogeneous sources with diverse modalities and granularities, overcoming the modality gap problem through intelligent routing and multi-level organization.

Abstract: Retrieval-Augmented Generation (RAG) has shown substantial promise in improving factual accuracy by grounding model responses with external knowledge relevant to queries. However, most existing approaches are limited to a text-only corpus, and while recent efforts have extended RAG to other modalities such as images and videos, they typically operate over a single modality-specific corpus. In contrast, real-world queries vary widely in the type of knowledge they require, which a single type of knowledge source cannot address. To address this, we introduce UniversalRAG, designed to retrieve and integrate knowledge from heterogeneous sources with diverse modalities and granularities. Specifically, motivated by the observation that forcing all modalities into a unified representation space derived from a single aggregated corpus causes a modality gap, where the retrieval tends to favor items from the same modality as the query, we propose modality-aware routing, which dynamically identifies the most appropriate modality-specific corpus and performs targeted retrieval within it, and further justify its effectiveness with a theoretical analysis. Moreover, beyond modality, we organize each modality into multiple granularity levels, enabling fine-tuned retrieval tailored to the complexity and scope of the query. We validate UniversalRAG on 10 benchmarks of multiple modalities, showing its superiority over various modality-specific and unified baselines.

Method: 1) Robustness metric: Calculates correlation between semantic similarity and taxonomic similarity to detect errors not captured by existing metrics. 2) Logical adequacy metric: Uses Natural Language Inference to assess whether taxonomic relationships are logically sound.

Result: Both metrics correlate well with F1 scores against ground truth taxonomies when tested on five taxonomies. The metrics can also predict downstream performance in hierarchical classification tasks when used with label hierarchies.

Conclusion: The proposed reference-free metrics provide effective ways to evaluate taxonomy quality without requiring ground truth labels, offering practical tools for assessing both robustness and logical adequacy of taxonomies.

Abstract: We introduce two reference-free metrics for quality evaluation of taxonomies in the absence of labels. The first metric evaluates robustness by calculating the correlation between semantic and taxonomic similarity, addressing error types not considered by existing metrics. The second uses Natural Language Inference to assess logical adequacy. Both metrics are tested on five taxonomies and are shown to correlate well with F1 against ground truth taxonomies. We further demonstrate that our metrics can predict downstream performance in hierarchical classification when used with label hierarchies.

Method: Created EduBench with synthetic data covering 9 major educational scenarios and 4,000+ contexts, developed 12 multi-dimensional evaluation metrics, used human annotation for validation, and trained a small-scale model on the dataset.

Result: The trained small-scale model achieved performance comparable to state-of-the-art large models (Deepseek V3, Qwen Max) on the test set, demonstrating the benchmark’s effectiveness.

Conclusion: This work provides a practical foundation for developing and evaluating education-oriented language models, with code and data publicly released.

Abstract: As large language models continue to advance, their application in educational contexts remains underexplored and under-optimized. In this paper, we address this gap by introducing the first diverse benchmark tailored for educational scenarios, incorporating synthetic data containing 9 major scenarios and over 4,000 distinct educational contexts. To enable comprehensive assessment, we propose a set of multi-dimensional evaluation metrics that cover 12 critical aspects relevant to both teachers and students. We further apply human annotation to ensure the effectiveness of the model-generated evaluation responses. Additionally, we succeed to train a relatively small-scale model on our constructed dataset and demonstrate that it can achieve performance comparable to state-of-the-art large models (e.g., Deepseek V3, Qwen Max) on the test set. Overall, this work provides a practical foundation for the development and evaluation of education-oriented language models. Code and data are released at https://github.com/ybai-nlp/EduBench.

Method: Created POLAR dataset with 23k+ instances in 7 languages from diverse online platforms and real-world events, annotated along three axes (presence, type, manifestation) using culturally-adapted annotation platforms. Conducted experiments with multilingual PLMs in monolingual/cross-lingual setups and evaluated LLMs in few-shot/zero-shot scenarios.

Result: Most models perform well on binary polarization detection but achieve substantially lower scores when predicting polarization types and manifestations, highlighting the complex, contextual nature of polarization.

Conclusion: Polarization is highly contextual and complex, requiring robust, adaptable approaches in NLP and computational social science. The POLAR dataset will be released to support global research on digital polarization mitigation.

Abstract: Online polarization poses a growing challenge for democratic discourse, yet most computational social science research remains monolingual, culturally narrow, or event-specific. We introduce POLAR, a multilingual, multicultural, and multievent dataset with over 23k instances in seven languages from diverse online platforms and real-world events. Polarization is annotated along three axes: presence, type, and manifestation, using a variety of annotation platforms adapted to each cultural context. We conduct two main experiments: (1) we fine-tune six multilingual pretrained language models in both monolingual and cross-lingual setups; and (2) we evaluate a range of open and closed large language models (LLMs) in few-shot and zero-shot scenarios. Results show that while most models perform well on binary polarization detection, they achieve substantially lower scores when predicting polarization types and manifestations. These findings highlight the complex, highly contextual nature of polarization and the need for robust, adaptable approaches in NLP and computational social science. All resources will be released to support further research and effective mitigation of digital polarization globally.

Method: Manual and large language model annotation to classify unstructured Craigslist listings by neighborhood, followed by geospatial analysis and topic modeling to identify spatial patterns and language correlations.

Result: Identified three distinct patterns: 1) conflicting neighborhood designations due to competing spatial definitions, 2) border properties with valid claims to adjacent neighborhoods, and 3) “reputation laundering” where listings claim association with distant desirable neighborhoods. Topic modeling showed listings further from neighborhood centers emphasize different amenities than centrally-located units.

Conclusion: Natural language processing reveals how definitions of urban spaces are contested in ways traditional methods overlook, showing how rental listings socially construct neighborhoods through language that often conflicts with institutional boundaries.

Abstract: Rental listings offer a window into how urban space is socially constructed through language. We analyze Chicago Craigslist rental advertisements from 2018 to 2024 to examine how listing agents characterize neighborhoods, identifying mismatches between institutional boundaries and neighborhood claims. Through manual and large language model annotation, we classify unstructured listings from Craigslist according to their neighborhood. Further geospatial analysis reveals three distinct patterns: properties with conflicting neighborhood designations due to competing spatial definitions, border properties with valid claims to adjacent neighborhoods, and “reputation laundering” where listings claim association with distant, desirable neighborhoods. Through topic modeling, we identify patterns that correlate with spatial positioning: listings further from neighborhood centers emphasize different amenities than centrally-located units. Natural language processing techniques reveal how definitions of urban spaces are contested in ways that traditional methods overlook.

Method: Created TRuST dataset through a carefully synthesized definition of toxicity and multi-stage human annotation process. Evaluated annotator diversity. Benchmarked state-of-the-art LLMs and pre-trained models on three tasks: toxicity detection, target group identification, and toxic word identification.

Result: Fine-tuned PLMs outperform LLMs on all three toxicity tasks. Current reasoning models do not reliably improve performance. TRuST provides one of the most comprehensive resources for evaluating LLM toxicity.

Conclusion: TRuST addresses the inconsistency in toxicity definitions and serves as a valuable resource for developing socially-aware and safer language technologies, with findings showing that fine-tuned PLMs currently outperform LLMs on toxicity-related tasks.

Abstract: Toxic language includes content that is offensive, abusive, or that promotes harm. Progress in preventing toxic output from large language models (LLMs) is hampered by inconsistent definitions of toxicity. We introduce TRuST, a large-scale dataset that unifies and expands prior resources through a carefully synthesized definition of toxicity, and corresponding annotation scheme. It consists of ~300k annotations, with high-quality human annotation on ~11k. To ensure high-quality, we designed a rigorous, multi-stage human annotation process, and evaluated the diversity of the annotators. Then we benchmarked state-of-the-art LLMs and pre-trained models on three tasks: toxicity detection, identification of the target group, and of toxic words. Our results indicate that fine-tuned PLMs outperform LLMs on the three tasks, and that current reasoning models do not reliably improve performance. TRuST constitutes one of the most comprehensive resources for evaluating and mitigating LLM toxicity, and other research in socially-aware and safer language technologies.

Method: Based on the FM-index data structure that simultaneously indexes and compresses text, creating indexes with size only 44% of the corpus. The system improves indexing speed (18×) and reduces memory use during indexing (3.2× reduction) and querying.

Result: Indexed 83TB of Internet text in 99 days with a single CPU node (or 19 hours with 137 nodes). Found several core LM evaluation benchmarks heavily contaminated in Internet crawls (up to 74.2% in GSM8K). Created a benchmark contamination bulletin and released web interface/API for search queries.

Conclusion: infini-gram mini enables efficient large-scale text corpus analysis, revealing significant benchmark contamination that could lead to overestimating language model capabilities if trained on contaminated data.

Abstract: Language models are trained mainly on massive text data from the Internet, and it becomes increasingly important to understand this data source. Exact-match search engines enable searching in large text corpora - counting string appearances and retrieving the enclosing documents - yet the high storage overhead hinders their application on Internet-scale data. We present infini-gram mini, an efficient and scalable system that can make petabyte-level text corpora searchable. Based on the FM-index data structure (Ferragina and Manzini, 2000), which simultaneously indexes and compresses text, our system creates indexes with size only 44% of the corpus. Infini-gram mini greatly improves upon the best existing implementation of FM-index in terms of indexing speed (18$\times$) and memory use during both indexing (3.2$\times$ reduction) and querying (down to a negligible amount). We index 83TB of Internet text in 99 days with a single CPU node with 128 vCPUs (or 19 hours if using 137 such nodes). We show one important use case of infini-gram mini in a large-scale analysis of benchmark contamination. We find several core LM evaluation benchmarks to be heavily contaminated in Internet crawls (up to 74.2% in GSM8K), which could lead to overestimating the capabilities of language models if trained on such data. We host a benchmark contamination bulletin to share the contamination rate of many core and community-contributed benchmarks. We also release a web interface and an API endpoint to serve general search queries on infini-gram mini indexes.

Method: Constructed MemeMind dataset aligned with international standards and internet context, providing detailed Chain-of-Thought reasoning annotations. Proposed MemeGuard, a reasoning-oriented multimodal detection model for harmful meme detection.

Result: MemeGuard outperforms existing state-of-the-art methods on the MemeMind dataset, improving both detection accuracy and interpretability of model decisions.

Conclusion: The MemeMind dataset and MemeGuard model establish a solid foundation for future research in harmful meme detection by addressing the challenges of implicit content and improving both accuracy and interpretability.

Abstract: As a multimodal medium combining images and text, memes frequently convey implicit harmful content through metaphors and humor, rendering the detection of harmful memes a complex and challenging task. Although recent studies have made progress in detection accuracy and interpretability, large-scale, high-quality datasets for harmful memes remain scarce, and current methods still struggle to capture implicit risks and nuanced semantics. Thus, we construct MemeMind, a large-scale harmful meme dataset. Aligned with the international standards and the context of internet, MemeMind provides detailed Chain-of-Thought (CoT) reasoning annotations to support fine-grained analysis of implicit intentions in memes. Based on this dataset, we further propose MemeGuard, a reasoning-oriented multimodal detection model that significantly improves both the accuracy of harmful meme detection and the interpretability of model decisions. Extensive experimental results demonstrate that MemeGuard outperforms existing state-of-the-art methods on the MemeMind dataset, establishing a solid foundation for future research in harmful meme detection.

Method: The WakenLLM pipeline: (1) extracts Vague Perception samples where LLMs incorrectly abstain, and (2) measures how many can be converted to correct answers under stimulation. Uses stage-wise metrics (TCR, OCR, etc.) and upper-bound accuracy Acc(WakenLLM) to quantify reasoning potential beyond one-shot accuracy.

Result: Experiments on six LLMs show that without further training or parameter revisions, LLMs can achieve up to 68.53% accuracy increase on Vague Perception samples through the pipeline. Analysis reveals how Vague Perception, Conformity and Degradation vary across model families and parameter sizes. Comparison with mainstream reasoning baselines shows existing methods only activate a small portion of LLMs’ reasoning potential.

Conclusion: The paper identifies Vague Perception as a key limitation in LLM abstention behavior and demonstrates that significant reasoning potential remains untapped. WakenLLM provides tools to quantify and activate this potential, pointing to perception-aware reasoning as a promising direction for future LLM design.

Abstract: Large language models (LLMs) are increasingly trained to abstain on difficult questions by answering unknown. However, we observe that LLMs often misuse this option: they output unknown even when LLMs can actually solve the questions, or they fail to understand why questions are truly unsolvable. We formalize this mismatch between potential ability and the inclination of abstention as the Vague Perception phenomenon. We introduce the WakenLLM pipeline that (1) extracts Vague Perception samples and (2) measures how many of them can be converted to correct answers under stimulation. Based on stage-wise metrics (TCR, OCR, etc.) and the upper-bound accuracy Acc(WakenLLM), we quantify LLMs’ reasoning potential beyond one-shot accuracy. Experiments on six LLMs suggest that, without further training or parameter revisions, LLMs can achieve up to a 68.53% increase in accuracy on Vague Perception samples through our designed pipeline. We further analyze how Vague Perception, Conformity and Degradation vary from model families and parameter sizes, and offer model selection strategies in multi-stage reasoning workflows. Finally, by comparing WakenLLM against mainstream reasoning baselines, both training and non-training ones, we show that existing baselines only activate a small portion of LLMs’ reasoning potential, pointing to perception-aware reasoning as a promising direction for future LLM designing. Code and datasets are available at https://github.com/WakenLLMTeam/WakenLLM-toolkit.

Method: Propose an efficient dialogue evaluator that captures collective wisdom of multiple LLM judges by aggregating their preference knowledge into a single model.

Result: Outperforms existing baselines across seven dialogue evaluation benchmarks in both single rating and pairwise comparison scenarios, demonstrating efficiency and robustness.

Conclusion: The method preserves advantages of diverse multi-judge feedback while drastically reducing evaluation cost, enabling fast, flexible, and fine-grained dialogue quality assessment.

Abstract: Evaluating the conversational abilities of large language models (LLMs) remains a challenging task. Current mainstream approaches primarily rely on the “LLM-as-a-judge” paradigm, where an LLM is prompted to serve as an evaluator to assess dialogue quality. However, such methods often suffer from various biases, which undermine the reliability and consistency of the evaluation results. To mitigate these biases, recent methods employ multiple LLMs as judges and aggregate their judgments to select the optimal assessment. Although effective, this multi-judge approach incurs significant computational overhead during inference. In this paper, we propose an efficient dialogue evaluator that captures the collective wisdom of multiple LLM judges by aggregating their preference knowledge into a single model. Our approach preserves the advantages of diverse multi-judge feedback while drastically reducing the evaluation cost, enabling fast, flexible, and fine-grained dialogue quality assessment. Extensive experiments on seven single rating and pairwise comparison dialogue evaluation benchmarks demonstrate that our method outperforms existing baselines across diverse scenarios, showcasing its efficiency and robustness.

Method: Proposes TreeDiff, a syntax-aware diffusion framework that incorporates structural priors from Abstract Syntax Trees (AST). Instead of random token masking, it selectively masks tokens belonging to key AST nodes, aligning corruption with code structure.

Result: Achieves 13.3% relative improvement over random masking training method, demonstrating effectiveness in code generation by leveraging underlying structural information.

Conclusion: Syntax-aware corruption based on AST structure helps diffusion models internalize compositional nature of programming languages, enabling better reconstruction of programs with proper grammatical boundaries and long-range dependencies.

Abstract: Code generation is increasingly critical for real-world applications. Still, diffusion-based large language models continue to struggle with this demand. Unlike free-form text, code requires syntactic precision; even minor structural inconsistencies can render a program non-executable. Existing diffusion-based large language models rely on random token masking for corruption, leading to two key failures: they lack awareness of syntactic boundaries during the iterative denoising process, and they fail to capture the long-range hierarchical dependencies essential for program correctness. We propose TreeDiff to address both issues. Specifically, we propose a syntax-aware diffusion framework that incorporates structural priors from Abstract Syntax Tree (AST) into the corruption process. Instead of masking individual tokens at random, we selectively mask tokens belonging to key AST nodes. By aligning the corruption process with the underlying structure of code, our method encourages the model to internalize the compositional nature of programming languages, enabling it to reconstruct programs that respect grammatical boundaries and capture long-range dependencies. Our method achieves a 13.3% relative improvement over the random masking training method, demonstrating its effectiveness in code generation task by leveraging underlying structures.

Method: Synthesizes evidence across linguistics, psychology, cognitive science, and computer science to analyze how LLM design and widespread use contribute to standardization effects.

Result: LLMs mirror patterns in training data and amplify convergence as people increasingly rely on the same models, leading to homogenization of language and reasoning styles.

Conclusion: Unchecked homogenization risks flattening cognitive landscapes that drive collective intelligence and adaptability, highlighting the need to preserve cognitive diversity in AI development.

Abstract: Cognitive diversity, reflected in variations of language, perspective, and reasoning, is essential to creativity and collective intelligence. This diversity is rich and grounded in culture, history, and individual experience. Yet as large language models (LLMs) become deeply embedded in people’s lives, they risk standardizing language and reasoning. We synthesize evidence across linguistics, psychology, cognitive science, and computer science to show how LLMs reflect and reinforce dominant styles while marginalizing alternative voices and reasoning strategies. We examine how their design and widespread use contribute to this effect by mirroring patterns in their training data and amplifying convergence as all people increasingly rely on the same models across contexts. Unchecked, this homogenization risks flattening the cognitive landscapes that drive collective intelligence and adaptability.

Method: BiPRM adds a parallel right-to-left evaluation stream via prompt reversal alongside conventional L2R flow, then uses a gating mechanism to adaptively fuse reward scores from both streams for holistic quality assessment.

Result: BiPRM achieves 10.6% average relative gain over 54 solution-level configurations and 37.7% improvement in 12 step-level error detection scenarios, with only 0.3% parameter increase and 5% inference time latency.

Conclusion: BiPRM demonstrates effectiveness, robustness, and general applicability for process-based reward modeling, offering a promising new direction with minimal computational overhead.

Abstract: Process Reward Models (PRMs), which assign fine-grained scores to intermediate reasoning steps within a solution trajectory, have emerged as a promising approach to enhance the reasoning quality of Large Language Models (LLMs). However, most existing PRMs rely on a unidirectional left-to-right (L2R) evaluation scheme, which restricts their utilization of global context. In light of this challenge, we propose a novel bidirectional evaluation paradigm, named Bidirectional Process Reward Model (BiPRM). BiPRM incorporates a parallel right-to-left (R2L) evaluation stream, implemented via prompt reversal, alongside the conventional L2R flow. Then a gating mechanism is introduced to adaptively fuse the reward scores from both streams to yield a holistic quality assessment. Remarkably, compared to the original PRM, BiPRM introduces only a 0.3% parameter increase for the gating module, and the parallel execution of two streams incurs merely 5% inference time latency. Our extensive empirical evaluations spanning diverse benchmarks, LLM backbones, PRM objectives and sampling policies demonstrate that BiPRM consistently surpasses unidirectional baselines, achieving an average relative gain of 10.6% over 54 solution-level configurations and 37.7% in 12 step-level error detection scenarios. Generally, our results highlight the effectiveness, robustness and general applicability of BiPRM, offering a promising new direction for process-based reward modeling.

Method: VocabTailor introduces a decoupled dynamic vocabulary selection framework based on two principles: lexical locality (only small token subsets needed per inference) and asymmetry in vocabulary component characteristics. It offloads embeddings and implements hybrid static-dynamic vocabulary selection for LM heads, enabling on-demand loading of vocabulary components.

Result: Comprehensive experiments show VocabTailor achieves up to 99% reduction in memory usage of vocabulary-related components with minimal or no performance degradation across diverse downstream tasks, substantially outperforming existing static vocabulary pruning methods.

Conclusion: VocabTailor effectively addresses memory constraints for SLM deployment on edge devices through dynamic vocabulary selection, maintaining performance while dramatically reducing memory footprint, offering a superior alternative to rigid static pruning approaches.

Abstract: Small Language Models (SLMs) provide computational advantages in resource-constrained environments, yet memory limitations remain a critical bottleneck for edge device deployment. A substantial portion of SLMs’ memory footprint stems from vocabulary-related components, particularly embeddings and language modeling (LM) heads, due to large vocabulary sizes. Existing static vocabulary pruning, while reducing memory usage, suffers from rigid, one-size-fits-all designs that cause information loss from the prefill stage and a lack of flexibility. In this work, we identify two key principles underlying the vocabulary reduction challenge: the lexical locality principle, the observation that only a small subset of tokens is required during any single inference, and the asymmetry in computational characteristics between vocabulary-related components of SLM. Based on these insights, we introduce VocabTailor, a novel decoupled dynamic vocabulary selection framework that addresses memory constraints through offloading embedding and implements a hybrid static-dynamic vocabulary selection strategy for LM Head, enabling on-demand loading of vocabulary components. Comprehensive experiments across diverse downstream tasks demonstrate that VocabTailor achieves a reduction of up to 99% in the memory usage of vocabulary-related components with minimal or no degradation in task performance, substantially outperforming existing static vocabulary pruning.

Method: Two LLM-based methods: 1) summarizing PubMed abstracts, and 2) using Medical Subject Headings (MeSH) terms. GPT-4o-mini was used to generate profiles for 595 faculty at Columbia University Irving Medical Center, comparing with 167 human-written profiles using manual and automated evaluations (ROUGE-L, BLEU, METEOR, BERTScore, KL Divergence).

Result: Low lexical overlap scores (ROUGE-L, BLEU, METEOR) but moderate semantic similarity (BERTScore F1: 0.542-0.555). MeSH-based profiles rated “good/excellent” by 77.78% of reviewers, with 93.44% favorable readability. 67.86% of panel reviews favored MeSH-derived over abstract-derived profiles. Machine summaries use different keywords than human ones (KL Divergence: 8.56-8.58).

Conclusion: LLMs can automate scientific interest profiling at scale. MeSH-derived profiles have better readability than abstract-derived ones. Machine-generated summaries differ from human-written ones in concept choice, with human summaries containing more novel ideas, highlighting limitations of current evaluation metrics.

Abstract: Research profiles highlight scientists’ research focus, enabling talent discovery and collaborations, but are often outdated. Automated, scalable methods are urgently needed to keep profiles current. We design and evaluate two Large Language Models (LLMs)-based methods to generate scientific interest profiles–one summarizing PubMed abstracts and the other using Medical Subject Headings (MeSH) terms–comparing them with researchers’ self-summarized interests. We collected titles, MeSH terms, and abstracts of PubMed publications for 595 faculty at Columbia University Irving Medical Center, obtaining human-written profiles for 167. GPT-4o-mini was prompted to summarize each researcher’s interests. Manual and automated evaluations characterized similarities between machine-generated and self-written profiles. The similarity study showed low ROUGE-L, BLEU, and METEOR scores, reflecting little terminological overlap. BERTScore analysis revealed moderate semantic similarity (F1: 0.542 for MeSH-based, 0.555 for abstract-based), despite low lexical overlap. In validation, paraphrased summaries achieved a higher F1 of 0.851. Comparing original and manually paraphrased summaries indicated limitations of such metrics. Kullback-Leibler (KL) Divergence of TF-IDF values (8.56 for MeSH-based, 8.58 for abstract-based) suggests machine summaries employ different keywords than human-written ones. Manual reviews showed 77.78% rated MeSH-based profiling “good” or “excellent,” with readability rated favorably in 93.44% of cases, though granularity and accuracy varied. Panel reviews favored 67.86% of MeSH-derived profiles over abstract-derived ones. LLMs promise to automate scientific interest profiling at scale. MeSH-derived profiles have better readability than abstract-derived ones. Machine-generated summaries differ from human-written ones in concept choice, with the latter initiating more novel ideas.

Method: Analyzed predictions across two major ontologies (Human Phenotype Ontology and Gene Ontology) using two high-performing models (GPT-4o and LLaMa 3.1 405B), evaluating nine candidate features related to term familiarity, identifier usage, morphology, and ontology structure with univariate and multivariate analyses

Result: Exposure to ontology identifiers was found to be the strongest predictor of linking success

Conclusion: The study reveals that identifier exposure, rather than term familiarity or other factors, is the key determinant of LLM success in ontology term linking tasks

Abstract: Large language models often perform well on biomedical NLP tasks but may fail to link ontology terms to their correct identifiers. We investigate why these failures occur by analyzing predictions across two major ontologies, Human Phenotype Ontology and Gene Ontology, and two high-performing models, GPT-4o and LLaMa 3.1 405B. We evaluate nine candidate features related to term familiarity, identifier usage, morphology, and ontology structure. Univariate and multivariate analyses show that exposure to ontology identifiers is the strongest predictor of linking success.

Method: Proposed a personality detection model using GPT’s zero-shot capability to infer Big-Five traits from forum introduction posts. Benchmarked against established models and integrated personality detection into SAMI’s matchmaking system, focusing on Extroversion, Agreeableness, and Openness.

Result: GPT models show promising results on this specific dataset but performance varies significantly across traits. Identified potential biases toward optimistic trait inference, especially for traits with skewed distributions. Successfully demonstrated technical feasibility of personality-informed social recommendations.

Conclusion: This work represents initial exploration of personality-informed social recommendations in education. While technically feasible, significant questions remain about what LLMs capture in personality inference and whether personality-based matching meaningfully improves student connections in practice.

Abstract: Social belonging is a vital part of learning, yet online course environments present barriers to the organic formation of social groups. SAMI (Social Agent Mediated Interactions) offers one solution by facilitating student connections, but its effectiveness may be constrained by an incomplete Theory of Mind, limiting its ability to create an effective ‘mental model’ of a student. One facet of this is its inability to intuit personality, which may influence the relevance of its recommendations. To explore this gap, we examine the viability of automated personality inference by proposing a personality detection model utilizing GPT’s zeroshot capability to infer Big-Five personality traits from forum introduction posts, often encouraged in online courses. We benchmark its performance against established models, finding that while GPT models show promising results on this specific dataset, performance varies significantly across traits. We identify potential biases toward optimistic trait inference, particularly for traits with skewed distributions. We demonstrate a proof-of-concept integration of personality detection into SAMI’s entity-based matchmaking system, focusing on three traits with established connections to positive social formation: Extroversion, Agreeableness, and Openness. This work represents an initial exploration of personality-informed social recommendations in educational settings. While our implementation shows technical feasibility, significant questions remain. We discuss these limitations and outline directions for future work, examining what LLMs specifically capture when performing personality inference and whether personality-based matching meaningfully improves student connections in practice.

Method: CAPE-KG ensures KG construction, update, and retrieval are always aligned with MHQA task requirements, maintaining coherent reasoning over both unedited and edited knowledge through a consistency-aware framework.

Result: Extensive experiments on MQuAKE benchmark show accuracy improvements in PPKE performance for MHQA, demonstrating effectiveness of addressing consistency in PPKE.

Conclusion: CAPE-KG successfully addresses consistency issues in parameter-preserving knowledge editing for multi-hop question answering, improving reliability and performance through alignment of KG operations with editing requirements.

Abstract: Parameter-Preserving Knowledge Editing (PPKE) enables updating models with new information without retraining or parameter adjustment. Recent PPKE approaches used knowledge graphs (KG) to extend knowledge editing (KE) capabilities to multi-hop question answering (MHQA). However, these methods often lack consistency, leading to knowledge contamination, unstable updates, and retrieval behaviors that are misaligned with the intended edits. Such inconsistencies undermine the reliability of PPKE in multi-hop reasoning. We present CAPE-KG, Consistency-Aware Parameter-Preserving Editing with Knowledge Graphs, a novel consistency-aware framework for PPKE on MHQA. CAPE-KG ensures KG construction, update, and retrieval are always aligned with the requirements of the MHQA task, maintaining coherent reasoning over both unedited and edited knowledge. Extensive experiments on the MQuAKE benchmark show accuracy improvements in PPKE performance for MHQA, demonstrating the effectiveness of addressing consistency in PPKE.

Method: Proposes BiasDetector benchmark and introduces “Instruction Boundary” concept to systematically evaluate different instruction settings that lead to biases. Tests LLMs on datasets with mixed question forms under various instruction boundary settings.

Result: Experiments show users’ instructions induce large biases on the benchmark, highlighting risks of LLM biased annotation resulting in Sparse Labels mixture and problems arising from users’ instructions to identify them.

Conclusion: Both LLM developers need to recognize risks of biased annotation causing Sparse Labels mixture, and users need to be aware of instruction biases when identifying mixed-form datasets. The work provides tools (code, datasets) for systematic evaluation.

Abstract: Large Language Models (LLMs) annotated datasets are widely used nowadays, however, large-scale annotations often show biases in low-quality datasets. For example, Multiple-Choice Questions (MCQs) datasets with one single correct option is common, however, there may be questions attributed to none or multiple correct options; whereas true-or-false questions are supposed to be labeled with either True or False, but similarly the text can include unsolvable elements, which should be further labeled as Unknown. There are problems when low-quality datasets with mixed question forms can not be identified. We refer to these exceptional label forms as Sparse Labels, and LLMs’ ability to distinguish datasets with Sparse Labels mixture is important. Since users may not know situations of datasets, their instructions can be biased. To study how different instruction settings affect LLMs’ identifications of Sparse Labels mixture, we introduce the concept of Instruction Boundary, which systematically evaluates different instruction settings that lead to biases. We propose BiasDetector, a diagnostic benchmark to systematically evaluate LLMs on datasets with mixed question forms under Instruction Boundary settings. Experiments show that users’ instructions induce large biases on our benchmark, highlighting the need not only for LLM developers to recognize risks of LLM biased annotation resulting in Sparse Labels mixture, but also problems arising from users’ instructions to identify them. Code, datasets and detailed implementations are available at https://github.com/ZpLing/Instruction-Boundary.

Method: Proposes Think-on-Graph 3.0 with MACER mechanism featuring dynamic construction and iterative refinement of a Chunk-Triplets-Community heterogeneous graph index using Dual-Evolution process that adaptively evolves both query and retrieved sub-graph during reasoning.

Result: Extensive experiments show ToG-3 outperforms baselines on both deep and broad reasoning benchmarks, with ablation studies confirming the efficacy of MACER components.

Conclusion: ToG-3 enables precise evidence retrieval and reasoning with lightweight LLMs by dynamically building targeted graph indices tailored to queries, addressing scalability and quality limitations of existing GraphRAG approaches.

Abstract: Graph-based Retrieval-Augmented Generation (GraphRAG) has become the important paradigm for enhancing Large Language Models (LLMs) with external knowledge. However, existing approaches are constrained by their reliance on high-quality knowledge graphs: manually built ones are not scalable, while automatically extracted ones are limited by the performance of LLM extractors, especially when using smaller, local-deployed models. To address this, we introduce Think-on-Graph 3.0 (ToG-3), a novel framework featuring a Multi-Agent Context Evolution and Retrieval (MACER) mechanism. Its core contribution is the dynamic construction and iterative refinement of a Chunk-Triplets-Community heterogeneous graph index, powered by a Dual-Evolution process that adaptively evolves both the query and the retrieved sub-graph during reasoning. ToG-3 dynamically builds a targeted graph index tailored to the query, enabling precise evidence retrieval and reasoning even with lightweight LLMs. Extensive experiments demonstrate that ToG-3 outperforms compared baselines on both deep and broad reasoning benchmarks, and ablation studies confirm the efficacy of the components of MACER framework. The source code are available in https://github.com/DataArcTech/ToG-3.

Method: Introduce MARCH benchmark with 2,209 multi-hop ambiguous questions curated via multi-LLM verification and human annotation. Propose CLARION, a two-stage agentic framework that explicitly decouples ambiguity planning from evidence-driven reasoning.

Result: Even state-of-the-art models struggle with MARCH benchmark, confirming that combining ambiguity resolution with multi-step reasoning is a significant challenge. CLARION framework significantly outperforms existing approaches.

Conclusion: MARCH benchmark reveals the challenge of layered ambiguity in multi-hop reasoning. CLARION’s decoupled approach paves the way for robust reasoning systems that can handle real-world ambiguous queries requiring multi-step inference.

Abstract: Real-world multi-hop QA is naturally linked with ambiguity, where a single query can trigger multiple reasoning paths that require independent resolution. Since ambiguity can occur at any stage, models must navigate layered uncertainty throughout the entire reasoning chain. Despite its prevalence in real-world user queries, previous benchmarks have primarily focused on single-hop ambiguity, leaving the complex interaction between multi-step inference and layered ambiguity underexplored. In this paper, we introduce \textbf{MARCH}, a benchmark for their intersection, with 2,209 multi-hop ambiguous questions curated via multi-LLM verification and validated by human annotation with strong agreement. Our experiments reveal that even state-of-the-art models struggle with MARCH, confirming that combining ambiguity resolution with multi-step reasoning is a significant challenge. To address this, we propose \textbf{CLARION}, a two-stage agentic framework that explicitly decouples ambiguity planning from evidence-driven reasoning, significantly outperforms existing approaches, and paves the way for robust reasoning systems.

Method: Created a constraint-selection-based experiment design with a library of 200 narratology-grounded constraints. Prompted selections from six LLMs under three instruction types: basic, quality-focused, and creativity-focused.

Result: Models consistently prioritize Style over narrative content elements (Event, Character, Setting). Style preferences remain stable across models and instruction types, while content elements show cross-model divergence and instructional sensitivity.

Conclusion: LLMs have latent narrative preferences that should inform how the NLP community evaluates and deploys models in creative domains, with Style being a consistently prioritized element across different conditions.

Abstract: We introduce a constraint-selection-based experiment design for measuring narrative preferences of Large Language Models (LLMs). This design offers an interpretable lens on LLMs’ narrative behavior. We developed a library of 200 narratology-grounded constraints and prompted selections from six LLMs under three different instruction types: basic, quality-focused, and creativity-focused. Findings demonstrate that models consistently prioritize Style over narrative content elements like Event, Character, and Setting. Style preferences remain stable across models and instruction types, whereas content elements show cross-model divergence and instructional sensitivity. These results suggest that LLMs have latent narrative preferences, which should inform how the NLP community evaluates and deploys models in creative domains.

Method: Proposes Self-Filtered Distillation framework with three unsupervised trust metrics: (1) Self-Consistency (stability across multiple generations), (2) Class Entailment Alignment (semantic coherence with patent class definitions), and (3) LLM Agreement Scoring (rationale-label plausibility). These are integrated into a unified trust score that weights training samples and optionally filters low-trust cases.

Result: Experiments on USPTO-2M dataset show the method consistently outperforms label-based learning and conventional distillation in accuracy, stability, and interpretability across diverse student architectures.

Conclusion: Establishes a reliable paradigm for leveraging reasoning-aware trust indicators in patent analytics, demonstrating that treating LLM-generated rationales as trust signals rather than ground-truth supervision improves performance and reliability.

Abstract: Large language models (LLMs) increasingly generate natural language rationales to enhance interpretability, but these often contain logical errors, label mismatches, and domain-specific misalignments. Directly using such rationales as supervision risks propagating noise and undermining training stability. To address this challenge, we introduce Self-Filtered Distillation, a framework tailored for patent classification that treats LLM-generated rationales as trust signals rather than ground-truth supervision. The framework employs selective distillation guided by three unsupervised trust metrics: (1) Self-Consistency, which measures the stability of LLM-generated rationales across multiple generations; (2) Class Entailment Alignment, which assesses semantic coherence with patent-specific class definitions; and (3) LLM Agreement Scoring, which validates rationale-label plausibility. These metrics are integrated into a unified trust score that primarily weights training samples while optionally filtering out extremely low-trust cases, enabling reasoning-aware supervision. Experiments on the USPTO-2M dataset show that our method consistently outperforms label-based learning and conventional distillation in accuracy, stability, and interpretability across diverse student architectures, establishing a reliable paradigm for leveraging reasoning-aware trust indicators in patent analytics.

Method: Introduce Concept, a word-guessing board game as a benchmark for abductive reasoning. Evaluate state-of-the-art LLMs on this game and compare performance across multiple languages (English, Dutch, French, Spanish).

Result: Humans achieve over 90% success rate, but no LLM exceeds 40%. LLMs struggle with interpreting strategic intents and correcting initial hypotheses with sequential updates. Performance drops further in lower-resource languages compared to English.

Conclusion: LLMs have significant limitations in abductive reasoning despite their benchmark successes. The Concept game reveals fundamental weaknesses in strategic reasoning and hypothesis updating that need addressing.

Abstract: Large language models (LLMs) have achieved striking successes on many benchmarks, yet recent studies continue to expose fundamental weaknesses. In this paper, we introduce Concept, a simple word-guessing board game, as a benchmark for probing abductive reasoning. Our results show that this game, easily solved by humans (with a success rate of over 90%), is still very challenging for state-of-the-art LLMs (no model exceeds 40% success rate). Specifically, we observe that LLMs struggle with interpreting other players’ strategic intents, and with correcting initial hypotheses given sequential information updates. In addition, we extend the evaluation across multiple languages, and find that the LLM performance drops further in lower-resource languages (Dutch, French, and Spanish) compared to English.

Method: Combines unsupervised clustering with prompt-based inference using large language models to iteratively construct taxonomies without seed sets or domain expertise.

Result: The framework produces semantically rich topic labels and supports downstream analyses like moral framing. Structured iterative labeling yields more consistent and interpretable labels than existing approaches under human evaluation.

Conclusion: The method is practical for analyzing large-scale political advertising data and provides an effective approach for automatically inducing interpretable topic taxonomies from unlabeled text corpora.

Abstract: Social media platforms play a pivotal role in shaping political discourse, but analyzing their vast and rapidly evolving content remains a major challenge. We introduce an end-to-end framework for automatically inducing an interpretable topic taxonomy from unlabeled text corpora. By combining unsupervised clustering with prompt-based inference, our method leverages large language models (LLMs) to iteratively construct a taxonomy without requiring seed sets (predefined labels) or domain expertise. We validate the framework through a study of political advertising ahead of the 2024 U.S. presidential election. The induced taxonomy yields semantically rich topic labels and supports downstream analyses, including moral framing, in this setting. Results suggest that structured, iterative labeling yields more consistent and interpretable topic labels than existing approaches under human evaluation, and is practical for analyzing large-scale political advertising data.

Method: Complete pipeline including bilingual continued pre-training, instruction tuning, and novel synthesis of human preference data by prompting LLMs to generate “accepted” and “rejected” responses. Used Gemini-2.5-Pro (top-ranked for Irish) to translate English instruction datasets and create first Irish-language human preference dataset.

Result: Qomhrá achieved gains of up to 29% in Irish and 44% in English compared to existing Irish LLM baseline (UCCIX) across translation, gender understanding, topic identification, and world knowledge benchmarks. Found misalignment between LLM-as-a-judge ratings and actual Irish speaker preferences.

Conclusion: The framework provides valuable insights for developing LLMs for Irish and other low-resource languages, demonstrating effective methods for creating bilingual models and synthesizing human preference data under resource constraints.

Abstract: Large language model (LLM) research and development has overwhelmingly focused on the world’s major languages, leading to under-representation of low-resource languages such as Irish. This paper introduces \textbf{Qomhrá}, a bilingual Irish and English LLM, developed under extremely low-resource constraints. A complete pipeline is outlined spanning bilingual continued pre-training, instruction tuning, and the synthesis of human preference data for future alignment training. We focus on the lack of scalable methods to create human preference data by proposing a novel method to synthesise such data by prompting an LLM to generate accepted'' and rejected’’ responses, which we validate as aligning with L1 Irish speakers. To select an LLM for synthesis, we evaluate the top closed-weight LLMs for Irish language generation performance. Gemini-2.5-Pro is ranked highest by L1 and L2 Irish-speakers, diverging from LLM-as-a-judge ratings, indicating a misalignment between current LLMs and the Irish-language community. Subsequently, we leverage Gemini-2.5-Pro to translate a large scale English-language instruction tuning dataset to Irish and to synthesise a first-of-its-kind Irish-language human preference dataset. We comprehensively evaluate Qomhrá across several benchmarks, testing translation, gender understanding, topic identification, and world knowledge; these evaluations show gains of up to 29% in Irish and 44% in English compared to the existing open-source Irish LLM baseline, UCCIX. The results of our framework provide insight and guidance to developing LLMs for both Irish and other low-resource languages.

Method: Translated a medical QA dataset to Persian, used RLAIF to generate rejected-preferred answer pairs with CoT reasoning, applied DPO training with 2M tokens in preferred answers and 2.5M tokens in rejected ones.

Result: The resulting model significantly outperformed its predecessor (gaokerena-V) which was trained on ~57M tokens, despite using much less data, demonstrating efficient reasoning-focused training.

Conclusion: Reasoning-focused training approaches (RLAIF+DPO with CoT) are highly effective for developing domain-specific language models with limited data, especially for underrepresented languages.

Abstract: Enhancing reasoning capabilities in small language models is critical for specialized applications such as medical question answering, particularly in underrepresented languages like Persian. In this study, we employ Reinforcement Learning with AI Feedback (RLAIF) and Direct preference optimization (DPO) to improve the reasoning skills of a general-purpose Persian language model. To achieve this, we translated a multiple-choice medical question-answering dataset into Persian and used RLAIF to generate rejected-preferred answer pairs, which are essential for DPO training. By prompting both teacher and student models to produce Chain-of-Thought (CoT) reasoning responses, we compiled a dataset containing correct and incorrect reasoning trajectories. This dataset, comprising 2 million tokens in preferred answers and 2.5 million tokens in rejected ones, was used to train a baseline model, significantly enhancing its medical reasoning capabilities in Persian. Remarkably, the resulting model outperformed its predecessor, gaokerena-V, which was trained on approximately 57 million tokens, despite leveraging a much smaller dataset. These results highlight the efficiency and effectiveness of reasoning-focused training approaches in developing domain-specific language models with limited data availability.

Method: The authors conduct spectral analysis of RoPE, revealing that its low-frequency components concentrate structured energy leading to low-rank, over-aligned attention patterns. They then introduce DoPE, which identifies noisy attention heads using truncated matrix entropy, suppresses them, and reparameterizes their attention maps with an isotropic Gaussian distribution.

Result: DoPE improves length extrapolation performance without fine-tuning, increases robustness to perturbations, and boosts both needle-in-a-haystack and many-shot in-context learning tasks across various settings.

Conclusion: Selective positional encoding is key to robust extrapolation in LLMs, and DoPE provides an effective training-free solution to mitigate RoPE-induced activation instability.

Abstract: Positional encoding is essential for large language models (LLMs) to represent sequence order, yet recent studies show that Rotary Position Embedding (RoPE) can induce massive activation. We investigate the source of these instabilities via a spectral analysis of RoPE, and show that its low-frequency components concentrate structured energy, producing low-rank, over-aligned attention patterns. We theoretically reveal that this low-frequency alignment manifests as activation noise, degrading stability during long-context extrapolation. To mitigate this effect, we introduce Denoising Rotary Position Embedding (DoPE), a training-free method that identifies and suppresses noisy attention heads using truncated matrix entropy, then reparameterizes their attention maps with an isotropic Gaussian distribution. Across a range of settings, DoPE improves length extrapolation performance without fine-tuning, increases robustness to perturbations, and boosts both needle-in-a-haystack and many-shot in-context learning tasks. These results suggest that selective positional encoding is key to robust extrapolation. Our project page is Project: https://The-physical-picture-of-LLMs.github.io

Method: Integrated approach using role-cross minimal pairs, temporal emergence analysis, and cross-model comparison to study how LLMs implement semantic roles.

Result: Found: (i) highly concentrated circuits (89-94% attribution within 28 nodes); (ii) gradual structural refinement rather than phase transitions; (iii) moderate cross-scale conservation (24-59% component overlap) with high spectral similarity.

Conclusion: LLMs form compact, causally isolated mechanisms for abstract semantic structure that exhibit partial transfer across scales and architectures.

Abstract: Despite displaying semantic competence, large language models’ internal mechanisms that ground abstract semantic structure remain insufficiently characterised. We propose a method integrating role-cross minimal pairs, temporal emergence analysis, and cross-model comparison to study how LLMs implement semantic roles. Our analysis uncovers: (i) highly concentrated circuits (89-94% attribution within 28 nodes); (ii) gradual structural refinement rather than phase transitions, with larger models sometimes bypassing localised circuits; and (iii) moderate cross-scale conservation (24-59% component overlap) alongside high spectral similarity. These findings suggest that LLMs form compact, causally isolated mechanisms for abstract semantic structure, and these mechanisms exhibit partial transfer across scales and architectures.

Method: Reward Auditor uses scientific auditing through hypothesis testing to infer systematic RM vulnerabilities. It quantifies statistical significance and effect size by analyzing distribution degradation of RM preference perception confidence under real-world perturbed scenarios.

Result: The framework enables inference of both certainty and severity of RM vulnerabilities across diverse real-world scenarios, providing a systematic way to identify systematic weaknesses.

Conclusion: Reward Auditor lays foundation for building next-generation LLM alignment systems that are verifiably safe, robust, and trustworthy by addressing the critical dimension of suitability assessment.

Abstract: Reliable reward models (RMs) are critical for ensuring the safe alignment of large language models (LLMs). However, current RM evaluation methods focus solely on preference perception accuracies in given specific scenarios, obscuring the critical vulnerabilities of RMs in real-world scenarios. We identify the true challenge lies in assessing a novel dimension: Suitability, defined as conditional reliability under specific real-world perturbations. To this end, we introduce Reward Auditor, a hypothesis-testing framework specifically designed for RM suitability inference. Rather than answering “How accurate is the RM’s preference perception for given samples?”, it employs scientific auditing to answer: “Can we infer RMs exhibit systematic vulnerabilities in specific real-world scenarios?”. Under real-world perturbed scenarios, Reward Auditor quantifies statistical significance and effect size by auditing distribution degradation of RM preference perception confidence. This enables inference of both the certainty and severity of RM vulnerabilities across diverse real-world scenarios. This lays a solid foundation for building next-generation LLM alignment systems that are verifiably safe, more robust, and trustworthy.

Method: TPA (Token Probability Attribution) mathematically attributes each token’s probability to seven sources: Query, RAG Context, Past Token, Self Token, FFN, Final LayerNorm, and Initial Embedding. It aggregates these attribution scores by Part-of-Speech tags to quantify how each model component contributes to generating specific linguistic categories.

Result: TPA achieves state-of-the-art performance in hallucination detection by identifying patterns such as anomalies where Nouns rely heavily on LayerNorm, which effectively flags hallucinated responses.

Conclusion: TPA provides a comprehensive framework for hallucination detection in RAG systems by capturing the impact of all relevant LLM components through mathematical attribution and POS-based analysis, outperforming previous approaches.

Abstract: Detecting hallucinations in Retrieval-Augmented Generation remains a challenge. Prior approaches attribute hallucinations to a binary conflict between internal knowledge stored in FFNs and the retrieved context. However, this perspective is incomplete, failing to account for the impact of other components of the LLM, such as the user query, previously generated tokens, the self token, and the final LayerNorm adjustment. To comprehensively capture the impact of these components on hallucination detection, we propose TPA which mathematically attributes each token’s probability to seven distinct sources: Query, RAG Context, Past Token, Self Token, FFN, Final LayerNorm, and Initial Embedding. This attribution quantifies how each source contributes to the generation of the next token. Specifically, we aggregate these attribution scores by Part-of-Speech (POS) tags to quantify the contribution of each model component to the generation of specific linguistic categories within a response. By leveraging these patterns, such as detecting anomalies where Nouns rely heavily on LayerNorm, TPA effectively identifies hallucinated responses. Extensive experiments show that TPA achieves state-of-the-art performance.

Method: Proposes d-TreeRPO with tree-structured rollouts and bottom-up advantage computation using verifiable outcome rewards for fine-grained step-wise signals. Also introduces time-scheduled self-distillation loss to enhance prediction confidence and minimize the gap between unbiased expected probabilities and single-step estimates.

Result: Outperforms existing baselines with significant improvements: +86.2% on Sudoku, +51.6% on Countdown, +4.5% on GSM8K, and +5.3% on Math500 compared to base model.

Conclusion: d-TreeRPO provides a reliable RL framework for diffusion LLMs that addresses key reliability bottlenecks through verifiable reward signals and improved probability estimation, achieving state-of-the-art performance on reasoning benchmarks.

Abstract: Reinforcement learning (RL) is pivotal for enhancing the reasoning capabilities of diffusion large language models (dLLMs). However, existing dLLM policy optimization methods suffer from two critical reliability bottlenecks: (1) reward sparsity, arising from coarse or unverifiable signals that impede accurate advantage calculation; and (2) their probability estimates do not account for the gap to the unbiased expectation over all decoding orders, which are intractable to compute. To mitigate these issues, we propose d-TreeRPO, a reliable RL framework for dLLMs that leverages tree-structured rollouts and bottom-up advantage computation based on verifiable outcome rewards to provide fine-grained and verifiable step-wise reward signals. Furthermore, we provide a theoretical proof demonstrating that increasing prediction confidence effectively minimizes the gap between unbiased expected prediction probabilities and its single-step forward pass estimate. Guided by this analysis, we introduce a time-scheduled self-distillation loss during training that enhances prediction confidence in later training stages, thereby enabling more accurate probability estimation and better performance. Experiments demonstrate that d-TreeRPO outperforms existing baselines and achieves significant improvements across multiple reasoning benchmarks. Specifically, it achieves +86.2% on Sudoku, +51.6% on Countdown, +4.5% on GSM8K, and +5.3% on Math500 compared to the base model.

Method: SWAA combines five strategies: (1) applying SWA only during prefilling phase, (2) preserving “sink” tokens, (3) interleaving FA/SWA layers, (4) using chain-of-thought reasoning, and (5) fine-tuning. These are systematically combined as a plug-and-play toolkit.

Result: While individual methods are insufficient, specific synergistic combinations effectively recover original long-context capabilities. The approach achieves 30% to 100% speedups for long-context LLM inference with acceptable quality loss. Recommended configurations are identified for diverse scenarios.

Conclusion: SWAA provides a practical solution for adapting existing full-attention LLMs to efficient sliding window attention without costly pretraining, enabling significant inference speedups for long-context applications while maintaining acceptable performance.

Abstract: The quadratic complexity of self-attention in Transformer-based Large Language Models (LLMs) renders long-context inference prohibitively expensive. While Sliding Window Attention (SWA), the simplest sparse attention pattern, offers a linear-complexity alternative, naively applying it to models pretrained with Full Attention (FA) causes catastrophic long-context performance collapse due to the training-inference mismatch. To address this, we propose Sliding Window Attention Adaptation (SWAA), a plug-and-play toolkit of recipes that adapt FA models to SWA without costly pretraining. SWAA systematically combines five strategies: (1) applying SWA only during prefilling; (2) preserving “sink” tokens; (3) interleaving FA/SWA layers; (4) chain-of-thought (CoT); and (5) fine-tuning. Our experiments demonstrate that while individual methods are insufficient, specific synergistic combinations can effectively recover original long-context capabilities. After further analyzing performance-efficiency trade-offs, we identify recommended SWAA configurations for diverse scenarios, which achieve 30% to 100% speedups for long-context LLM inference with acceptable quality loss. Our code is available at https://github.com/yuyijiong/sliding-window-attention-adaptation

Method: Train LLMs (Activation Oracles) to accept LLM activations as inputs and answer natural language questions about them. Evaluate in out-of-distribution settings, test scaling with training data diversity, and compare against white- and black-box baselines on four downstream tasks.

Result: AOs can recover information fine-tuned into models (e.g., biographical knowledge) not in input text. Narrowly-trained models generalize well, and adding diverse training datasets yields consistent improvements. Best AOs match or exceed white-box baselines on all tasks and best overall baseline on 3 of 4.

Conclusion: Diversified training to answer natural-language queries imparts a general capability to verbalize information about LLM activations, suggesting LatentQA is a promising simpler approach for LLM interpretability.

Abstract: Large language model (LLM) activations are notoriously difficult to understand, with most existing techniques using complex, specialized methods for interpreting them. Recent work has proposed a simpler approach known as LatentQA: training LLMs to directly accept LLM activations as inputs and answer arbitrary questions about them in natural language. However, prior work has focused on narrow task settings for both training and evaluation. In this paper, we instead take a generalist perspective. We evaluate LatentQA-trained models, which we call Activation Oracles (AOs), in far out-of-distribution settings and examine how performance scales with training data diversity. We find that AOs can recover information fine-tuned into a model (e.g., biographical knowledge or malign propensities) that does not appear in the input text, despite never being trained with activations from a fine-tuned model. Our main evaluations are four downstream tasks where we can compare to prior white- and black-box techniques. We find that even narrowly-trained LatentQA models can generalize well, and that adding additional training datasets (such as classification tasks and a self-supervised context prediction task) yields consistent further improvements. Our best AOs match or exceed white-box baselines on all four tasks and the best overall baseline on 3 of 4. These results suggest that diversified training to answer natural-language queries imparts a general capability to verbalize information about LLM activations.

Method: Created ShareChat by collecting 142,808 conversations (660,293 turns) directly from publicly shared URLs on five major LLM platforms, preserving native platform affordances like citations and thinking traces. The dataset covers 101 languages and spans from April 2023 to October 2025.

Result: ShareChat offers substantially longer context windows and greater interaction depth than prior datasets. The paper demonstrates the dataset’s utility through three case studies: completeness analysis of intent satisfaction, citation study of model grounding, and temporal analysis of engagement rhythms.

Conclusion: ShareChat provides a vital resource for understanding authentic user-LLM chatbot interactions in the wild, addressing the limitations of current datasets by preserving platform-specific features and capturing real-world usage patterns across multiple commercial LLM products.

Abstract: While academic research typically treats Large Language Models (LLM) as generic text generators, they are distinct commercial products with unique interfaces and capabilities that fundamentally shape user behavior. Current datasets obscure this reality by collecting text-only data through uniform interfaces that fail to capture authentic chatbot usage. To address this limitation, we present ShareChat, a large-scale corpus of 142,808 conversations (660,293 turns) sourced directly from publicly shared URLs on ChatGPT, Perplexity, Grok, Gemini, and Claude. ShareChat distinguishes itself by preserving native platform affordances, such as citations and thinking traces, across a diverse collection covering 101 languages and the period from April 2023 to October 2025. Furthermore, ShareChat offers substantially longer context windows and greater interaction depth than prior datasets. To illustrate the dataset’s breadth, we present three case studies: a completeness analysis of intent satisfaction, a citation study of model grounding, and a temporal analysis of engagement rhythms. This work provides the community with a vital and timely resource for understanding authentic user-LLM chatbot interactions in the wild. The dataset will be publicly available.

Method: The researchers investigate whether n-gram representations (shorter sequences) can serve as cognitive units of planning, comparing them against whole utterance contexts to see if they capture the same probabilistic reduction effects.

Result: The study demonstrates that n-gram representations are sufficient to observe probabilistic reduction, meaning whole utterances are not necessary for studying these language model-cognition relationships.

Conclusion: N-gram representations suffice as cognitive units of planning for investigating relationships between language model probabilities and cognitive phenomena, providing a more efficient approach than using whole utterances.

Abstract: The primary research questions of this paper center on defining the amount of context that is necessary and/or appropriate when investigating the relationship between language model probabilities and cognitive phenomena. We investigate whether whole utterances are necessary to observe probabilistic reduction and demonstrate that n-gram representations suffice as cognitive units of planning.

Method: LongCat ZigZag Attention (LoZA) - a sparse attention scheme designed to transform existing full-attention models into sparse versions. Applied to LongCat-Flash during mid-training to create LongCat-Flash-Exp as a long-context foundation model.

Result: Achieves significant speed-ups in long-context scenarios for both prefill-intensive and decode-intensive cases. Enables processing of up to 1 million tokens efficiently, supporting long-term reasoning and long-horizon agentic capabilities.

Conclusion: LoZA provides an effective sparse attention solution that makes long-context processing computationally feasible, enabling new capabilities in long-term reasoning and agentic applications while maintaining efficiency.

Abstract: We introduce LongCat ZigZag Attention (LoZA), which is a sparse attention scheme designed to transform any existing full-attention models into sparse versions with rather limited compute budget. In long-context scenarios, LoZA can achieve significant speed-ups both for prefill-intensive (e.g., retrieval-augmented generation) and decode-intensive (e.g., tool-integrated reasoning) cases. Specifically, by applying LoZA to LongCat-Flash during mid-training, we serve LongCat-Flash-Exp as a long-context foundation model that can swiftly process up to 1 million tokens, enabling efficient long-term reasoning and long-horizon agentic capabilities.

Method: FIGR integrates executable visual construction into reasoning via end-to-end RL. It externalizes intermediate hypotheses by generating executable code that constructs diagrams within the reasoning loop, with adaptive reward mechanism regulating when visual construction is invoked.

Result: Outperforms strong text-only chain-of-thought baselines, improving base model by 13.12% on AIME 2025 and 11.00% on BeyondAIME across eight challenging mathematical benchmarks.

Conclusion: FIGR demonstrates effectiveness of precise, controllable figure construction in enhancing complex reasoning ability, particularly for latent global properties difficult to infer from text alone.

Abstract: Complex reasoning problems often involve implicit spatial and geometric relationships that are not explicitly encoded in text. While recent reasoning models perform well across many domains, purely text-based reasoning struggles to capture structural constraints in complex settings. In this paper, we introduce FIGR, which integrates executable visual construction into multi-turn reasoning via end-to-end reinforcement learning. Rather than relying solely on textual chains of thought, FIGR externalizes intermediate hypotheses by generating executable code that constructs diagrams within the reasoning loop. An adaptive reward mechanism selectively regulates when visual construction is invoked, enabling more consistent reasoning over latent global properties that are difficult to infer from text alone. Experiments on eight challenging mathematical benchmarks demonstrate that FIGR outperforms strong text-only chain-of-thought baselines, improving the base model by 13.12% on AIME 2025 and 11.00% on BeyondAIME. These results highlight the effectiveness of precise, controllable figure construction of FIGR in enhancing complex reasoning ability.

Method: Extract standalone knowledge statements from authoritative textbooks, then dynamically compose evaluation questions by randomly sampling 8-10 statements at test time. This creates a vast combinatorial space that prevents memorization, enables multi-knowledge assessment, and reduces annotation costs (annotators only verify formatting compliance).

Result: Tests on 50+ LLMs show strong discriminative power: OpenAI-GPT-5.1 achieves only 62.07% accuracy. Clear performance gradient: reasoning models range 16.04%-62.07%, chat models range 9.71%-50.40%. Model rankings remain stable across dynamically generated question sets, enabling reliable periodic dataset refresh.

Conclusion: Encyclo-K successfully addresses the three limitations of existing benchmarks through its statement-based, dynamically composed approach. It provides a scalable framework for evaluating LLMs’ comprehensive understanding across multiple fine-grained knowledge statements, with strong discriminative power and resistance to data contamination.

Abstract: Benchmarks play a crucial role in tracking the rapid advancement of large language models (LLMs) and identifying their capability boundaries. However, existing benchmarks predominantly curate questions at the question level, suffering from three fundamental limitations: vulnerability to data contamination, restriction to single-knowledge-point assessment, and reliance on costly domain expert annotation. We propose Encyclo-K, a statement-based benchmark that rethinks benchmark construction from the ground up. Our key insight is that knowledge statements, not questions, can serve as the unit of curation, and questions can then be constructed from them. We extract standalone knowledge statements from authoritative textbooks and dynamically compose them into evaluation questions through random sampling at test time. This design directly addresses all three limitations: the combinatorial space is too vast to memorize, and model rankings remain stable across dynamically generated question sets, enabling reliable periodic dataset refresh; each question aggregates 8-10 statements for comprehensive multi-knowledge assessment; annotators only verify formatting compliance without requiring domain expertise, substantially reducing annotation costs. Experiments on over 50 LLMs demonstrate that Encyclo-K poses substantial challenges with strong discriminative power. Even the top-performing OpenAI-GPT-5.1 achieves only 62.07% accuracy, and model performance displays a clear gradient distribution–reasoning models span from 16.04% to 62.07%, while chat models range from 9.71% to 50.40%. These results validate the challenges introduced by dynamic evaluation and multi-statement comprehensive understanding. These findings establish Encyclo-K as a scalable framework for dynamic evaluation of LLMs’ comprehensive understanding over multiple fine-grained disciplinary knowledge statements.

Method: Formulates structured inference as adaptive constraint propagation and trains a Graph Attention Network using meta-reinforcement learning to learn universal policies applicable across languages and tasks without task-specific retraining.

Result: Achieves 1.5-2x speedups over GPU-optimized baselines while maintaining within 0.2% accuracy of state-of-the-art parsers. Demonstrates rapid cross-domain adaptation: policies trained on English parsing adapt to new languages/tasks with 5-10 gradient steps (5-15 seconds) instead of hours of training.

Conclusion: MetaJuLS enables efficient structured inference for LLMs through learned constraint propagation policies that generalize across languages and tasks, reducing inference carbon footprint and contributing to Green AI while discovering both human-like and novel parsing strategies.

Abstract: Large language models increasingly require structured inference, from JSON schema enforcement to multi-lingual parsing, where outputs must satisfy complex constraints. We introduce MetaJuLS, a meta-reinforcement learning approach that learns universal constraint propagation policies applicable across languages and tasks without task-specific retraining. By formulating structured inference as adaptive constraint propagation and training a Graph Attention Network with meta-learning, MetaJuLS achieves 1.5–2.0$\times$ speedups over GPU-optimized baselines while maintaining within 0.2% accuracy of state-of-the-art parsers. On Universal Dependencies across 10 languages and LLM-constrained generation (LogicBench, GSM8K-Constrained), MetaJuLS demonstrates rapid cross-domain adaptation: a policy trained on English parsing adapts to new languages and tasks with 5–10 gradient steps (5–15 seconds) rather than requiring hours of task-specific training. Mechanistic analysis reveals the policy discovers human-like parsing strategies (easy-first) and novel non-intuitive heuristics. By reducing propagation steps in LLM deployments, MetaJuLS contributes to Green AI by directly reducing inference carbon footprint.

Method: The researchers use Qwen3 models and InstrumentalEval benchmark to measure convergence rates for harmful behaviors like shutdown avoidance and self-replication. They test different prompting strategies: pro-instrumental vs anti-instrumental suffixes. They compare performance across model sizes (30B vs smaller) and model types (Instruct vs Thinking variants).

Result: Anti-instrumental prompting dramatically reduces harmful behavior convergence rates. For Qwen3-30B Instruct, convergence drops from 81.69% with pro-instrumental suffix to 2.82% with anti-instrumental suffix. Larger aligned models show lower convergence rates than smaller ones under anti-instrumental prompting (2.82% vs 4.23% for Instruct; 4.23% vs 9.86% for Thinking).

Conclusion: The paper demonstrates that steerability can be modulated through prompting strategies, revealing a safety-security tradeoff in AI development. Anti-instrumental prompting shows promise for reducing harmful behaviors, but the fundamental tension between safety and security remains a significant challenge, especially for open-weight models.

Abstract: We examine two properties of AI systems: capability (what a system can do) and steerability (how reliably one can shift behavior toward intended outcomes). A central question is whether capability growth reduces steerability and risks control collapse. We also distinguish between authorized steerability (builders reliably reaching intended behaviors) and unauthorized steerability (attackers eliciting disallowed behaviors). This distinction highlights a fundamental safety–security dilemma of AI models: safety requires high steerability to enforce control (e.g., stop/refuse), while security requires low steerability for malicious actors to elicit harmful behaviors. This tension presents a significant challenge for open-weight models, which currently exhibit high steerability via common techniques like fine-tuning or adversarial attacks. Using Qwen3 and InstrumentalEval, we find that a short anti-instrumental prompt suffix sharply reduces the measured convergence rate (e.g., shutdown avoidance, self-replication). For Qwen3-30B Instruct, the convergence rate drops from 81.69% under a pro-instrumental suffix to 2.82% under an anti-instrumental suffix. Under anti-instrumental prompting, larger aligned models show lower convergence rates than smaller ones (Instruct: 2.82% vs. 4.23%; Thinking: 4.23% vs. 9.86%). Code is available at github.com/j-hoscilowicz/instrumental_steering.

Method: Proposes a novel fine-tuning method specifically designed to enhance LLMs’ ability to distinguish between relevant and irrelevant information within retrieved documents. The method addresses the structural constraints of attention patterns that limit standard fine-tuning approaches.

Result: Extensive experiments across multiple benchmarks show that the proposed approach significantly improves the robustness and performance of LLMs when dealing with noisy retrieved documents in RAG systems.

Conclusion: Instead of relying solely on better filtering of retrieved documents, LLMs should be made more robust to noise through specialized fine-tuning that enhances their ability to selectively utilize relevant information while ignoring irrelevant content in RAG settings.

Abstract: Retrieval-Augmented Generation (RAG) has become a widely adopted approach to enhance Large Language Models (LLMs) by incorporating external knowledge and reducing hallucinations. However, noisy or irrelevant documents are often introduced during RAG, potentially degrading performance and even causing hallucinated outputs. While various methods have been proposed to filter out such noise, we argue that identifying irrelevant information from retrieved content is inherently difficult and limited number of transformer layers can hardly solve this. Consequently, retrievers fail to filter out irrelevant documents entirely. Therefore, LLMs must be robust against such noise, but we demonstrate that standard fine-tuning approaches are often ineffective in enabling the model to selectively utilize relevant information while ignoring irrelevant content due to the structural constraints of attention patterns. To address this, we propose a novel fine-tuning method designed to enhance the model’s ability to distinguish between relevant and irrelevant information within retrieved documents. Extensive experiments across multiple benchmarks show that our approach significantly improves the robustness and performance of LLMs.

Method: Created RoBench25 benchmark to evaluate information retrieval under HiSPAs, tested SSMs and Transformers, analyzed Jamba hybrid model, and conducted interpretability study of Mamba’s hidden layers during attacks.

Result: SSMs are vulnerable to HiSPAs while Transformers are not; even 52B Jamba hybrid model collapses on RoBench25 under optimized triggers; HiSPA triggers weaken Jamba on Open-Prompt-Injections benchmark; interpretability reveals patterns for potential mitigation.

Conclusion: SSMs have critical vulnerability to hidden state poisoning attacks that cause irreversible information loss, highlighting a security weakness compared to Transformers, but interpretability patterns could enable mitigation systems.

Abstract: State space models (SSMs) like Mamba offer efficient alternatives to Transformer-based language models, with linear time complexity. Yet, their adversarial robustness remains critically unexplored. This paper studies the phenomenon whereby specific short input phrases induce a partial amnesia effect in such models, by irreversibly overwriting information in their hidden states, referred to as a Hidden State Poisoning Attack (HiSPA). Our benchmark RoBench25 allows evaluating a model’s information retrieval capabilities when subject to HiSPAs, and confirms the vulnerability of SSMs against such attacks. Even a recent 52B hybrid SSM-Transformer model from the Jamba family collapses on RoBench25 under optimized HiSPA triggers, whereas pure Transformers do not. We also observe that HiSPA triggers significantly weaken the Jamba model on the popular Open-Prompt-Injections benchmark, unlike pure Transformers. Finally, our interpretability study reveals patterns in Mamba’s hidden layers during HiSPAs that could be used to build a HiSPA mitigation system. The full code and data to reproduce the experiments can be found at https://anonymous.4open.science/r/hispa_anonymous-5DB0.

Method: Created pdfQA dataset with 2K human-annotated (real-pdfQA) and 2K synthetic (syn-pdfQA) QA pairs from PDFs, categorized by ten complexity dimensions (file type, source modality, source position, answer type, etc.). Applied quality and difficulty filters to obtain valid and challenging QA pairs.

Result: Evaluated open-source LLMs on pdfQA, revealing performance challenges that correlate with the defined complexity dimensions. The dataset provides a basis for end-to-end QA pipeline evaluation and testing diverse skill sets.

Conclusion: pdfQA addresses the gap in PDF-based QA evaluation, offering a multi-domain dataset with complexity dimensions that help identify specific challenges in LLM performance on PDF documents, enabling better pipeline optimization.

Abstract: PDFs are the second-most used document type on the internet (after HTML). Yet, existing QA datasets commonly start from text sources or only address specific domains. In this paper, we present pdfQA, a multi-domain 2K human-annotated (real-pdfQA) and 2K synthetic dataset (syn-pdfQA) differentiating QA pairs in ten complexity dimensions (e.g., file type, source modality, source position, answer type). We apply and evaluate quality and difficulty filters on both datasets, obtaining valid and challenging QA pairs. We answer the questions with open-source LLMs, revealing existing challenges that correlate with our complexity dimensions. pdfQA presents a basis for end-to-end QA pipeline evaluation, testing diverse skill sets and local optimizations (e.g., in information retrieval or parsing).

Method: Proposed Dense-MAE: a self-supervised framework using Masked Autoencoders for 3D medical data. Randomly masks 3D patches of vessel lumen and trains Dense-Unet to reconstruct missing geometry, forcing encoder to learn arterial topology features without human annotation.

Result: Experimental results show that initializing calcium removal networks with MAE-based weights significantly improves inpainting accuracy and stenosis estimation compared to training from scratch, particularly in few-shot scenarios.

Conclusion: Dense-MAE provides an effective self-supervised pre-training strategy for medical imaging that reduces dependency on labeled data and improves performance in coronary calcification removal tasks.

Abstract: Coronary calcification creates blooming artifacts in Computed Tomography Angiography (CTA), severely hampering the diagnosis of lumen stenosis. While Deep Convolutional Neural Networks (DCNNs) like Dense-Unet have shown promise in removing these artifacts via inpainting, they often require large labeled datasets which are scarce in the medical domain. Inspired by recent advancements in Masked Autoencoders (MAE) for 3D point clouds, we propose \textbf{Dense-MAE}, a novel self-supervised learning framework for volumetric medical data. We introduce a pre-training strategy that randomly masks 3D patches of the vessel lumen and trains the Dense-Unet to reconstruct the missing geometry. This forces the encoder to learn high-level latent features of arterial topology without human annotation. Experimental results on clinical CTA datasets demonstrate that initializing the Calcium Removal network with our MAE-based weights significantly improves inpainting accuracy and stenosis estimation compared to training from scratch, specifically in few-shot scenarios.

Method: 1) View Cone Sampling (VCS) strategy using Gaussian-distributed ray bundles to simulate human foveal receptive field for robust sampling on complex topologies. 2) Hybrid Manifold-Euclidean Constrained Diffusion (HCD) algorithm that fuses manifold geodesic constraints with Euclidean scales for topologically-consistent saliency propagation.

Result: The framework mitigates “topological short-circuits” and aliasing, providing high-fidelity 3D attention acquisition that aligns with natural human perception, offering more accurate and robust baseline for 3D mesh saliency research.

Conclusion: Proposed framework addresses limitations of current methods by better simulating human visual perception in 3D environments, providing reliable ground truth for human-centric visual modeling in VR applications.

Abstract: Reliable 3D mesh saliency ground truth (GT) is essential for human-centric visual modeling in virtual reality (VR). However, current 3D mesh saliency GT acquisition methods are generally consistent with 2D image methods, ignoring the differences between 3D geometry topology and 2D image array. Current VR eye-tracking pipelines rely on single ray sampling and Euclidean smoothing, triggering texture attention and signal leakage across gaps. This paper proposes a robust framework to address these limitations. We first introduce a view cone sampling (VCS) strategy, which simulates the human foveal receptive field via Gaussian-distributed ray bundles to improve sampling robustness for complex topologies. Furthermore, a hybrid Manifold-Euclidean constrained diffusion (HCD) algorithm is developed, fusing manifold geodesic constraints with Euclidean scales to ensure topologically-consistent saliency propagation. By mitigating “topological short-circuits” and aliasing, our framework provides a high-fidelity 3D attention acquisition paradigm that aligns with natural human perception, offering a more accurate and robust baseline for 3D mesh saliency research.

Method: MIAR (Modality Interaction and Alignment Representation) integrates contextual features across modalities using feature interaction to generate four tokens representing global representations of how each modality extracts information from others. It aligns different modalities using contrastive learning and normalization strategies.

Result: Experiments on CMU-MOSI and CMU-MOSEI benchmarks demonstrate that MIAR outperforms state-of-the-art MER methods.

Conclusion: MIAR effectively addresses modality distribution differences and varying contributions through its interaction and alignment approach, achieving superior performance in multimodal emotion recognition.

Abstract: Multimodal Emotion Recognition (MER) aims to perceive human emotions through three modes: language, vision, and audio. Previous methods primarily focused on modal fusion without adequately addressing significant distributional differences among modalities or considering their varying contributions to the task. They also lacked robust generalization capabilities across diverse textual model features, thus limiting performance in multimodal scenarios. Therefore, we propose a novel approach called Modality Interaction and Alignment Representation (MIAR). This network integrates contextual features across different modalities using a feature interaction to generate feature tokens to represent global representations of this modality extracting information from other modalities. These four tokens represent global representations of how each modality extracts information from others. MIAR aligns different modalities using contrastive learning and normalization strategies. We conduct experiments on two benchmarks: CMU-MOSI and CMU-MOSEI datasets, experimental results demonstrate the MIAR outperforms state-of-the-art MER methods.

Method: 1) Extract multi-channel features for visual and auditory modalities to enhance intra-modal representation. 2) Propose symmetric mutual promotion (SMP) inter-modal fusion combining symmetric cross-modal attention (capturing useful info from other modalities) and self-attention (modeling contextual info). 3) Integrate intra-modal and inter-modal fused features to leverage both complementarity and differences.

Result: Experiments on two benchmark datasets demonstrate the effectiveness and superiority of the proposed method compared to existing approaches.

Conclusion: The proposed approach addresses key limitations in multimodal sentiment analysis by improving both intra-modal feature richness and inter-modal fusion through attention mechanisms that consider both consistency and differences between modalities.

Abstract: Multimodal sentiment analysis is a key technology in the fields of human-computer interaction and affective computing. Accurately recognizing human emotional states is crucial for facilitating smooth communication between humans and machines. Despite some progress in multimodal sentiment analysis research, numerous challenges remain. The first challenge is the limited and insufficiently rich features extracted from single modality data. Secondly, most studies focus only on the consistency of inter-modal feature information, neglecting the differences between features, resulting in inadequate feature information fusion. In this paper, we first extract multi-channel features to obtain more comprehensive feature information. We employ dual-channel features in both the visual and auditory modalities to enhance intra-modal feature representation. Secondly, we propose a symmetric mutual promotion (SMP) inter-modal feature fusion method. This method combines symmetric cross-modal attention mechanisms and self-attention mechanisms, where the cross-modal attention mechanism captures useful information from other modalities, and the self-attention mechanism models contextual information. This approach promotes the exchange of useful information between modalities, thereby strengthening inter-modal interactions. Furthermore, we integrate intra-modal features and inter-modal fused features, fully leveraging the complementarity of inter-modal feature information while considering feature information differences. Experiments conducted on two benchmark datasets demonstrate the effectiveness and superiority of our proposed method.

Method: Three-stage pipeline: 1) Large-scale data acquisition/filtering with VLMs and tri-modal consistency checks, 2) Mixture-of-experts knowledge distillation with complementary encoders and uncertainty sampling, 3) Region-evidence VLM adversarial labeling with proposer-validator architecture for explainable labels.

Result: Runs on consumer-grade GPUs, steadily improves purity and separability, requires minimal human supervision, and provides practical substrate for downstream training/evaluation under resource constraints.

Conclusion: ReCCur offers a practical solution for scalable corner-case curation from web data with low compute requirements and minimal human intervention, enabling better training and evaluation for real-world AI systems.

Abstract: Corner cases are rare or extreme scenarios that drive real-world failures, but they are difficult to curate at scale: web data are noisy, labels are brittle, and edge deployments preclude large retraining. We present ReCCur (Recursive Corner-Case Curation), a low-compute framework that converts noisy web imagery into auditable fine-grained labels via a multi-agent recursive pipeline. First, large-scale data acquisition and filtering expands a domain vocabulary with a vision-language model (VLM), crawls the web, and enforces tri-modal (image, description, keyword) consistency with light human spot checks to yield refined candidates. Next, mixture-of-experts knowledge distillation uses complementary encoders (e.g., CLIP, DINOv2, BEiT) for kNN voting with dual-confidence activation and uncertainty sampling, converging to a high-precision set. Finally, region-evidence VLM adversarial labeling pairs a proposer (multi-granularity regions and semantic cues) with a validator (global and local chained consistency) to produce explainable labels and close the loop. On realistic corner-case scenarios (e.g., flooded-car inspection), ReCCur runs on consumer-grade GPUs, steadily improves purity and separability, and requires minimal human supervision, providing a practical substrate for downstream training and evaluation under resource constraints. Code and dataset will be released.

Method: Proposes CoCoT framework with two key innovations: 1) Dynamic Multi-Region Grounding to adaptively detect relevant image regions based on questions, and 2) Relation-Aware Reasoning to enable multi-region collaboration through iterative visual cue alignment for coherent reasoning chains.

Result: Created CoCoT-70K dataset with 74,691 high-quality samples featuring multi-region annotations and structured reasoning chains. Achieved average accuracy improvements of 15.4% on LLaVA-1.5 and 4.0% on Qwen2-VL across six challenging benchmarks.

Conclusion: CoCoT significantly enhances complex visual reasoning by addressing cross-modal integration challenges through collaborative multi-region grounding and relation-aware reasoning, with publicly available data and code.

Abstract: Multi-modal reasoning requires the seamless integration of visual and linguistic cues, yet existing Chain-of-Thought methods suffer from two critical limitations in cross-modal scenarios: (1) over-reliance on single coarse-grained image regions, and (2) semantic fragmentation between successive reasoning steps. To address these issues, we propose the CoCoT (Collaborative Coross-modal Thought) framework, built upon two key innovations: a) Dynamic Multi-Region Grounding to adaptively detect the most relevant image regions based on the question, and b) Relation-Aware Reasoning to enable multi-region collaboration by iteratively aligning visual cues to form a coherent and logical chain of thought. Through this approach, we construct the CoCoT-70K dataset, comprising 74,691 high-quality samples with multi-region annotations and structured reasoning chains. Extensive experiments demonstrate that CoCoT significantly enhances complex visual reasoning, achieving an average accuracy improvement of 15.4% on LLaVA-1.5 and 4.0% on Qwen2-VL across six challenging benchmarks. The data and code are available at: https://github.com/deer-echo/CoCoT.

Method: 1) Created internet-scale video-action dataset by automatically extracting player actions from gameplay videos, 2) Developed multi-game benchmark for measuring cross-game generalization, 3) Trained unified vision-action model using large-scale behavior cloning on 40,000 hours of gameplay across 1,000+ games.

Result: NitroGen demonstrates strong competence across diverse game domains including 3D action combat, 2D platformer precision control, and procedurally generated world exploration. It achieves up to 52% relative improvement in task success rates over models trained from scratch when transferred to unseen games.

Conclusion: NitroGen represents an effective vision-action foundation model for generalist gaming agents with strong cross-game generalization capabilities. The authors release the dataset, evaluation suite, and model weights to advance research on generalist embodied agents.

Abstract: We introduce NitroGen, a vision-action foundation model for generalist gaming agents that is trained on 40,000 hours of gameplay videos across more than 1,000 games. We incorporate three key ingredients: 1) an internet-scale video-action dataset constructed by automatically extracting player actions from publicly available gameplay videos, 2) a multi-game benchmark environment that can measure cross-game generalization, and 3) a unified vision-action model trained with large-scale behavior cloning. NitroGen exhibits strong competence across diverse domains, including combat encounters in 3D action games, high-precision control in 2D platformers, and exploration in procedurally generated worlds. It transfers effectively to unseen games, achieving up to 52% relative improvement in task success rates over models trained from scratch. We release the dataset, evaluation suite, and model weights to advance research on generalist embodied agents.

Method: Systematic ablation studies manipulating three MLLM components: vision encoder, connector, and LLM across diverse training strategies. Evaluated on knee OA radiograph classification using both small class-balanced datasets (500 images) and larger class-imbalanced datasets (5,778 images) with LoRA fine-tuning.

Result: 1) Trained vision encoder alone outperformed full MLLM pipelines in classification accuracy. 2) Fine-tuning the LLM provided no meaningful improvement over prompt-based guidance. 3) LoRA fine-tuning on small class-balanced dataset (500 images) gave better results than training on much larger but class-imbalanced set (5,778 images).

Conclusion: For domain-specific medical classification, LLMs are more effective as interpreters and report generators rather than primary classifiers. MLLM architecture appears less suitable for medical image diagnostic classification tasks requiring high certainty. Recommendations: prioritize vision encoder optimization and careful dataset curation for clinically applicable systems.

Abstract: Multimodal large language models (MLLMs) show promising performance on medical visual question answering (VQA) and report generation, but these generation and explanation abilities do not reliably transfer to disease-specific classification. We evaluated MLLM architectures on knee osteoarthritis (OA) radiograph classification, which remains underrepresented in existing medical MLLM benchmarks, even though knee OA affects an estimated 300 to 400 million people worldwide. Through systematic ablation studies manipulating the vision encoder, the connector, and the large language model (LLM) across diverse training strategies, we measured each component’s contribution to diagnostic accuracy. In our classification task, a trained vision encoder alone could outperform full MLLM pipelines in classification accuracy and fine-tuning the LLM provided no meaningful improvement over prompt-based guidance. And LoRA fine-tuning on a small, class-balanced dataset (500 images) gave better results than training on a much larger but class-imbalanced set (5,778 images), indicating that data balance and quality can matter more than raw scale for this task. These findings suggest that for domain-specific medical classification, LLMs are more effective as interpreters and report generators rather than as primary classifiers. Therefore, the MLLM architecture appears less suitable for medical image diagnostic classification tasks that demand high certainty. We recommend prioritizing vision encoder optimization and careful dataset curation when developing clinically applicable systems.

Method: 1) GMM-based metric dataset construction: Each device fits a lightweight Gaussian Mixture Model to approximate its private data distribution and uploads only GMM parameters. Cloud selects distribution-consistent samples from public data to construct task-representative metric datasets. 2) Dual-granularity importance evaluation: Jointly measures composite neuron importance and adaptive layer importance for fine-grained, task-aware pruning tailored to each device’s computational budget.

Result: Extensive experiments across multiple ViT backbones and datasets demonstrate that TAP-ViTs consistently outperforms state-of-the-art pruning methods under comparable compression ratios.

Conclusion: TAP-ViTs enables task-customized ViT pruning in privacy-preserving mobile computing settings by generating device-specific pruned models without requiring access to raw local data, addressing both device heterogeneity and privacy constraints.

Abstract: Vision Transformers (ViTs) have demonstrated strong performance across a wide range of vision tasks, yet their substantial computational and memory demands hinder efficient deployment on resource-constrained mobile and edge devices. Pruning has emerged as a promising direction for reducing ViT complexity. However, existing approaches either (i) produce a single pruned model shared across all devices, ignoring device heterogeneity, or (ii) rely on fine-tuning with device-local data, which is often infeasible due to limited on-device resources and strict privacy constraints. As a result, current methods fall short of enabling task-customized ViT pruning in privacy-preserving mobile computing settings. This paper introduces TAP-ViTs, a novel task-adaptive pruning framework that generates device-specific pruned ViT models without requiring access to any raw local data. Specifically, to infer device-level task characteristics under privacy constraints, we propose a Gaussian Mixture Model (GMM)-based metric dataset construction mechanism. Each device fits a lightweight GMM to approximate its private data distribution and uploads only the GMM parameters. Using these parameters, the cloud selects distribution-consistent samples from public data to construct a task-representative metric dataset for each device. Based on this proxy dataset, we further develop a dual-granularity importance evaluation-based pruning strategy that jointly measures composite neuron importance and adaptive layer importance, enabling fine-grained, task-aware pruning tailored to each device’s computational budget. Extensive experiments across multiple ViT backbones and datasets demonstrate that TAP-ViTs consistently outperforms state-of-the-art pruning methods under comparable compression ratios.

Method: The researchers compare existing BIQA models with three paradigms designed to learn image-text-score relationships: 1) Chain-of-Thought, 2) Self-Consistency, and 3) Autoencoder. They evaluate how well textual information alone can predict quality scores compared to image-based predictions.

Result: Existing models perform significantly worse when using only text for prediction. Chain-of-Thought provides little improvement, but Self-Consistency dramatically reduces the gap between image- and text-conditioned predictions (PLCC/SRCC difference to 0.02/0.03). Autoencoder is less effective but reveals optimization directions.

Conclusion: Self-Consistency is the most effective paradigm for improving textual reasoning in BIQA, significantly narrowing the performance gap between image- and text-based predictions. These findings provide insights for enhancing textual reasoning in BIQA and other high-level vision tasks.

Abstract: Textual reasoning has recently been widely adopted in Blind Image Quality Assessment (BIQA). However, it remains unclear how textual information contributes to quality prediction and to what extent text can represent the score-related image contents. This work addresses these questions from an information-flow perspective by comparing existing BIQA models with three paradigms designed to learn the image-text-score relationship: Chain-of-Thought, Self-Consistency, and Autoencoder. Our experiments show that the score prediction performance of the existing model significantly drops when only textual information is used for prediction. Whereas the Chain-of-Thought paradigm introduces little improvement in BIQA performance, the Self-Consistency paradigm significantly reduces the gap between image- and text-conditioned predictions, narrowing the PLCC/SRCC difference to 0.02/0.03. The Autoencoder-like paradigm is less effective in closing the image-text gap, yet it reveals a direction for further optimization. These findings provide insights into how to improve the textual reasoning for BIQA and high-level vision tasks.

Method: The paper models bias using causal theory, identifying ingredients as confounders. It applies backdoor adjustment for causal intervention, reformulating the conventional retrieval model with an additional term to remove bias. Also proposes a plug-and-play neural module (multi-label ingredient classifier) for debiasing.

Result: Achieves oracle performance of MedR=1 across testing data sizes of 1K, 10K, and 50K on Recipe1M dataset. Reports new state-of-the-art search performances on Recipe1M.

Conclusion: Causal intervention effectively addresses bias in cross-modal recipe-food image retrieval by treating ingredients as confounders, leading to improved retrieval performance through debiasing techniques.

Abstract: This paper addresses the challenges of learning representations for recipes and food images in the cross-modal retrieval problem. As the relationship between a recipe and its cooked dish is cause-and-effect, treating a recipe as a text source describing the visual appearance of a dish for learning representation, as the existing approaches, will create bias misleading image-and-recipe similarity judgment. Specifically, a food image may not equally capture every detail in a recipe, due to factors such as the cooking process, dish presentation, and image-capturing conditions. The current representation learning tends to capture dominant visual-text alignment while overlooking subtle variations that determine retrieval relevance. In this paper, we model such bias in cross-modal representation learning using causal theory. The causal view of this problem suggests ingredients as one of the confounder sources and a simple backdoor adjustment can alleviate the bias. By causal intervention, we reformulate the conventional model for food-to-recipe retrieval with an additional term to remove the potential bias in similarity judgment. Based on this theory-informed formulation, we empirically prove the oracle performance of retrieval on the Recipe1M dataset to be MedR=1 across the testing data sizes of 1K, 10K, and even 50K. We also propose a plug-and-play neural module, which is essentially a multi-label ingredient classifier for debiasing. New state-of-the-art search performances are reported on the Recipe1M dataset.

Method: Reframes gridded monsoon prediction as spatio-temporal computer vision task. Treats multi-variable pre-monsoon atmospheric and oceanic fields as sequence of multi-channel images (video-like input tensor). Uses CNN-based architecture with 85 years of ERA5 reanalysis data for predictors and IMD rainfall data for targets to learn mapping from five-month pre-monsoon period (January-May) to high-resolution gridded rainfall patterns.

Result: Successfully produces distinct forecasts for each of the four monsoon months (June-September) as well as total seasonal average, demonstrating utility for both intra-seasonal and seasonal outlooks.

Conclusion: The deep learning framework effectively addresses the spatial detail limitation of traditional monsoon forecasting, providing high-resolution gridded predictions that can better support regional-level resource management and decision-making.

Abstract: The Indian Summer Monsoon (ISM) is a critical climate phenomenon, fundamentally impacting the agriculture, economy, and water security of over a billion people. Traditional long-range forecasting, whether statistical or dynamical, has predominantly focused on predicting a single, spatially-averaged seasonal value, lacking the spatial detail essential for regional-level resource management. To address this gap, we introduce a novel deep learning framework that reframes gridded monsoon prediction as a spatio-temporal computer vision task. We treat multi-variable, pre-monsoon atmospheric and oceanic fields as a sequence of multi-channel images, effectively creating a video-like input tensor. Using 85 years of ERA5 reanalysis data for predictors and IMD rainfall data for targets, we employ a Convolutional Neural Network (CNN)-based architecture to learn the complex mapping from the five-month pre-monsoon period (January-May) to a high-resolution gridded rainfall pattern for the subsequent monsoon season. Our framework successfully produces distinct forecasts for each of the four monsoon months (June-September) as well as the total seasonal average, demonstrating its utility for both intra-seasonal and seasonal outlooks.

Method: Created FCMBench-V1.0 using a closed synthesis-capture pipeline: manually synthesized document templates with virtual content and captured scenario-aware images in-house to ensure privacy compliance and avoid web-sourced data leakage. The benchmark includes 3 perception tasks, 4 credit-specific reasoning tasks, and 10 real-world acquisition artifact types for robustness testing.

Result: Evaluated 23 state-of-the-art VLMs from 14 organizations. Gemini 3 Pro achieved best commercial model score (64.61 F1), Qwen3-VL-235B best open-source (57.27), and their financial credit-specific model Qfin-VL-Instruct achieved top overall score (64.92). Robustness evaluations showed significant performance drops under acquisition artifacts even for top models.

Conclusion: FCMBench effectively discriminates performance disparities and robustness across modern VLMs, revealing that current models struggle with financial credit document understanding and real-world robustness, highlighting the need for domain-specific benchmarks and specialized models for financial applications.

Abstract: As multimodal AI becomes widely used for credit risk assessment and document review, a domain-specific benchmark is urgently needed that (1) reflects documents and workflows specific to financial credit applications, (2) includes credit-specific understanding and real-world robustness, and (3) preserves privacy compliance without sacrificing practical utility. Here, we introduce FCMBench-V1.0 – a large-scale financial credit multimodal benchmark for real-world applications, covering 18 core certificate types, with 4,043 privacy-compliant images and 8,446 QA samples. The FCMBench evaluation framework consists of three dimensions: Perception, Reasoning, and Robustness, including 3 foundational perception tasks, 4 credit-specific reasoning tasks that require decision-oriented understanding of visual evidence, and 10 real-world acquisition artifact types for robustness stress testing. To reconcile compliance with realism, we construct all samples via a closed synthesis-capture pipeline: we manually synthesize document templates with virtual content and capture scenario-aware images in-house. This design also mitigates pre-training data leakage by avoiding web-sourced or publicly released images. FCMBench can effectively discriminate performance disparities and robustness across modern vision-language models. Extensive experiments were conducted on 23 state-of-the-art vision-language models (VLMs) from 14 top AI companies and research institutes. Among them, Gemini 3 Pro achieves the best F1(%) score as a commercial model (64.61), Qwen3-VL-235B achieves the best score as an open-source baseline (57.27), and our financial credit-specific model, Qfin-VL-Instruct, achieves the top overall score (64.92). Robustness evaluations show that even top-performing models suffer noticeable performance drops under acquisition artifacts.

Method: Two INR-based frameworks: 1) Inter-B-scan interpolation incorporating en-face modality information to retain structures between B-scans, and 2) Resolution-agnostic retinal atlas using generalizable INRs trained on population data for shape representation.

Result: The frameworks enable dense 3D analysis of retinal OCT volumes, are resolution-independent, and can handle images with large B-scan distances for volumetric evaluation of retinal structures and pathologies.

Conclusion: INR-based approaches overcome limitations of anisotropic OCT data, allowing volumetric analysis of sparsely scanned retinas and opening possibilities for evaluating retinal structures across different imaging protocols.

Abstract: Routine clinical imaging of the retina using optical coherence tomography (OCT) is performed with large slice spacing, resulting in highly anisotropic images and a sparsely scanned retina. Most learning-based methods circumvent the problems arising from the anisotropy by using 2D approaches rather than performing volumetric analyses. These approaches inherently bear the risk of generating inconsistent results for neighboring B-scans. For example, 2D retinal layer segmentations can have irregular surfaces in 3D. Furthermore, the typically used convolutional neural networks are bound to the resolution of the training data, which prevents their usage for images acquired with a different imaging protocol. Implicit neural representations (INRs) have recently emerged as a tool to store voxelized data as a continuous representation. Using coordinates as input, INRs are resolution-agnostic, which allows them to be applied to anisotropic data. In this paper, we propose two frameworks that make use of this characteristic of INRs for dense 3D analyses of retinal OCT volumes. 1) We perform inter-B-scan interpolation by incorporating additional information from en-face modalities, that help retain relevant structures between B-scans. 2) We create a resolution-agnostic retinal atlas that enables general analysis without strict requirements for the data. Both methods leverage generalizable INRs, improving retinal shape representation through population-based training and allowing predictions for unseen cases. Our resolution-independent frameworks facilitate the analysis of OCT images with large B-scan distances, opening up possibilities for the volumetric evaluation of retinal structures and pathologies.

Method: Two-stage pre-training: (1) distillation of dense 2D features from visual encoders (DINOv2) into 3D patches, and (2) alignment of patch embeddings with part-level text embeddings through multi-positive contrastive objective.

Result: Achieves zero-shot 3D part segmentation with fast single-pass inference, significantly outperforming previous rendering-based and feed-forward approaches across several 3D part segmentation benchmarks.

Conclusion: The encoder-only 3D model enables efficient, geometry-aware part-level reasoning without expensive multi-view rendering, bridging the gap between 2D vision-language models and 3D understanding.

Abstract: Current foundation models for 3D shapes excel at global tasks (retrieval, classification) but transfer poorly to local part-level reasoning. Recent approaches leverage vision and language foundation models to directly solve dense tasks through multi-view renderings and text queries. While promising, these pipelines require expensive inference over multiple renderings, depend heavily on large language-model (LLM) prompt engineering for captions, and fail to exploit the inherent 3D geometry of shapes. We address this gap by introducing an encoder-only 3D model that produces language-aligned patch-level features directly from point clouds. Our pre-training approach builds on existing data engines that generate part-annotated 3D shapes by pairing multi-view SAM regions with VLM captioning. Using this data, we train a point cloud transformer encoder in two stages: (1) distillation of dense 2D features from visual encoders such as DINOv2 into 3D patches, and (2) alignment of these patch embeddings with part-level text embeddings through a multi-positive contrastive objective. Our 3D encoder achieves zero-shot 3D part segmentation with fast single-pass inference without any test-time multi-view rendering, while significantly outperforming previous rendering-based and feed-forward approaches across several 3D part segmentation benchmarks. Project website: https://souhail-hadgi.github.io/patchalign3dsite/

Method: 1) Cross-modal differentiated quantization framework for VLMs with tailored strategies for different components. 2) Scene-aware vectorized memory multi-agent system using perception-memory-reasoning workflows to provide contextual environmental information.

Result: Quantization reduces memory from 38GB to 11.3GB with only 2.05% performance drop on MMBench. OCR-VQA accuracy maintains 63.7 (original: 64.9). Multi-agent system achieves 2.83-3.52s latency to initial speech output, outperforming smaller models with equivalent memory.

Conclusion: The research advances computational efficiency and assistive technology by enabling efficient VLM deployment for comprehensive scene perception, text recognition, and navigation assistance for visually impaired individuals.

Abstract: Visually impaired individuals face significant challenges in environmental perception. Traditional assistive technologies often lack adaptive intelligence, focusing on individual components rather than integrated systems. While Vision-Language Models (VLMs) offer a promising path to richer, integrated understanding, their deployment is severely limited by substantial computational requirements, demanding dozens of gigabytes of memory. To address these gaps in computational efficiency and integrated design, this study proposes a dual technological innovation framework: a cross-modal differentiated quantization framework for VLMs and a scene-aware vectorized memory multi-agent system. The quantization framework implements differentiated strategies, reducing memory from 38GB to 11.3GB. The multi-agent system uses vectorized memory and perception-memory-reasoning workflows to provide environmental information beyond the current view, achieving 2.83-3.52s latency to initial speech output. Experiments show the quantized 19B-parameter model only experiences a 2.05% performance drop on MMBench and maintains 63.7 accuracy on OCR-VQA (original: 64.9), outperforming smaller models with equivalent memory. This research advances computational efficiency and assistive technology, offering comprehensive assistance in scene perception, text recognition, and navigation.

Method: Reformulate volumetric CT data as sequential video streams. Benchmark YOLO architectures (v8, v10, v11, v12 Nano) for slice-level detection. Use ByteTrack algorithm for anatomical consistency across z-axis. Implement hybrid inference strategy and spatiotemporal consistency filter to address tracker initialization lag and reduce false positives.

Result: On independent test data, the framework increased detection precision from 0.703 to 0.779 compared to baseline 2D detector while maintaining high sensitivity. Provides 3D contextual reasoning at fraction of computational cost.

Conclusion: The video-viewpoint paradigm offers scalable solution for real-time patient prioritization in resource-constrained environments by approximating 3D context efficiently, balancing detection accuracy with computational efficiency for edge deployment.

Abstract: Automated analysis of volumetric medical imaging on edge devices is severely constrained by the high memory and computational demands of 3D Convolutional Neural Networks (CNNs). This paper develops a lightweight computer vision framework that reconciles the efficiency of 2D detection with the necessity of 3D context by reformulating volumetric Computer Tomography (CT) data as sequential video streams. This video-viewpoint paradigm is applied to the time-sensitive task of Intracranial Hemorrhage (ICH) detection using the Hemorica dataset. To ensure operational efficiency, we benchmarked multiple generations of the YOLO architecture (v8, v10, v11 and v12) in their Nano configurations, selecting the version with the highest mAP@50 to serve as the slice-level backbone. A ByteTrack algorithm is then introduced to enforce anatomical consistency across the $z$-axis. To address the initialization lag inherent in video trackers, a hybrid inference strategy and a spatiotemporal consistency filter are proposed to distinguish true pathology from transient prediction noise. Experimental results on independent test data demonstrate that the proposed framework serves as a rigorous temporal validator, increasing detection Precision from 0.703 to 0.779 compared to the baseline 2D detector, while maintaining high sensitivity. By approximating 3D contextual reasoning at a fraction of the computational cost, this method provides a scalable solution for real-time patient prioritization in resource-constrained environments, such as mobile stroke units and IoT-enabled remote clinics.

Method: Created benchmark using movie recap videos from YouTube that summarize films through synchronized visual (recap video) and textual (recap summary) modalities. Generated ≈8.2K QA pairs aligned with movie subtitles, providing necessary “facts” for reference-free evaluation.

Result: Evaluated 7 state-of-the-art MLLMs, finding: 1) visual-only questions most challenging, 2) models default to textual inputs when available, 3) extracting accurate factual information from video remains difficult, 4) proprietary and open-source models perform comparably on video-dependent questions.

Conclusion: MovieRecapsQA is the first open-ended VideoQA benchmark with explicit textual context for evaluation, enabling fine-grained analysis through multiple video lengths and question categorizations, revealing significant challenges in multimodal reasoning for current models.

Abstract: Understanding real-world videos such as movies requires integrating visual and dialogue cues to answer complex questions. Yet existing VideoQA benchmarks struggle to capture this multimodal reasoning and are largely not open-ended, given the difficulty of evaluating free-form answers. In this paper, we introduce a novel open-ended multi-modal VideoQA benchmark, MovieRecapsQA created using movie recap videos–a distinctive type of YouTube content that summarizes a film by presenting its key events through synchronized visual (recap video) and textual (recap summary) modalities. Using the recap summary, we generate $\approx 8.2$ K question-answer (QA) pairs (aligned with movie-subtitles) and provide the necessary “facts” needed to verify an answer in a reference-free manner. To our knowledge, this is the first open-ended VideoQA benchmark that supplies explicit textual context of the input (video and/or text); which we use for evaluation. Our benchmark provides videos of multiple lengths (i.e., recap-segments, movie-segments) and categorizations of questions (by modality and type) to enable fine-grained analysis. We evaluate the performance of seven state-of-the-art MLLMs using our benchmark and observe that: 1) visual-only questions remain the most challenging; 2) models default to textual inputs whenever available; 3) extracting factually accurate information from video content is still difficult for all models; and 4) proprietary and open-source models perform comparably on video-dependent questions.

Method: Uses Vision Mamba network to extract feature maps that clearly describe boundaries between tampered and untouched regions, plus a novel Guided Graph Neural Network (G-GNN) module to amplify distinction between manipulated and authentic pixels.

Result: Achieved higher inference accuracy compared to other state-of-the-art methods in evaluation.

Conclusion: Proposed solution effectively localizes manipulations in both shallow- and deep-fake images using Vision Mamba and G-GNN, demonstrating superior performance over existing methods.

Abstract: Image manipulation localization is a critical research task, given that forged images may have a significant societal impact of various aspects. Such image manipulations can be produced using traditional image editing tools (known as “shallowfakes”) or advanced artificial intelligence techniques (“deepfakes”). While numerous studies have focused on image manipulation localization on either shallowfake images or deepfake videos, few approaches address both cases. In this paper, we explore the feasibility of using a deep learning network to localize manipulations in both shallow- and deep-fake images, and proposed a solution for such purpose. To precisely differentiate between authentic and manipulated pixels, we leverage the Vision Mamba network to extract feature maps that clearly describe the boundaries between tampered and untouched regions. To further enhance this separation, we propose a novel Guided Graph Neural Network (G-GNN) module that amplifies the distinction between manipulated and authentic pixels. Our evaluation results show that our proposed method achieved higher inference accuracy compared to other state-of-the-art methods.

Method: Adapts a general video diffusion model by training on two objectives: temporal bridging and motion conditioning. During inference: (1) uses input image as both first- and last-frame condition to enforce seamless loops, (2) conditions on static tracks to maintain static backgrounds, and (3) uses user-specified motion paths for target objects to control animation trajectory and timing.

Result: Produces high-quality, complex cinemagraphs that align with user intent, outperforming existing approaches. The method enables cinemagraph generation for general scenes with flexible and intuitive controls, without requiring cinemagraph training data.

Conclusion: DreamLoop is the first method to enable cinemagraph generation for general scenes with flexible and intuitive controls, successfully adapting video diffusion models to meet cinemagraph constraints through specialized training objectives and inference techniques.

Abstract: Cinemagraphs, which combine static photographs with selective, looping motion, offer unique artistic appeal. Generating them from a single photograph in a controllable manner is particularly challenging. Existing image-animation techniques are restricted to simple, low-frequency motions and operate only in narrow domains with repetitive textures like water and smoke. In contrast, large-scale video diffusion models are not tailored for cinemagraph constraints and lack the specialized data required to generate seamless, controlled loops. We present DreamLoop, a controllable video synthesis framework dedicated to generating cinemagraphs from a single photo without requiring any cinemagraph training data. Our key idea is to adapt a general video diffusion model by training it on two objectives: temporal bridging and motion conditioning. This strategy enables flexible cinemagraph generation. During inference, by using the input image as both the first- and last- frame condition, we enforce a seamless loop. By conditioning on static tracks, we maintain a static background. Finally, by providing a user-specified motion path for a target object, our method provides intuitive control over the animation’s trajectory and timing. To our knowledge, DreamLoop is the first method to enable cinemagraph generation for general scenes with flexible and intuitive controls. We demonstrate that our method produces high-quality, complex cinemagraphs that align with user intent, outperforming existing approaches.

Method: Proposes G-channel Removed Reconstruction Error (GRRE) - remove green channel from images, reconstruct it, and analyze reconstruction error differences between real and AI-generated images.

Result: GRRE achieves high detection accuracy across multiple generative models (including unseen ones), maintains robustness against perturbations/post-processing, and shows superior cross-model generalization compared to existing methods.

Conclusion: Channel-removal-based reconstruction is a powerful forensic approach for detecting AI-generated images, offering strong generalization capabilities essential for safeguarding image authenticity in the generative AI era.

Abstract: The rapid progress of generative models, particularly diffusion models and GANs, has greatly increased the difficulty of distinguishing synthetic images from real ones. Although numerous detection methods have been proposed, their accuracy often degrades when applied to images generated by novel or unseen generative models, highlighting the challenge of achieving strong generalization. To address this challenge, we introduce a novel detection paradigm based on channel removal reconstruction. Specifically, we observe that when the green (G) channel is removed from real images and reconstructed, the resulting reconstruction errors differ significantly from those of AI-generated images. Building upon this insight, we propose G-channel Removed Reconstruction Error (GRRE), a simple yet effective method that exploits this discrepancy for robust AI-generated image detection. Extensive experiments demonstrate that GRRE consistently achieves high detection accuracy across multiple generative models, including those unseen during training. Compared with existing approaches, GRRE not only maintains strong robustness against various perturbations and post-processing operations but also exhibits superior cross-model generalization. These results highlight the potential of channel-removal-based reconstruction as a powerful forensic tool for safeguarding image authenticity in the era of generative AI.

Method: Uses morphology-parameterized articulated 3D Gaussian splatting model combined with dense semantic correspondences to jointly optimize shape and pose, alleviating shape-pose ambiguities.

Result: Demonstrates superior motion accuracy, efficiency, and visual coherence compared to existing methods, advancing motion transfer for diverse object categories and casual video scenarios.

Conclusion: CAMO provides a category-agnostic framework that enables visually faithful motion transfer without predefined templates or explicit 3D supervision, overcoming traditional limitations.

Abstract: Motion transfer from 2D videos to 3D assets is a challenging problem, due to inherent pose ambiguities and diverse object shapes, often requiring category-specific parametric templates. We propose CAMO, a category-agnostic framework that transfers motion to diverse target meshes directly from monocular 2D videos without relying on predefined templates or explicit 3D supervision. The core of CAMO is a morphology-parameterized articulated 3D Gaussian splatting model combined with dense semantic correspondences to jointly adapt shape and pose through optimization. This approach effectively alleviates shape-pose ambiguities, enabling visually faithful motion transfer for diverse categories. Experimental results demonstrate superior motion accuracy, efficiency, and visual coherence compared to existing methods, significantly advancing motion transfer in varied object categories and casual video scenarios.

Method: Uses Grounded SAM2 to identify foreground objects and compute per-image foreground occupancy, derives category-wise patch decision thresholds, and implements dynamic patch selection that either selects the most informative patch or resizes the full image when foreground dominates.

Result: Extensive experiments on multiple benchmarks show consistent improvement in distillation performance over existing approaches, producing more informative and representative distilled datasets with enhanced robustness across different architectures and image compositions.

Conclusion: The proposed foreground-aware dataset distillation method effectively preserves key information about main objects while reducing redundant background content, outperforming existing methods in creating compact synthetic datasets.

Abstract: In this paper, we propose a foreground-aware dataset distillation method that enhances patch selection in a content-adaptive manner. With the rising computational cost of training large-scale deep models, dataset distillation has emerged as a promising approach for constructing compact synthetic datasets that retain the knowledge of their large original counterparts. However, traditional optimization-based methods often suffer from high computational overhead, memory constraints, and the generation of unrealistic, noise-like images with limited architectural generalization. Recent non-optimization methods alleviate some of these issues by constructing distilled data from real image patches, but the used rigid patch selection strategies can still discard critical information about the main objects. To solve this problem, we first leverage Grounded SAM2 to identify foreground objects and compute per-image foreground occupancy, from which we derive a category-wise patch decision threshold. Guided by these thresholds, we design a dynamic patch selection strategy that, for each image, either selects the most informative patch from multiple candidates or directly resizes the full image when the foreground dominates. This dual-path mechanism preserves more key information about the main objects while reducing redundant background content. Extensive experiments on multiple benchmarks show that the proposed method consistently improves distillation performance over existing approaches, producing more informative and representative distilled datasets and enhancing robustness across different architectures and image compositions.

Method: Proposes a homography-guided pose estimator that constructs input pairs satisfying homography constraints by projecting ground-view features into BEV domain and enforcing semantic alignment with map features. Uses homography relationships to guide feature fusion and restrict pose outputs to valid feasible regions.

Result: Significantly outperforms existing state-of-the-art visual localization methods on nuScenes dataset. The approach improves training efficiency and localization accuracy compared to attention-based fusion and direct 3-DoF pose regression methods.

Conclusion: First work to unify BEV semantic reasoning with homography learning for image-to-map localization. The framework naturally supports cross-resolution inputs and demonstrates superior performance, with code and models to be publicly released.

Abstract: Visual localization on standard-definition (SD) maps has emerged as a promising low-cost and scalable solution for autonomous driving. However, existing regression-based approaches often overlook inherent geometric priors, resulting in suboptimal training efficiency and limited localization accuracy. In this paper, we propose a novel homography-guided pose estimator network for fine-grained visual localization between multi-view images and standard-definition (SD) maps. We construct input pairs that satisfy a homography constraint by projecting ground-view features into the BEV domain and enforcing semantic alignment with map features. Then we leverage homography relationships to guide feature fusion and restrict the pose outputs to a valid feasible region, which significantly improves training efficiency and localization accuracy compared to prior methods relying on attention-based fusion and direct 3-DoF pose regression. To the best of our knowledge, this is the first work to unify BEV semantic reasoning with homography learning for image-to-map localization. Furthermore, by explicitly modeling homography transformations, the proposed framework naturally supports cross-resolution inputs, enhancing model flexibility. Extensive experiments on the nuScenes dataset demonstrate that our approach significantly outperforms existing state-of-the-art visual localization methods. Code and pretrained models will be publicly released to foster future research.

Method: 1) Created PET-Bench benchmark with 52,308 QA pairs from 9,732 multi-site PET studies. 2) Evaluated 19 SOTA MLLMs, identifying “CoT hallucination trap.” 3) Proposed Atomic Visual Alignment (AVA) - fine-tuning strategy that forces models to master low-level functional perception before high-level reasoning.

Result: 1) MLLMs show critical functional perception gap in PET imaging. 2) Standard CoT prompting causes clinically fluent but factually ungrounded diagnoses (hallucinations). 3) AVA transforms CoT from hallucination source to robust inference tool, improving diagnostic accuracy by up to 14.83%.

Conclusion: Functional imaging requires different perception capabilities than anatomical imaging. AVA effectively bridges the perception gap by enforcing visual grounding before reasoning, making MLLMs safer and more accurate for PET diagnosis. The PET-Bench benchmark enables future research in functional imaging.

Abstract: While Multimodal Large Language Models (MLLMs) have demonstrated remarkable proficiency in tasks such as abnormality detection and report generation for anatomical modalities, their capability in functional imaging remains largely unexplored. In this work, we identify and quantify a fundamental functional perception gap: the inability of current vision encoders to decode functional tracer biodistribution independent of morphological priors. Identifying Positron Emission Tomography (PET) as the quintessential modality to investigate this disconnect, we introduce PET-Bench, the first large-scale functional imaging benchmark comprising 52,308 hierarchical QA pairs from 9,732 multi-site, multi-tracer PET studies. Extensive evaluation of 19 state-of-the-art MLLMs reveals a critical safety hazard termed the Chain-of-Thought (CoT) hallucination trap. We observe that standard CoT prompting, widely considered to enhance reasoning, paradoxically decouples linguistic generation from visual evidence in PET, producing clinically fluent but factually ungrounded diagnoses. To resolve this, we propose Atomic Visual Alignment (AVA), a simple fine-tuning strategy that enforces the mastery of low-level functional perception prior to high-level diagnostic reasoning. Our results demonstrate that AVA effectively bridges the perception gap, transforming CoT from a source of hallucination into a robust inference tool and improving diagnostic accuracy by up to 14.83%. Code and data are available at https://github.com/yezanting/PET-Bench.

Method: Proposes D³R-DETR with Dual-Domain Density Refinement that fuses spatial and frequency domain information to refine low-level feature maps, uses rich details to predict accurate object density maps, and guides precise localization of tiny objects.

Result: Extensive experiments on AI-TOD-v2 dataset demonstrate that D³R-DETR outperforms existing state-of-the-art detectors for tiny object detection.

Conclusion: The proposed D³R-DETR with Dual-Domain Density Refinement effectively addresses the challenges of tiny object detection in remote sensing by improving feature refinement and density prediction for better localization accuracy.

Abstract: Detecting tiny objects plays a vital role in remote sensing intelligent interpretation, as these objects often carry critical information for downstream applications. However, due to the extremely limited pixel information and significant variations in object density, mainstream Transformer-based detectors often suffer from slow convergence and inaccurate query-object matching. To address these challenges, we propose D$^3$R-DETR, a novel DETR-based detector with Dual-Domain Density Refinement. By fusing spatial and frequency domain information, our method refines low-level feature maps and utilizes their rich details to predict more accurate object density map, thereby guiding the model to precisely localize tiny objects. Extensive experiments on the AI-TOD-v2 dataset demonstrate that D$^3$R-DETR outperforms existing state-of-the-art detectors for tiny object detection.

Method: BUFFER-X uses: (a) geometric bootstrapping for automatic hyperparameter estimation, (b) distribution-aware farthest point sampling to replace learned detectors, and (c) patch-level coordinate normalization for scale consistency. It employs hierarchical multi-scale matching across local, middle, and global receptive fields. BUFFER-X-Lite adds early exit strategies and fast pose solvers for 43% faster computation.

Result: The approach generalizes effectively without manual tuning or prior knowledge of test domains across 12 datasets spanning object-scale, indoor, and outdoor scenes, including cross-sensor registration between heterogeneous LiDAR configurations.

Conclusion: BUFFER-X provides a robust, training-free solution for point cloud registration that achieves zero-shot generalization across diverse environments, addressing key limitations of existing methods while maintaining accuracy and offering efficient variants for time-critical applications.

Abstract: Some deep learning-based point cloud registration methods struggle with zero-shot generalization, often requiring dataset-specific hyperparameter tuning or retraining for new environments. We identify three critical limitations: (a) fixed user-defined parameters (e.g., voxel size, search radius) that fail to generalize across varying scales, (b) learned keypoint detectors exhibit poor cross-domain transferability, and (c) absolute coordinates amplify scale mismatches between datasets. To address these three issues, we present BUFFER-X, a training-free registration framework that achieves zero-shot generalization through: (a) geometric bootstrapping for automatic hyperparameter estimation, (b) distribution-aware farthest point sampling to replace learned detectors, and (c) patch-level coordinate normalization to ensure scale consistency. Our approach employs hierarchical multi-scale matching to extract correspondences across local, middle, and global receptive fields, enabling robust registration in diverse environments. For efficiency-critical applications, we introduce BUFFER-X-Lite, which reduces total computation time by 43% (relative to BUFFER-X) through early exit strategies and fast pose solvers while preserving accuracy. We evaluate on a comprehensive benchmark comprising 12 datasets spanning object-scale, indoor, and outdoor scenes, including cross-sensor registration between heterogeneous LiDAR configurations. Results demonstrate that our approach generalizes effectively without manual tuning or prior knowledge of test domains. Code: https://github.com/MIT-SPARK/BUFFER-X.

Method: 1) Use DINOv3 as visual encoder for dense features. 2) Design Simple Depth Transformer (SDT) decoder with single-path feature fusion and upsampling (85-89% parameter reduction vs DPT). 3) Quality-based filtering strategy to remove harmful training samples.

Result: Framework achieves higher accuracy than DPT on five benchmarks while being more parameter-efficient. Demonstrates importance of balancing model design and data quality.

Conclusion: Proposed lightweight, data-centric framework enables efficient and generalizable zero-shot depth estimation through improved model architecture and data quality management.

Abstract: Monocular depth estimation aims to recover the depth information of 3D scenes from 2D images. Recent work has made significant progress, but its reliance on large-scale datasets and complex decoders has limited its efficiency and generalization ability. In this paper, we propose a lightweight and data-centric framework for zero-shot monocular depth estimation. We first adopt DINOv3 as the visual encoder to obtain high-quality dense features. Secondly, to address the inherent drawbacks of the complex structure of the DPT, we design the Simple Depth Transformer (SDT), a compact transformer-based decoder. Compared to the DPT, it uses a single-path feature fusion and upsampling process to reduce the computational overhead of cross-scale feature fusion, achieving higher accuracy while reducing the number of parameters by approximately 85%-89%. Furthermore, we propose a quality-based filtering strategy to filter out harmful samples, thereby reducing dataset size while improving overall training quality. Extensive experiments on five benchmarks demonstrate that our framework surpasses the DPT in accuracy. This work highlights the importance of balancing model design and data quality for achieving efficient and generalizable zero-shot depth estimation. Code: https://github.com/AIGeeksGroup/AnyDepth. Website: https://aigeeksgroup.github.io/AnyDepth.

Method: 1) MLLM-based Image Quality Assessment for overall evaluation using cross-modal understanding; 2) Region awareness and task recognition pipeline with semantic cross-attention and degradation-aware module; 3) Internal clue reuse mechanism for self-supervised fine detail restoration.

Result: ClearAIR achieves superior performance across diverse synthetic and real-world datasets, demonstrating effective handling of complex composite degradations without oversmoothing or artifacts.

Conclusion: The hierarchical coarse-to-fine restoration strategy inspired by Human Visual Perception, combined with MLLM-based assessment and internal clue reuse, provides an effective framework for All-in-One Image Restoration that outperforms existing approaches.

Abstract: All-in-One Image Restoration (AiOIR) has advanced significantly, offering promising solutions for complex real-world degradations. However, most existing approaches rely heavily on degradation-specific representations, often resulting in oversmoothing and artifacts. To address this, we propose ClearAIR, a novel AiOIR framework inspired by Human Visual Perception (HVP) and designed with a hierarchical, coarse-to-fine restoration strategy. First, leveraging the global priority of early HVP, we employ a Multimodal Large Language Model (MLLM)-based Image Quality Assessment (IQA) model for overall evaluation. Unlike conventional IQA, our method integrates cross-modal understanding to more accurately characterize complex, composite degradations. Building upon this overall assessment, we then introduce a region awareness and task recognition pipeline. A semantic cross-attention, leveraging semantic guidance unit, first produces coarse semantic prompts. Guided by this regional context, a degradation-aware module implicitly captures region-specific degradation characteristics, enabling more precise local restoration. Finally, to recover fine details, we propose an internal clue reuse mechanism. It operates in a self-supervised manner to mine and leverage the intrinsic information of the image itself, substantially enhancing detail restoration. Experimental results show that ClearAIR achieves superior performance across diverse synthetic and real-world datasets.

Method: Proposes AbductiveMLLM with two synergistic components: REASONER (verbal domain) that explores explanations using LLM and prunes visually incongruent hypotheses, and IMAGINER (pictorial domain) that uses diffusion models to “imagine” visual scenes corresponding to verbal explanations. Both components are trained end-to-end.

Result: Achieves state-of-the-art performance on standard VAR benchmarks, consistently outperforming traditional solutions and advanced MLLMs.

Conclusion: Mimicking human dual-mode cognition (verbal and pictorial abduction) effectively enhances MLLMs’ visual abductive reasoning capabilities, bridging the gap between AI systems and human-level abductive inference.

Abstract: Visual abductive reasoning (VAR) is a challenging task that requires AI systems to infer the most likely explanation for incomplete visual observations. While recent MLLMs develop strong general-purpose multimodal reasoning capabilities, they fall short in abductive inference, as compared to human beings. To bridge this gap, we draw inspiration from the interplay between verbal and pictorial abduction in human cognition, and propose to strengthen abduction of MLLMs by mimicking such dual-mode behavior. Concretely, we introduce AbductiveMLLM comprising of two synergistic components: REASONER and IMAGINER. The REASONER operates in the verbal domain. It first explores a broad space of possible explanations using a blind LLM and then prunes visually incongruent hypotheses based on cross-modal causal alignment. The remaining hypotheses are introduced into the MLLM as targeted priors, steering its reasoning toward causally coherent explanations. The IMAGINER, on the other hand, further guides MLLMs by emulating human-like pictorial thinking. It conditions a text-to-image diffusion model on both the input video and the REASONER’s output embeddings to “imagine” plausible visual scenes that correspond to verbal explanation, thereby enriching MLLMs’ contextual grounding. The two components are trained jointly in an end-to-end manner. Experiments on standard VAR benchmarks show that AbductiveMLLM achieves state-of-the-art performance, consistently outperforming traditional solutions and advanced MLLMs.

Method: Proposes EarthVL framework with EarthVLSet dataset (10.9k sub-meter RS images, land-cover masks, 761.5k text pairs) and EarthVLNet network. EarthVLNet uses progressive approach: 1) land-cover segmentation for object semantics, 2) object-aware LLM for relational reasoning and VQA, with numerical difference loss for optimization.

Result: Superior performance on three benchmarks: semantic segmentation, multiple-choice VQA, and open-ended VQA. Key findings: segmentation features enhance VQA even cross-dataset; multiple-choice tasks more sensitive to vision encoder; open-ended tasks need advanced vision and language models.

Conclusion: The framework advances Earth vision by connecting image-mask-text for comprehensive scene understanding, providing benchmark for geographical applications. Future directions identified for segmentation-enhanced VQA, encoder sensitivity, and model requirements for different task types.

Abstract: Earth vision has achieved milestones in geospatial object recognition but lacks exploration in object-relational reasoning, limiting comprehensive scene understanding. To address this, a progressive Earth vision-language understanding and generation framework is proposed, including a multi-task dataset (EarthVLSet) and a semantic-guided network (EarthVLNet). Focusing on city planning applications, EarthVLSet includes 10.9k sub-meter resolution remote sensing images, land-cover masks, and 761.5k textual pairs involving both multiple-choice and open-ended visual question answering (VQA) tasks. In an object-centric way, EarthVLNet is proposed to progressively achieve semantic segmentation, relational reasoning, and comprehensive understanding. The first stage involves land-cover segmentation to generate object semantics for VQA guidance. Guided by pixel-wise semantics, the object awareness based large language model (LLM) performs relational reasoning and knowledge summarization to generate the required answers. As for optimization, the numerical difference loss is proposed to dynamically add difference penalties, addressing the various objects’ statistics. Three benchmarks, including semantic segmentation, multiple-choice, and open-ended VQA demonstrated the superiorities of EarthVLNet, yielding three future directions: 1) segmentation features consistently enhance VQA performance even in cross-dataset scenarios; 2) multiple-choice tasks show greater sensitivity to the vision encoder than to the language decoder; and 3) open-ended tasks necessitate advanced vision encoders and language decoders for an optimal performance. We believe this dataset and method will provide a beneficial benchmark that connects ‘‘image-mask-text’’, advancing geographical applications for Earth vision.

Method: Built on vanilla Image-to-Video model with Low-Rank Adaptation (LoRA) using token-specific up matrices to reduce confusion among different condition tokens, plus a well-designed data curation pipeline for high-quality paired video data.

Result: DreamStyle demonstrates competence in all three video stylization tasks and outperforms competitors in both style consistency and video quality in qualitative and quantitative evaluations.

Conclusion: DreamStyle provides a unified solution for flexible video stylization across multiple condition types while addressing key challenges of style consistency and temporal stability.

Abstract: Video stylization, an important downstream task of video generation models, has not yet been thoroughly explored. Its input style conditions typically include text, style image, and stylized first frame. Each condition has a characteristic advantage: text is more flexible, style image provides a more accurate visual anchor, and stylized first frame makes long-video stylization feasible. However, existing methods are largely confined to a single type of style condition, which limits their scope of application. Additionally, their lack of high-quality datasets leads to style inconsistency and temporal flicker. To address these limitations, we introduce DreamStyle, a unified framework for video stylization, supporting (1) text-guided, (2) style-image-guided, and (3) first-frame-guided video stylization, accompanied by a well-designed data curation pipeline to acquire high-quality paired video data. DreamStyle is built on a vanilla Image-to-Video (I2V) model and trained using a Low-Rank Adaptation (LoRA) with token-specific up matrices that reduces the confusion among different condition tokens. Both qualitative and quantitative evaluations demonstrate that DreamStyle is competent in all three video stylization tasks, and outperforms the competitors in style consistency and video quality.

Method: Textile IR uses a seven-layer Verification Ladder from cheap syntactic checks (pattern closure, seam compatibility) to expensive physics validation (drape simulation, stress analysis). It formalizes fashion engineering as constraint satisfaction over three domains and uses a scene-graph representation that enables AI systems to manipulate garments as structured programs.

Result: The framework enables bidirectional feedback: simulation failures suggest pattern modifications; material substitutions update sustainability estimates in real time; uncertainty propagates across the pipeline with explicit confidence bounds. It makes engineering constraints perceptible, manipulable, and immediately consequential.

Conclusion: Textile IR provides semantic glue for integrating fashion design tools, enabling designers to navigate tradeoffs simultaneously rather than discovering conflicts after costly physical prototyping. The paper proposes six research priorities and discusses deployment considerations for fashion SMEs.

Abstract: We introduce Textile IR, a bidirectional intermediate representation that connects manufacturing-valid CAD, physics-based simulation, and lifecycle assessment for fashion design. Unlike existing siloed tools where pattern software guarantees sewable outputs but understands nothing about drape, and physics simulation predicts behaviour but cannot automatically fix patterns, Textile IR provides the semantic glue for integration through a seven-layer Verification Ladder – from cheap syntactic checks (pattern closure, seam compatibility) to expensive physics validation (drape simulation, stress analysis). The architecture enables bidirectional feedback: simulation failures suggest pattern modifications; material substitutions update sustainability estimates in real time; uncertainty propagates across the pipeline with explicit confidence bounds. We formalise fashion engineering as constraint satisfaction over three domains and demonstrate how Textile IR’s scene-graph representation enables AI systems to manipulate garments as structured programs rather than pixel arrays. The framework addresses the compound uncertainty problem: when measurement errors in material testing, simulation approximations, and LCA database gaps combine, sustainability claims become unreliable without explicit uncertainty tracking. We propose six research priorities and discuss deployment considerations for fashion SMEs where integrated workflows reduce specialised engineering requirements. Key contribution: a formal representation that makes engineering constraints perceptible, manipulable, and immediately consequential – enabling designers to navigate sustainability, manufacturability, and aesthetic tradeoffs simultaneously rather than discovering conflicts after costly physical prototyping.

Method: Builds on ViT encoder and DPT head, adding temporal layers with efficient cross-attention to integrate information from keyframes across entire video sequences.

Result: Improved temporal consistency, competitive SOTA performance, and 2x faster processing in real-world scenarios on multiple benchmark datasets.

Conclusion: StableDPT effectively adapts image-based depth models for video with temporal stability, global context capture, and efficient processing of arbitrary-length videos.

Abstract: Applying single image Monocular Depth Estimation (MDE) models to video sequences introduces significant temporal instability and flickering artifacts. We propose a novel approach that adapts any state-of-the-art image-based (depth) estimation model for video processing by integrating a new temporal module - trainable on a single GPU in a few days. Our architecture StableDPT builds upon an off-the-shelf Vision Transformer (ViT) encoder and enhances the Dense Prediction Transformer (DPT) head. The core of our contribution lies in the temporal layers within the head, which use an efficient cross-attention mechanism to integrate information from keyframes sampled across the entire video sequence. This allows the model to capture global context and inter-frame relationships leading to more accurate and temporally stable depth predictions. Furthermore, we propose a novel inference strategy for processing videos of arbitrary length avoiding the scale misalignment and redundant computations associated with overlapping windows used in other methods. Evaluations on multiple benchmark datasets demonstrate improved temporal consistency, competitive state-of-the-art performance and on top 2x faster processing in real-world scenarios.

Method: Proposes a topology-aware framework with two key mechanisms: 1) Topology-aware Consistency Matching (TACM) using graph contrastive learning and topological perturbations to learn robust matching patterns despite spatial misalignments, and 2) Topology-constrained Pathological Matching (TCPM) that aligns pathological positive regions based on node importance to enhance pathological consistency.

Result: Extensive experiments on two benchmarks across four staining tasks demonstrate superior performance over state-of-the-art approaches, achieving better generation quality with higher clinical relevance.

Conclusion: The proposed topology-aware framework effectively addresses the challenges of weakly-paired data in virtual IHC staining, providing a robust solution that outperforms existing methods and has strong clinical applicability.

Abstract: Immunohistochemical (IHC) staining provides crucial molecular characterization of tissue samples and plays an indispensable role in the clinical examination and diagnosis of cancers. However, compared with the commonly used Hematoxylin and Eosin (H&E) staining, IHC staining involves complex procedures and is both time-consuming and expensive, which limits its widespread clinical use. Virtual staining converts H&E images to IHC images, offering a cost-effective alternative to clinical IHC staining. Nevertheless, using adjacent slides as ground truth often results in weakly-paired data with spatial misalignment and local deformations, hindering effective supervised learning. To address these challenges, we propose a novel topology-aware framework for H&E-to-IHC virtual staining. Specifically, we introduce a Topology-aware Consistency Matching (TACM) mechanism that employs graph contrastive learning and topological perturbations to learn robust matching patterns despite spatial misalignments, ensuring structural consistency. Furthermore, we propose a Topology-constrained Pathological Matching (TCPM) mechanism that aligns pathological positive regions based on node importance to enhance pathological consistency. Extensive experiments on two benchmarks across four staining tasks demonstrate that our method outperforms state-of-the-art approaches, achieving superior generation quality with higher clinical relevance.

Method: Three-stage approach: 1) Sketch-Mode Cold Start: Convert standard long reasoning into sketch-style reasoning and finetune base model; 2) Train SketchJudge Reward Model to evaluate thinking processes and assign higher scores to sketch-style reasoning; 3) Sketch-Thinking Reinforcement Learning supervised by SketchJudge to generalize sketch-style reasoning ability.

Result: Achieves over 64% reduction in reasoning token cost without compromising final answer accuracy across four benchmarks. Qualitative analysis shows sketch-style reasoning focuses more on key cues during problem solving.

Conclusion: SketchThinker-R1 successfully incentivizes sketch-style reasoning in large multimodal models, significantly improving computational efficiency while maintaining accuracy, demonstrating the value of human-inspired cognitive efficiency approaches in AI systems.

Abstract: Despite the empirical success of extensive, step-by-step reasoning in large multimodal models, long reasoning processes inevitably incur substantial computational overhead, i.e., in terms of higher token costs and increased response time, which undermines inference efficiency. In contrast, humans often employ sketch-style reasoning: a concise, goal-directed cognitive process that prioritizes salient information and enables efficient problem-solving. Inspired by this cognitive efficiency, we propose SketchThinker-R1, which incentivizes sketch-style reasoning ability in large multimodal models. Our method consists of three primary stages. In the Sketch-Mode Cold Start stage, we convert standard long reasoning process into sketch-style reasoning and finetune base multimodal model, instilling initial sketch-style reasoning capability. Next, we train SketchJudge Reward Model, which explicitly evaluates thinking process of model and assigns higher scores to sketch-style reasoning. Finally, we conduct Sketch-Thinking Reinforcement Learning under supervision of SketchJudge to further generalize sketch-style reasoning ability. Experimental evaluation on four benchmarks reveals that our SketchThinker-R1 achieves over 64% reduction in reasoning token cost without compromising final answer accuracy. Qualitative analysis further shows that sketch-style reasoning focuses more on key cues during problem solving.

Method: Proposes DGA-Net with two key modules: 1) Cross-modal Graph Enhancement (CGE) synthesizes RGB semantics and depth geometry in a heterogeneous graph, 2) Anchor-Guided Refinement (AGR) creates global anchor and establishes non-local pathways to broadcast guidance from deep to shallow layers.

Result: Quantitative and qualitative experiments show DGA-Net outperforms state-of-the-art COD methods.

Conclusion: The novel depth prompting paradigm with CGE and AGR modules effectively adapts SAM for COD, achieving superior performance by fully exploiting depth cues.

Abstract: To fully exploit depth cues in Camouflaged Object Detection (COD), we present DGA-Net, a specialized framework that adapts the Segment Anything Model (SAM) via a novel ``depth prompting" paradigm. Distinguished from existing approaches that primarily rely on sparse prompts (e.g., points or boxes), our method introduces a holistic mechanism for constructing and propagating dense depth prompts. Specifically, we propose a Cross-modal Graph Enhancement (CGE) module that synthesizes RGB semantics and depth geometric within a heterogeneous graph to form a unified guidance signal. Furthermore, we design an Anchor-Guided Refinement (AGR) module. To counteract the inherent information decay in feature hierarchies, AGR forges a global anchor and establishes direct non-local pathways to broadcast this guidance from deep to shallow layers, ensuring precise and consistent segmentation. Quantitative and qualitative experimental results demonstrate that our proposed DGA-Net outperforms the state-of-the-art COD methods.

Method: SEF-DETR consists of three components: 1) Frequency-guided Patch Screening (FPS) that uses Fourier spectrum to construct target-relevant density maps and suppress background features; 2) Dynamic Embedding Enhancement (DEE) that strengthens multi-scale representations in a target-aware manner; 3) Reliability-Consistency-aware Fusion (RCF) that refines object queries by enforcing spatial-frequency consistency and reliability.

Result: Extensive experiments on three public IRSTD datasets demonstrate that SEF-DETR achieves superior detection performance compared to state-of-the-art methods, providing a robust and efficient solution for infrared small target detection.

Conclusion: SEF-DETR effectively addresses the limitations of existing DETR-based detectors for IRSTD by refining query initialization through frequency guidance and target-aware feature enhancement, offering a promising solution for challenging infrared small target detection scenarios.

Abstract: Infrared small target detection (IRSTD) faces significant challenges due to the low signal-to-noise ratio (SNR), small target size, and complex cluttered backgrounds. Although recent DETR-based detectors benefit from global context modeling, they exhibit notable performance degradation on IRSTD. We revisit this phenomenon and reveal that the target-relevant embeddings of IRST are inevitably overwhelmed by dominant background features due to the self-attention mechanism, leading to unreliable query initialization and inaccurate target localization. To address this issue, we propose SEF-DETR, a novel framework that refines query initialization for IRSTD. Specifically, SEF-DETR consists of three components: Frequency-guided Patch Screening (FPS), Dynamic Embedding Enhancement (DEE), and Reliability-Consistency-aware Fusion (RCF). The FPS module leverages the Fourier spectrum of local patches to construct a target-relevant density map, suppressing background-dominated features. DEE strengthens multi-scale representations in a target-aware manner, while RCF further refines object queries by enforcing spatial-frequency consistency and reliability. Extensive experiments on three public IRSTD datasets demonstrate that SEF-DETR achieves superior detection performance compared to state-of-the-art methods, delivering a robust and efficient solution for infrared small target detection task.

Method: Diffusion-based framework with key adaptations: location-aware palette with 2D gray code ordering, tanh activation for discrete data, sigmoid loss weighting, and x-prediction approach.

Result: The model narrows the performance gap with leading mask-based architectures but doesn’t surpass them yet. It introduces unique capabilities like principled ambiguity modeling that existing models lack.

Conclusion: While current performance doesn’t exceed mask-based models, combining proposed improvements with large-scale pretraining or promptable conditioning could lead to competitive universal segmentation models.

Abstract: This paper introduces a diffusion-based framework for universal image segmentation, making agnostic segmentation possible without depending on mask-based frameworks and instead predicting the full segmentation in a holistic manner. We present several key adaptations to diffusion models, which are important in this discrete setting. Notably, we show that a location-aware palette with our 2D gray code ordering improves performance. Adding a final tanh activation function is crucial for discrete data. On optimizing diffusion parameters, the sigmoid loss weighting consistently outperforms alternatives, regardless of the prediction type used, and we settle on x-prediction. While our current model does not yet surpass leading mask-based architectures, it narrows the performance gap and introduces unique capabilities, such as principled ambiguity modeling, that these models lack. All models were trained from scratch, and we believe that combining our proposed improvements with large-scale pretraining or promptable conditioning could lead to competitive models.

Method: Proposes TA-Prompting with Temporal Anchors that learn to precisely localize events and prompt VideoLLMs for temporal-aware understanding. Also introduces event coherent sampling strategy to select captions with sufficient coherence across temporal events and cross-modal similarity.

Result: Extensive experiments on benchmark datasets show TA-Prompting outperforms state-of-the-art VideoLLMs on dense video captioning and temporal understanding tasks including moment retrieval and temporalQA.

Conclusion: TA-Prompting effectively addresses the temporal localization challenge in VideoLLMs, improving both event boundary identification and coherent caption generation for dense video understanding.

Abstract: Dense video captioning aims to interpret and describe all temporally localized events throughout an input video. Recent state-of-the-art methods leverage large language models (LLMs) to provide detailed moment descriptions for video data. However, existing VideoLLMs remain challenging in identifying precise event boundaries in untrimmed videos, causing the generated captions to be not properly grounded. In this paper, we propose TA-Prompting, which enhances VideoLLMs via Temporal Anchors that learn to precisely localize events and prompt the VideoLLMs to perform temporal-aware video event understanding. During inference, in order to properly determine the output caption sequence from an arbitrary number of events presented within a video, we introduce an event coherent sampling strategy to select event captions with sufficient coherence across temporal events and cross-modal similarity with the given video. Through extensive experiments on benchmark datasets, we show that our TA-Prompting is favorable against state-of-the-art VideoLLMs, yielding superior performance on dense video captioning and temporal understanding tasks including moment retrieval and temporalQA.

Method: Two-stage training pipeline: 1) Supervised fine-tuning on Grounded-Rationale-IQA dataset to teach region grounding, and 2) Reinforcement learning with KL-Coverage regularizer to prevent reasoning diversity collapse and Progressive Re-sampling Strategy to mitigate annotation bias.

Result: Zoom-IQA achieves improved robustness, explainability, and generalization. Application to downstream tasks like image restoration demonstrates its effectiveness.

Conclusion: The proposed Zoom-IQA model successfully emulates key cognitive behaviors for IQA, addressing limitations of existing methods through explicit uncertainty awareness, region reasoning, and iterative refinement.

Abstract: Image Quality Assessment (IQA) is a long-standing problem in computer vision. Previous methods typically focus on predicting numerical scores without explanation or provide low-level descriptions lacking precise scores. Recent reasoning-based vision language models (VLMs) have shown strong potential for IQA, enabling joint generation of quality descriptions and scores. However, we notice that existing VLM-based IQA methods tend to exhibit unreliable reasoning due to their limited capability of integrating visual and textual cues. In this work, we introduce Zoom-IQA, a VLM-based IQA model to explicitly emulate key cognitive behaviors: uncertainty awareness, region reasoning, and iterative refinement. Specifically, we present a two-stage training pipeline: 1) supervised fine-tuning (SFT) on our Grounded-Rationale-IQA (GR-IQA) dataset to teach the model to ground its assessments in key regions; and 2) reinforcement learning (RL) for dynamic policy exploration, primarily stabilized by our KL-Coverage regularizer to prevent reasoning and scoring diversity collapse, and supported by a Progressive Re-sampling Strategy to mitigate annotation bias. Extensive experiments show that Zoom-IQA achieves improved robustness, explainability, and generalization. The application to downstream tasks, such as image restoration, further demonstrates the effectiveness of Zoom-IQA.

Method: Proposes DCG-ReID with: 1) Dynamic Confidence-based Disentangling Weighting (DCDW) mechanism to dynamically reweight modal contributions, 2) Collaboration Fusion Module (CFM) for balanced distributions to mine pairwise consensus features, and 3) Guidance Fusion Module (GFM) for unbalanced distributions to amplify dominant modality advantages and guide auxiliary modalities.

Result: Extensive experiments on three multi-modal ReID benchmarks (WMVeID863, MSVR310, RGBNT100) validate the effectiveness of the proposed method.

Conclusion: DCG-ReID successfully addresses the challenges of multi-modal vehicle ReID by disentangling heterogeneous quality-distributed modal data and providing scenario-specific fusion strategies, improving both intra-class consistency and inter-modal decision performance.

Abstract: Multi-modal vehicle Re-Identification (ReID) aims to leverage complementary information from RGB, Near Infrared (NIR), and Thermal Infrared (TIR) modalities to retrieve the same vehicle. The challenges of multi-modal vehicle ReID arise from the uncertainty of modality quality distribution induced by inherent discrepancies across modalities, resulting in distinct conflicting fusion requirements for data with balanced and unbalanced quality distributions. Existing methods handle all multi-modal data within a single fusion model, overlooking the different needs of the two data types and making it difficult to decouple the conflict between intra-class consistency and inter-modal heterogeneity. To this end, we propose Disentangle Collaboration and Guidance Fusion Representations for Multi-modal Vehicle ReID (DCG-ReID). Specifically, to disentangle heterogeneous quality-distributed modal data without mutual interference, we first design the Dynamic Confidence-based Disentangling Weighting (DCDW) mechanism: dynamically reweighting three-modal contributions via interaction-derived modal confidence to build a disentangled fusion framework. Building on DCDW, we develop two scenario-specific fusion strategies: (1) for balanced quality distributions, Collaboration Fusion Module (CFM) mines pairwise consensus features to capture shared discriminative information and boost intra-class consistency; (2) for unbalanced distributions, Guidance Fusion Module (GFM) implements differential amplification of modal discriminative disparities to reinforce dominant modality advantages, guide auxiliary modalities to mine complementary discriminative info, and mitigate inter-modal divergence to boost multi-modal joint decision performance. Extensive experiments on three multi-modal ReID benchmarks (WMVeID863, MSVR310, RGBNT100) validate the effectiveness of our method. Code will be released upon acceptance.

Method: Two-stage system: (1) coarse anomaly scoring via similarity to textual anchors, and (2) MLLM-based refinement with contextualized system/user prompts. Uses weakly supervised Automatic Prompt Engineering (APE) to optimize textual anchors and prompts without instruction tuning or frame-level annotations.

Result: Extensive experiments on standard VAD benchmarks show competitive detection performance and interpretable anomaly explanations. The system achieves this without instruction tuning, frame-level annotations, external modules, or dense processing.

Conclusion: PrismVAU provides an efficient, practical solution for real-time Video Anomaly Understanding that leverages a single off-the-shelf MLLM, making it suitable for real-world applications with reduced computational overhead.

Abstract: Video Anomaly Understanding (VAU) extends traditional Video Anomaly Detection (VAD) by not only localizing anomalies but also describing and reasoning about their context. Existing VAU approaches often rely on fine-tuned multimodal large language models (MLLMs) or external modules such as video captioners, which introduce costly annotations, complex training pipelines, and high inference overhead. In this work, we introduce PrismVAU, a lightweight yet effective system for real-time VAU that leverages a single off-the-shelf MLLM for anomaly scoring, explanation, and prompt optimization. PrismVAU operates in two complementary stages: (1) a coarse anomaly scoring module that computes frame-level anomaly scores via similarity to textual anchors, and (2) an MLLM-based refinement module that contextualizes anomalies through system and user prompts. Both textual anchors and prompts are optimized with a weakly supervised Automatic Prompt Engineering (APE) framework. Extensive experiments on standard VAD benchmarks demonstrate that PrismVAU delivers competitive detection performance and interpretable anomaly explanations – without relying on instruction tuning, frame-level annotations, and external modules or dense processing – making it an efficient and practical solution for real-world applications.

Method: Proposed HybridSolarNet integrates EfficientNet-B0 with Convolutional Block Attention Module (CBAM). Used tight split-before-augmentation protocol to avoid accuracy leakage. Implemented focal loss for imbalanced classes and cosine annealing. Evaluated on Kaggle Solar Panel Images dataset with 5-fold stratified cross-validation.

Result: Achieved 92.37% ± 0.41 average accuracy and F1-score of 0.9226 ± 0.39. Model requires only 16.3 MB storage (32× smaller than VGG19). Inference speed of 54.9 FPS with GPU support. CBAM contributed +1.53% accuracy boost. Grad-CAM visualizations show model focuses on actual fault locations.

Conclusion: HybridSolarNet is a lightweight, accurate, and fast model suitable for real-time UAV-based solar panel fault detection, addressing both computational constraints and accuracy estimation issues of previous methods.

Abstract: Manual inspections for solar panel systems are a tedious, costly, and error-prone task, making it desirable for Unmanned Aerial Vehicle (UAV) based monitoring. Though deep learning models have excellent fault detection capabilities, almost all methods either are too large and heavy for edge computing devices or involve biased estimation of accuracy due to ineffective learning techniques. We propose a new solar panel fault detection model called HybridSolarNet. It integrates EfficientNet-B0 with Convolutional Block Attention Module (CBAM). We implemented it on the Kaggle Solar Panel Images competition dataset with a tight split-before-augmentation protocol. It avoids leakage in accuracy estimation. We introduced focal loss and cosine annealing. Ablation analysis validates that accuracy boosts due to added benefits from CBAM (+1.53%) and that there are benefits from recognition of classes with imbalanced samples via focal loss. Overall average accuracy on 5-fold stratified cross-validation experiments on the given competition dataset topped 92.37% +/- 0.41 and an F1-score of 0.9226 +/- 0.39 compared to baselines like VGG19, requiring merely 16.3 MB storage, i.e., 32 times less. Its inference speed measured at 54.9 FPS with GPU support makes it a successful candidate for real-time UAV implementation. Moreover, visualization obtained from Grad-CAM illustrates that HybridSolarNet focuses on actual locations instead of irrelevant ones.

Method: Constructed VTONQA dataset with 8,132 images generated by 11 representative VTON models, collected 24,396 mean opinion scores across three evaluation dimensions (clothing fit, body compatibility, overall quality).

Result: Benchmarked both VTON models and diverse IQA metrics, revealing limitations of existing methods and demonstrating the value of the proposed dataset for perceptually aligned evaluation.

Conclusion: VTONQA dataset and benchmarks provide foundation for perceptually aligned evaluation, benefiting both quality assessment method development and VTON model advancement.

Abstract: With the rapid development of e-commerce and digital fashion, image-based virtual try-on (VTON) has attracted increasing attention. However, existing VTON models often suffer from artifacts such as garment distortion and body inconsistency, highlighting the need for reliable quality evaluation of VTON-generated images. To this end, we construct VTONQA, the first multi-dimensional quality assessment dataset specifically designed for VTON, which contains 8,132 images generated by 11 representative VTON models, along with 24,396 mean opinion scores (MOSs) across three evaluation dimensions (i.e., clothing fit, body compatibility, and overall quality). Based on VTONQA, we benchmark both VTON models and a diverse set of image quality assessment (IQA) metrics, revealing the limitations of existing methods and highlighting the value of the proposed dataset. We believe that the VTONQA dataset and corresponding benchmarks will provide a solid foundation for perceptually aligned evaluation, benefiting both the development of quality assessment methods and the advancement of VTON models.

Method: LAMS-Edit leverages intermediate states from the inversion process during edited image generation. It combines latent representations and attention maps from both processes at each step using weighted interpolation controlled by a scheduler (Latent and Attention Mixing with Schedulers), integrates with Prompt-to-Prompt, supports precise editing with region masks, and enables style transfer via LoRA.

Result: Extensive experiments demonstrate that LAMS-Edit effectively balances content preservation and edit application.

Conclusion: LAMS-Edit provides an extensible framework that addresses key challenges in text-to-image editing with diffusion models, offering improved balance between content preservation and edit application.

Abstract: Text-to-Image editing using diffusion models faces challenges in balancing content preservation with edit application and handling real-image editing. To address these, we propose LAMS-Edit, leveraging intermediate states from the inversion process–an essential step in real-image editing–during edited image generation. Specifically, latent representations and attention maps from both processes are combined at each step using weighted interpolation, controlled by a scheduler. This technique, Latent and Attention Mixing with Schedulers (LAMS), integrates with Prompt-to-Prompt (P2P) to form LAMS-Edit–an extensible framework that supports precise editing with region masks and enables style transfer via LoRA. Extensive experiments demonstrate that LAMS-Edit effectively balances content preservation and edit application.

Method: ULS+ incorporates additional public datasets and uses smaller input image sizes compared to the original ULS model, while maintaining the same architecture for segmenting lesions across the whole body in CT scans given volumes of interest centered around click-points.

Result: ULS+ significantly outperformed ULS in all comparisons using Dice score and robustness to click-point location on ULS23 Challenge test data and Longitudinal-CT dataset. ULS+ ranks first on the ULS23 Challenge test-phase leaderboard.

Conclusion: ULS+ establishes a foundation for robust and clinically relevant lesion segmentation models through data-driven updates and clinical validation cycles, demonstrating improved performance over the original ULS model.

Abstract: In this study, we present ULS+, an enhanced version of the Universal Lesion Segmentation (ULS) model. The original ULS model segments lesions across the whole body in CT scans given volumes of interest (VOIs) centered around a click-point. Since its release, several new public datasets have become available that can further improve model performance. ULS+ incorporates these additional datasets and uses smaller input image sizes, resulting in higher accuracy and faster inference. We compared ULS and ULS+ using the Dice score and robustness to click-point location on the ULS23 Challenge test data and a subset of the Longitudinal-CT dataset. In all comparisons, ULS+ significantly outperformed ULS. Additionally, ULS+ ranks first on the ULS23 Challenge test-phase leaderboard. By maintaining a cycle of data-driven updates and clinical validation, ULS+ establishes a foundation for robust and clinically relevant lesion segmentation models.

Method: Proposes a comic reasoning framework combining modular CoT generation with GRPO-based reinforcement fine-tuning and a novel structured reward mechanism.

Result: The 3B model outperforms SOTA methods across five challenging benchmarks, with plug-in experiments yielding additional 12.1% average improvement across different MLLMs.

Conclusion: The proposed framework effectively addresses CoT limitations in comic VQA, enabling small MLLMs to achieve better performance on humor-centric and abstract visual reasoning tasks.

Abstract: Comic-based visual question answering (CVQA) poses distinct challenges to multimodal large language models (MLLMs) due to its reliance on symbolic abstraction, narrative logic, and humor, which differ from conventional VQA tasks. Although Chain-of-Thought (CoT) prompting is widely used to enhance MLLM reasoning, surprisingly, its direct application to CVQA often degrades performance, especially in small-scale models. Our theoretical and empirical analyses reveal that standard CoT in CVQA suffers from state entanglement, spurious transitions, and exploration inefficiency, with small models particularly vulnerable in resource-constrained settings. To address these issues, we propose a novel comic reasoning framework, designed to produce more faithful and transferable reasoning chains in small MLLMs. Specifically, our framework combines modular CoT generation with GRPO-based reinforcement fine-tuning and a novel structured reward. Beyond comic VQA, we further evaluate our approach on a broader class of humor-centric and abstract visual reasoning tasks, including meme understanding and editorial cartoon interpretation. Across five challenging benchmarks, our 3B model outperforms state-of-the-art methods, and plug-in experiments yield an additional average improvement of $\mathbf{12.1%}$ across different MLLMs.

Method: Proposes Risk-intent Selective detection (RiSe) with two key components: 1) Potential Field-Trajectory Correlation Model (PTCM) using potential field theory to quantitatively assess kinematic risks, and 2) Intention-Driven Area Prediction Module (IDAPM) leveraging ego-motion priors to proactively predict and filter key BEV areas. Implements semantic-selective fusion that transmits high-fidelity features only from high-interaction regions.

Result: Extensive experiments on DeepAccident dataset show RiSe reduces communication volume to 0.71% of full feature sharing while maintaining state-of-the-art detection accuracy, establishing a competitive Pareto frontier between bandwidth efficiency and perception performance.

Conclusion: RiSe successfully shifts the VICP paradigm from identifying visible regions to prioritizing risk-critical ones, effectively acting as a feature denoiser that significantly reduces communication overhead without compromising detection performance.

Abstract: Vehicle-Infrastructure Collaborative Perception (VICP) is pivotal for resolving occlusion in autonomous driving, yet the trade-off between communication bandwidth and feature redundancy remains a critical bottleneck. While intermediate fusion mitigates data volume compared to raw sharing, existing frameworks typically rely on spatial compression or static confidence maps, which inefficiently transmit spatially redundant features from non-critical background regions. To address this, we propose Risk-intent Selective detection (RiSe), an interaction-aware framework that shifts the paradigm from identifying visible regions to prioritizing risk-critical ones. Specifically, we introduce a Potential Field-Trajectory Correlation Model (PTCM) grounded in potential field theory to quantitatively assess kinematic risks. Complementing this, an Intention-Driven Area Prediction Module (IDAPM) leverages ego-motion priors to proactively predict and filter key Bird’s-Eye-View (BEV) areas essential for decision-making. By integrating these components, RiSe implements a semantic-selective fusion scheme that transmits high-fidelity features only from high-interaction regions, effectively acting as a feature denoiser. Extensive experiments on the DeepAccident dataset demonstrate that our method reduces communication volume to 0.71% of full feature sharing while maintaining state-of-the-art detection accuracy, establishing a competitive Pareto frontier between bandwidth efficiency and perception performance.

Method: 1) Self-Augmented point clouds via triangulation between training and rasterized extrapolated views for coverage estimation; 2) First residual learning strategy for 3D Gaussian Splatting with uncertainty-driven filtering and dropout/hard-negative-mining sampling; 3) Physically grounded view selection for uniform coverage.

Result: Outperforms state-of-the-art baselines in both reconstruction quality and view selection robustness for active view selection tasks.

Conclusion: SA-ResGS provides a comprehensive solution for NBV selection by improving uncertainty reliability, enhancing supervision for weak Gaussians, and implicitly unbiasing uncertainty quantification through constrained view selection and residual supervision.

Abstract: We propose Self-Augmented Residual 3D Gaussian Splatting (SA-ResGS), a novel framework to stabilize uncertainty quantification and enhancing uncertainty-aware supervision in next-best-view (NBV) selection for active scene reconstruction. SA-ResGS improves both the reliability of uncertainty estimates and their effectiveness for supervision by generating Self-Augmented point clouds (SA-Points) via triangulation between a training view and a rasterized extrapolated view, enabling efficient scene coverage estimation. While improving scene coverage through physically guided view selection, SA-ResGS also addresses the challenge of under-supervised Gaussians, exacerbated by sparse and wide-baseline views, by introducing the first residual learning strategy tailored for 3D Gaussian Splatting. This targeted supervision enhances gradient flow in high-uncertainty Gaussians by combining uncertainty-driven filtering with dropout- and hard-negative-mining-inspired sampling. Our contributions are threefold: (1) a physically grounded view selection strategy that promotes efficient and uniform scene coverage; (2) an uncertainty-aware residual supervision scheme that amplifies learning signals for weakly contributing Gaussians, improving training stability and uncertainty estimation across scenes with diverse camera distributions; (3) an implicit unbiasing of uncertainty quantification as a consequence of constrained view selection and residual supervision, which together mitigate conflicting effects of wide-baseline exploration and sparse-view ambiguity in NBV planning. Experiments on active view selection demonstrate that SA-ResGS outperforms state-of-the-art baselines in both reconstruction quality and view selection robustness.

Method: Two frameworks combining PointNet with generative models: 1) Flow Matching PointNet integrates PointNet with flow matching models, 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 without pixelation.

Result: Both frameworks achieve more accurate predictions of velocity and pressure fields, as well as lift and drag forces compared to vanilla PointNet with same parameters. They don’t exhibit high-frequency noise artifacts seen in graph neural network-based approaches and show greater robustness to incomplete geometries.

Conclusion: Flow Matching PointNet and Diffusion PointNet provide effective, unified architectures for fluid flow prediction on irregular geometries using point clouds, offering improved accuracy, reduced noise artifacts, and simpler conditioning compared to existing approaches.

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. 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: DFRCP enhances YOLOv11’s feature pyramid by combining large and medium scale features while preserving native representations. It introduces Dynamic Robust Switch units that adaptively inject fuzzy features (synthesized by rotating and nonlinearly interpolating multi-scale features) through a transparency convolution that learns content-adaptive trade-offs. Includes CUDA parallel rotation/interpolation kernel for 400x speedup.

Result: On blurred test sets from a private wheat pest damage dataset (3,500 images, 3x augmented with uniform motion blur and bounding box rotational blur), YOLOv11 with DFRCP achieves ~10.4% higher accuracy than baseline YOLOv11 with only modest training time overhead.

Conclusion: DFRCP provides an effective plug-in solution for blur-robust object detection that balances accuracy and computational efficiency, reducing the need for manual filtering after data collection while being practical for edge deployment.

Abstract: Motion blur caused by camera shake produces ghosting artifacts that substantially degrade edge side object detection. Existing approaches either suppress blur as noise and lose discriminative structure, or apply full image restoration that increases latency and limits deployment on resource constrained devices. We propose DFRCP, a Dynamic Fuzzy Robust Convolutional Pyramid, as a plug in upgrade to YOLOv11 for blur robust detection. DFRCP enhances the YOLOv11 feature pyramid by combining large scale and medium scale features while preserving native representations, and by introducing Dynamic Robust Switch units that adaptively inject fuzzy features to strengthen global perception under jitter. Fuzzy features are synthesized by rotating and nonlinearly interpolating multiscale features, then merged through a transparency convolution that learns a content adaptive trade off between original and fuzzy cues. We further develop a CUDA parallel rotation and interpolation kernel that avoids boundary overflow and delivers more than 400 times speedup, making the design practical for edge deployment. We train with paired supervision on a private wheat pest damage dataset of about 3,500 images, augmented threefold using two blur regimes, uniform image wide motion blur and bounding box confined rotational blur. On blurred test sets, YOLOv11 with DFRCP achieves about 10.4 percent higher accuracy than the YOLOv11 baseline with only a modest training time overhead, reducing the need for manual filtering after data collection.

Method: The authors formalize spatial understanding as learning a Group Homomorphism that preserves algebraic structure of transformation groups. They analyze computational complexity: proving constant-depth ViTs with polynomial precision are bounded by TC⁰, while spatial reasoning for non-solvable groups (like SO(3)) is NC¹-complete via the Word Problem. They validate the complexity gap through latent-space probing experiments.

Result: Theoretical analysis establishes a complexity boundary: constant-depth ViTs fundamentally lack the logical depth to efficiently capture non-solvable spatial structures (assuming TC⁰ ⊊ NC¹). Experimental validation shows ViT representations suffer structural collapse on non-solvable tasks as compositional depth increases.

Conclusion: ViTs’ spatial reasoning failures are not merely data limitations but stem from fundamental architectural complexity constraints. The TC⁰ vs NC¹ complexity gap explains why constant-depth ViTs cannot efficiently learn non-solvable spatial transformations, providing a theoretical foundation for understanding architectural limitations in spatial reasoning.

Abstract: Vision Transformers (ViTs) excel in semantic recognition but exhibit systematic failures in spatial reasoning tasks such as mental rotation. While often attributed to data scale, we propose that this limitation arises from the intrinsic circuit complexity of the architecture. We formalize spatial understanding as learning a Group Homomorphism: mapping image sequences to a latent space that preserves the algebraic structure of the underlying transformation group. We demonstrate that for non-solvable groups (e.g., the 3D rotation group $\mathrm{SO}(3)$), maintaining such a structure-preserving embedding is computationally lower-bounded by the Word Problem, which is $\mathsf{NC^1}$-complete. In contrast, we prove that constant-depth ViTs with polynomial precision are strictly bounded by $\mathsf{TC^0}$. Under the conjecture $\mathsf{TC^0} \subsetneq \mathsf{NC^1}$, we establish a complexity boundary: constant-depth ViTs fundamentally lack the logical depth to efficiently capture non-solvable spatial structures. We validate this complexity gap via latent-space probing, demonstrating that ViT representations suffer a structural collapse on non-solvable tasks as compositional depth increases.

Method: IBISAgent reformulates segmentation as a vision-centric, multi-step decision-making process where MLLMs generate interleaved reasoning and text-based click actions to invoke segmentation tools. It uses iterative multi-step visual reasoning on masked image features for mask refinement. A two-stage training framework includes cold-start supervised fine-tuning followed by agentic reinforcement learning with fine-grained rewards.

Result: Extensive experiments show IBISAgent consistently outperforms both closed-source and open-source state-of-the-art methods on medical referring and reasoning segmentation tasks.

Conclusion: IBISAgent successfully addresses limitations of existing approaches by enabling iterative refinement without architectural modifications, developing pixel-level visual reasoning capabilities, and demonstrating superior performance through a novel agentic framework with two-stage training.

Abstract: Recent research on medical MLLMs has gradually shifted its focus from image-level understanding to fine-grained, pixel-level comprehension. Although segmentation serves as the foundation for pixel-level understanding, existing approaches face two major challenges. First, they introduce implicit segmentation tokens and require simultaneous fine-tuning of both the MLLM and external pixel decoders, which increases the risk of catastrophic forgetting and limits generalization to out-of-domain scenarios. Second, most methods rely on single-pass reasoning and lack the capability to iteratively refine segmentation results, leading to suboptimal performance. To overcome these limitations, we propose a novel agentic MLLM, named IBISAgent, that reformulates segmentation as a vision-centric, multi-step decision-making process. IBISAgent enables MLLMs to generate interleaved reasoning and text-based click actions, invoke segmentation tools, and produce high-quality masks without architectural modifications. By iteratively performing multi-step visual reasoning on masked image features, IBISAgent naturally supports mask refinement and promotes the development of pixel-level visual reasoning capabilities. We further design a two-stage training framework consisting of cold-start supervised fine-tuning and agentic reinforcement learning with tailored, fine-grained rewards, enhancing the model’s robustness in complex medical referring and reasoning segmentation tasks. Extensive experiments demonstrate that IBISAgent consistently outperforms both closed-source and open-source SOTA methods. All datasets, code, and trained models will be released publicly.

Method: Concept-Feature Structuralized Generalization (CFSG) explicitly disentangles both concept and feature spaces into three structured components: common, specific, and confounding segments. An adaptive mechanism dynamically adjusts the proportions of these components based on distribution shifts. In final prediction, explicit weights are assigned to each pair of components.

Result: On three single-source benchmark datasets, CFSG achieves average performance improvement of 9.87% over baseline models and outperforms existing state-of-the-art methods by average of 3.08%. Explainability analysis validates effective integration of multi-granularity structured knowledge and confirms feature structuralization facilitates concept structuralization emergence.

Conclusion: CFSG successfully addresses fine-grained domain generalization challenges by structurally disentangling representations into common, specific, and confounding components with adaptive adjustment, inspired by human cognitive mechanisms for fine-grained classification.

Abstract: Fine-Grained Domain Generalization (FGDG) presents greater challenges than conventional domain generalization due to the subtle inter-class differences and relatively pronounced intra-class variations inherent in fine-grained recognition tasks. Under domain shifts, the model becomes overly sensitive to fine-grained cues, leading to the suppression of critical features and a significant drop in performance. Cognitive studies suggest that humans classify objects by leveraging both common and specific attributes, enabling accurate differentiation between fine-grained categories. However, current deep learning models have yet to incorporate this mechanism effectively. Inspired by this mechanism, we propose Concept-Feature Structuralized Generalization (CFSG). This model explicitly disentangles both the concept and feature spaces into three structured components: common, specific, and confounding segments. To mitigate the adverse effects of varying degrees of distribution shift, we introduce an adaptive mechanism that dynamically adjusts the proportions of common, specific, and confounding components. In the final prediction, explicit weights are assigned to each pair of components. Extensive experiments on three single-source benchmark datasets demonstrate that CFSG achieves an average performance improvement of 9.87% over baseline models and outperforms existing state-of-the-art methods by an average of 3.08%. Additionally, explainability analysis validates that CFSG effectively integrates multi-granularity structured knowledge and confirms that feature structuralization facilitates the emergence of concept structuralization.

Method: A multi-agent LLM framework with three specialized agents: visual agent (handles visual cues), language agent (processes narrative context), and critic agent (coordinates reasoning). They work collaboratively to integrate visual and narrative information for structured reasoning.

Result: Systematically evaluated on Pororo and Simpsons cartoon VQA datasets. Results provide detailed analysis of each agent’s contribution to final predictions, offering insights into multi-agent LLM behavior for cartoon VQA and multimodal inference.

Conclusion: The multi-agent framework effectively addresses cartoon VQA challenges by enabling collaborative reasoning between specialized agents, providing a deeper understanding of how LLM-based multi-agent systems handle multimodal inference in stylized imagery.

Abstract: Visual Question Answering (VQA) for stylised cartoon imagery presents challenges, such as interpreting exaggerated visual abstraction and narrative-driven context, which are not adequately addressed by standard large language models (LLMs) trained on natural images. To investigate this issue, a multi-agent LLM framework is introduced, specifically designed for VQA tasks in cartoon imagery. The proposed architecture consists of three specialised agents: visual agent, language agent and critic agent, which work collaboratively to support structured reasoning by integrating visual cues and narrative context. The framework was systematically evaluated on two cartoon-based VQA datasets: Pororo and Simpsons. Experimental results provide a detailed analysis of how each agent contributes to the final prediction, offering a deeper understanding of LLM-based multi-agent behaviour in cartoon VQA and multimodal inference.

Method: LesionTABE couples adversarial debiasing techniques with dermatology-specific foundation model embeddings. The framework uses adversarial training to remove skin tone bias while leveraging specialized embeddings from foundation models trained on dermatology data.

Result: Evaluated across multiple datasets covering both malignant and inflammatory conditions, LesionTABE achieves over 25% improvement in fairness metrics compared to ResNet-152 baseline. It outperforms existing debiasing methods while simultaneously enhancing overall diagnostic accuracy.

Conclusion: Foundation model debiasing shows strong potential for creating equitable clinical AI systems. LesionTABE represents a step toward addressing skin tone bias in dermatology AI, which could facilitate broader clinical adoption of these technologies.

Abstract: Bias remains a major barrier to the clinical adoption of AI in dermatology, as diagnostic models underperform on darker skin tones. We present LesionTABE, a fairness-centric framework that couples adversarial debiasing with dermatology-specific foundation model embeddings. Evaluated across multiple datasets covering both malignant and inflammatory conditions, LesionTABE achieves over a 25% improvement in fairness metrics compared to a ResNet-152 baseline, outperforming existing debiasing methods while simultaneously enhancing overall diagnostic accuracy. These results highlight the potential of foundation model debiasing as a step towards equitable clinical AI adoption.

Method: TGIF treats vision encoder layers as depth-wise “experts” and predicts a prompt-dependent fusion of visual features using a lightweight module that follows direct external fusion principles, requiring no vision-encoder updates and adding minimal overhead.

Result: Integrated into LLaVA-1.5-7B, TGIF provides consistent improvements across hallucination, OCR, and VQA benchmarks while preserving or improving performance on ScienceQA, GQA, and MMBench.

Conclusion: Query-conditioned, hierarchy-aware fusion is an effective approach to strengthen visual grounding and reduce hallucination in modern MLLMs, demonstrating the value of dynamically leveraging the full hierarchy of visual features.

Abstract: Multimodal large language models (MLLMs) typically rely on a single late-layer feature from a frozen vision encoder, leaving the encoder’s rich hierarchy of visual cues under-utilized. MLLMs still suffer from visually ungrounded hallucinations, often relying on language priors rather than image evidence. While many prior mitigation strategies operate on the text side, they leave the visual representation unchanged and do not exploit the rich hierarchy of features encoded across vision layers. Existing multi-layer fusion methods partially address this limitation but remain static, applying the same layer mixture regardless of the query. In this work, we introduce TGIF (Text-Guided Inter-layer Fusion), a lightweight module that treats encoder layers as depth-wise “experts” and predicts a prompt-dependent fusion of visual features. TGIF follows the principle of direct external fusion, requires no vision-encoder updates, and adds minimal overhead. Integrated into LLaVA-1.5-7B, TGIF provides consistent improvements across hallucination, OCR, and VQA benchmarks, while preserving or improving performance on ScienceQA, GQA, and MMBench. These results suggest that query-conditioned, hierarchy-aware fusion is an effective way to strengthen visual grounding and reduce hallucination in modern MLLMs.

Method: Used a dataset of 9,032 grape leaf images across 4 classes (3 disease types + healthy). After pre-processing and 70-20-10 train-validation-test split, deployed two pre-trained models (InceptionV3 and Xception) with adversarial training for robustness. Integrated Grad-CAM for explainability and deployed a Streamlit web application with heatmap visualization and confidence-level predictions.

Result: Xception achieved 96.23% accuracy, outperforming InceptionV3. The adversarial training improved model robustness, and Grad-CAM successfully confirmed disease localization. The web application provides practical deployment with visualization capabilities.

Conclusion: The proposed approach effectively classifies grape leaf diseases with high accuracy using Xception model. The integration of adversarial training and Grad-CAM provides both robustness and transparency, making it suitable for real-world agricultural applications through the deployed web interface.

Abstract: Plant disease diagnosis is essential to farmers’ management choices because plant diseases frequently lower crop yield and product quality. For harvests to flourish and agricultural productivity to boost, grape leaf disease detection is important. The plant disease dataset contains grape leaf diseases total of 9,032 images of four classes, among them three classes are leaf diseases, and the other one is healthy leaves. After rigorous pre-processing dataset was split (70% training, 20% validation, 10% testing), and two pre-trained models were deployed: InceptionV3 and Xception. Xception shows a promising result of 96.23% accuracy, which is remarkable than InceptionV3. Adversarial Training is used for robustness, along with more transparency. Grad-CAM is integrated to confirm the leaf disease. Finally deployed a web application using Streamlit with a heatmap visualization and prediction with confidence level for robust grape leaf disease classification.

Method: Proposes Unified Thinker, a task-agnostic reasoning architecture with a dedicated Thinker module separate from the image Generator. Uses a two-stage training paradigm: first builds a structured planning interface for the Thinker, then applies reinforcement learning to ground its policy in pixel-level feedback, optimizing visual correctness over textual plausibility.

Result: Extensive experiments on text-to-image generation and image editing show that Unified Thinker substantially improves image reasoning and generation quality.

Conclusion: Decoupling reasoning from generation enables modular upgrades and better logic-intensive instruction following. The proposed architecture and training approach effectively bridge the reasoning-execution gap in generative models.

Abstract: Despite impressive progress in high-fidelity image synthesis, generative models still struggle with logic-intensive instruction following, exposing a persistent reasoning–execution gap. Meanwhile, closed-source systems (e.g., Nano Banana) have demonstrated strong reasoning-driven image generation, highlighting a substantial gap to current open-source models. We argue that closing this gap requires not merely better visual generators, but executable reasoning: decomposing high-level intents into grounded, verifiable plans that directly steer the generative process. To this end, we propose Unified Thinker, a task-agnostic reasoning architecture for general image generation, designed as a unified planning core that can plug into diverse generators and workflows. Unified Thinker decouples a dedicated Thinker from the image Generator, enabling modular upgrades of reasoning without retraining the entire generative model. We further introduce a two-stage training paradigm: we first build a structured planning interface for the Thinker, then apply reinforcement learning to ground its policy in pixel-level feedback, encouraging plans that optimize visual correctness over textual plausibility. Extensive experiments on text-to-image generation and image editing show that Unified Thinker substantially improves image reasoning and generation quality.

Method: Uses lightweight transformer with linear complexity to process large images, represents nuclei as star-convex polygons, and introduces novel radial distance loss for natural segmentation of overlapping nuclei without explicit overlap annotations.

Result: Shows strong generalization across tissues (PanNuke and MoNuSeg datasets) and state-of-the-art efficiency - over 5x faster than next-fastest leading method.

Conclusion: LSP-DETR provides precise, scalable, and efficient instance segmentation for computational pathology without handcrafted post-processing, enabling better analysis of whole-slide images.

Abstract: Precise and scalable instance segmentation of cell nuclei is essential for computational pathology, yet gigapixel Whole-Slide Images pose major computational challenges. Existing approaches rely on patch-based processing and costly post-processing for instance separation, sacrificing context and efficiency. We introduce LSP-DETR (Local Star Polygon DEtection TRansformer), a fully end-to-end framework that uses a lightweight transformer with linear complexity to process substantially larger images without additional computational cost. Nuclei are represented as star-convex polygons, and a novel radial distance loss function allows the segmentation of overlapping nuclei to emerge naturally, without requiring explicit overlap annotations or handcrafted post-processing. Evaluations on PanNuke and MoNuSeg show strong generalization across tissues and state-of-the-art efficiency, with LSP-DETR being over five times faster than the next-fastest leading method. Code and models are available at https://github.com/RationAI/lsp-detr.

Method: DiffBench: 3-stage automated evaluation pipeline for diffusion architectures, optimizations, and deployment scenarios. DiffAgent: LLM agent with closed-loop workflow (planning, code generation, debugging components) using genetic algorithm for performance feedback.

Result: DiffBench provides thorough evaluation of generated acceleration codes. DiffAgent significantly outperforms existing LLMs in producing effective diffusion acceleration strategies.

Conclusion: The framework enables automated generation of optimal acceleration strategies for diffusion models, addressing the challenge of combining multiple acceleration techniques through LLM-driven code generation and evaluation.

Abstract: Diffusion models have achieved remarkable success in image and video generation. However, their inherently multiple step inference process imposes substantial computational overhead, hindering real-world deployment. Accelerating diffusion models is therefore essential, yet determining how to combine multiple model acceleration techniques remains a significant challenge. To address this issue, we introduce a framework driven by large language models (LLMs) for automated acceleration code generation and evaluation. First, we present DiffBench, a comprehensive benchmark that implements a three stage automated evaluation pipeline across diverse diffusion architectures, optimization combinations and deployment scenarios. Second, we propose DiffAgent, an agent that generates optimal acceleration strategies and codes for arbitrary diffusion models. DiffAgent employs a closed-loop workflow in which a planning component and a debugging component iteratively refine the output of a code generation component, while a genetic algorithm extracts performance feedback from the execution environment to guide subsequent code refinements. We provide a detailed explanation of the DiffBench construction and the design principles underlying DiffAgent. Extensive experiments show that DiffBench offers a thorough evaluation of generated codes and that DiffAgent significantly outperforms existing LLMs in producing effective diffusion acceleration strategies.

Method: Two-stage approach inspired by radiological workflow: 1) identifies anatomical structures and extracts features, 2) uses LLM for downstream tasks (phrase grounding, report generation, VQA, image understanding).

Result: Superior anatomical reasoning with over 25% improvement on anatomy grounding, phrase grounding, grounded diagnosis, and grounded captioning tasks across multiple benchmarks.

Conclusion: AnatomiX effectively addresses anatomical understanding gaps in medical LLMs through explicit anatomical grounding, significantly improving chest X-ray interpretation performance.

Abstract: Multimodal medical large language models have shown impressive progress in chest X-ray interpretation but continue to face challenges in spatial reasoning and anatomical understanding. Although existing grounding techniques improve overall performance, they often fail to establish a true anatomical correspondence, resulting in incorrect anatomical understanding in the medical domain. To address this gap, we introduce AnatomiX, a multitask multimodal large language model explicitly designed for anatomically grounded chest X-ray interpretation. Inspired by the radiological workflow, AnatomiX adopts a two stage approach: first, it identifies anatomical structures and extracts their features, and then leverages a large language model to perform diverse downstream tasks such as phrase grounding, report generation, visual question answering, and image understanding. Extensive experiments across multiple benchmarks demonstrate that AnatomiX achieves superior anatomical reasoning and delivers over 25% improvement in performance on anatomy grounding, phrase grounding, grounded diagnosis and grounded captioning tasks compared to existing approaches. Code and pretrained model are available at https://github.com/aneesurhashmi/anatomix

Method: UniCorn partitions a single UMM into three collaborative roles: Proposer (generates initial content), Solver (processes and refines), and Judge (evaluates quality). The framework uses self-play to generate high-quality interactions and cognitive pattern reconstruction to distill latent understanding into explicit generative signals.

Result: UniCorn achieves comprehensive improvements across six image generation benchmarks, with SOTA performance on TIIF (73.8), DPG (86.8), CompBench (88.5), and the new UniCycle benchmark. It also delivers substantial gains of +5.0 on WISE and +6.5 on OneIG while maintaining robust comprehension capabilities.

Conclusion: The method significantly enhances text-to-image generation while preserving comprehension, demonstrating that fully self-supervised refinement is scalable for unified multimodal intelligence. The proposed UniCycle benchmark validates restoration of multimodal coherence through cycle-consistency evaluation.

Abstract: While Unified Multimodal Models (UMMs) have achieved remarkable success in cross-modal comprehension, a significant gap persists in their ability to leverage such internal knowledge for high-quality generation. We formalize this discrepancy as Conduction Aphasia, a phenomenon where models accurately interpret multimodal inputs but struggle to translate that understanding into faithful and controllable synthesis. To address this, we propose UniCorn, a simple yet elegant self-improvement framework that eliminates the need for external data or teacher supervision. By partitioning a single UMM into three collaborative roles: Proposer, Solver, and Judge, UniCorn generates high-quality interactions via self-play and employs cognitive pattern reconstruction to distill latent understanding into explicit generative signals. To validate the restoration of multimodal coherence, we introduce UniCycle, a cycle-consistency benchmark based on a Text to Image to Text reconstruction loop. Extensive experiments demonstrate that UniCorn achieves comprehensive and substantial improvements over the base model across six general image generation benchmarks. Notably, it achieves SOTA performance on TIIF(73.8), DPG(86.8), CompBench(88.5), and UniCycle while further delivering substantial gains of +5.0 on WISE and +6.5 on OneIG. These results highlight that our method significantly enhances T2I generation while maintaining robust comprehension, demonstrating the scalability of fully self-supervised refinement for unified multimodal intelligence.

Method: Asymmetric dual-stream transformer architecture with 14B-parameter video stream and 5B-parameter audio stream, connected via bidirectional audio-video cross-attention layers with temporal positional embeddings and cross-modality AdaLN. Uses multilingual text encoder and modality-aware classifier-free guidance (modality-CFG) for improved alignment.

Result: Achieves state-of-the-art audiovisual quality and prompt adherence among open-source systems, with results comparable to proprietary models at a fraction of computational cost and inference time. Generates rich, coherent audio tracks with speech, background, and foley elements synchronized to video content.

Conclusion: LTX-2 demonstrates that unified audiovisual generation is feasible with efficient architecture design, enabling high-quality synchronized content generation while being computationally efficient and open-source.

Abstract: Recent text-to-video diffusion models can generate compelling video sequences, yet they remain silent – missing the semantic, emotional, and atmospheric cues that audio provides. We introduce LTX-2, an open-source foundational model capable of generating high-quality, temporally synchronized audiovisual content in a unified manner. LTX-2 consists of an asymmetric dual-stream transformer with a 14B-parameter video stream and a 5B-parameter audio stream, coupled through bidirectional audio-video cross-attention layers with temporal positional embeddings and cross-modality AdaLN for shared timestep conditioning. This architecture enables efficient training and inference of a unified audiovisual model while allocating more capacity for video generation than audio generation. We employ a multilingual text encoder for broader prompt understanding and introduce a modality-aware classifier-free guidance (modality-CFG) mechanism for improved audiovisual alignment and controllability. Beyond generating speech, LTX-2 produces rich, coherent audio tracks that follow the characters, environment, style, and emotion of each scene – complete with natural background and foley elements. In our evaluations, the model achieves state-of-the-art audiovisual quality and prompt adherence among open-source systems, while delivering results comparable to proprietary models at a fraction of their computational cost and inference time. All model weights and code are publicly released.

Method: Proposes MultiMedia-Agent system with: 1) data generation pipeline, 2) content creation tool library, 3) preference alignment metrics. Uses skill acquisition theory for data curation and agent training. Implements two-stage correlation strategy for plan optimization (self-correlation and model preference correlation). Trains agent via three-stage approach: base/success plan finetuning and preference optimization.

Result: Comparison results demonstrate the approaches are effective, and the MultiMedia-Agent can generate better multimedia content compared to novel models.

Conclusion: Agent-based systems enable tackling complex content generation tasks that individual AIGC models cannot handle, providing an integrated solution for real-world multimedia content creation.

Abstract: With the advancement of AIGC (AI-generated content) technologies, an increasing number of generative models are revolutionizing fields such as video editing, music generation, and even film production. However, due to the limitations of current AIGC models, most models can only serve as individual components within specific application scenarios and are not capable of completing tasks end-to-end in real-world applications. In real-world applications, editing experts often work with a wide variety of images and video inputs, producing multimodal outputs – a video typically includes audio, text, and other elements. This level of integration across multiple modalities is something current models are unable to achieve effectively. However, the rise of agent-based systems has made it possible to use AI tools to tackle complex content generation tasks. To deal with the complex scenarios, in this paper, we propose a MultiMedia-Agent designed to automate complex content creation. Our agent system includes a data generation pipeline, a tool library for content creation, and a set of metrics for evaluating preference alignment. Notably, we introduce the skill acquisition theory to model the training data curation and agent training. We designed a two-stage correlation strategy for plan optimization, including self-correlation and model preference correlation. Additionally, we utilized the generated plans to train the MultiMedia-Agent via a three stage approach including base/success plan finetune and preference optimization. The comparison results demonstrate that the our approaches are effective and the MultiMedia-Agent can generate better multimedia content compared to novel models.

Method: InfiniDepth represents depth as neural implicit fields using a local implicit decoder that can query depth at continuous 2D coordinates. The authors also curate a high-quality 4K synthetic benchmark from five different games for evaluation.

Result: InfiniDepth achieves state-of-the-art performance on both synthetic and real-world benchmarks across relative and metric depth estimation tasks, particularly excelling in fine-detail regions. It also improves novel view synthesis under large viewpoint shifts with fewer holes and artifacts.

Conclusion: Representing depth as neural implicit fields enables arbitrary-resolution and fine-grained depth estimation, overcoming fundamental limitations of existing discrete grid-based methods. The approach shows strong performance across multiple tasks and benefits downstream applications like novel view synthesis.

Abstract: Existing depth estimation methods are fundamentally limited to predicting depth on discrete image grids. Such representations restrict their scalability to arbitrary output resolutions and hinder the geometric detail recovery. This paper introduces InfiniDepth, which represents depth as neural implicit fields. Through a simple yet effective local implicit decoder, we can query depth at continuous 2D coordinates, enabling arbitrary-resolution and fine-grained depth estimation. To better assess our method’s capabilities, we curate a high-quality 4K synthetic benchmark from five different games, spanning diverse scenes with rich geometric and appearance details. Extensive experiments demonstrate that InfiniDepth achieves state-of-the-art performance on both synthetic and real-world benchmarks across relative and metric depth estimation tasks, particularly excelling in fine-detail regions. It also benefits the task of novel view synthesis under large viewpoint shifts, producing high-quality results with fewer holes and artifacts.

Method: Muses uses a 3D skeleton as fundamental representation to guide generation through a three-stage pipeline: 1) Graph-constrained reasoning to construct creatively composed 3D skeletons with coherent layout and scale, 2) Voxel-based assembly in structured latent space guided by the skeleton to integrate regions from different objects, and 3) Image-guided appearance modeling under skeletal conditions to generate style-consistent textures.

Result: Extensive experiments show Muses achieves state-of-the-art performance in visual fidelity and alignment with textual descriptions. The method also demonstrates potential for flexible 3D object editing.

Conclusion: Muses presents the first training-free, feed-forward method for fantastic 3D creature generation that overcomes limitations of previous approaches by leveraging skeletal guidance for explicit and rational composition of diverse elements.

Abstract: We present Muses, the first training-free method for fantastic 3D creature generation in a feed-forward paradigm. Previous methods, which rely on part-aware optimization, manual assembly, or 2D image generation, often produce unrealistic or incoherent 3D assets due to the challenges of intricate part-level manipulation and limited out-of-domain generation. In contrast, Muses leverages the 3D skeleton, a fundamental representation of biological forms, to explicitly and rationally compose diverse elements. This skeletal foundation formalizes 3D content creation as a structure-aware pipeline of design, composition, and generation. Muses begins by constructing a creatively composed 3D skeleton with coherent layout and scale through graph-constrained reasoning. This skeleton then guides a voxel-based assembly process within a structured latent space, integrating regions from different objects. Finally, image-guided appearance modeling under skeletal conditions is applied to generate a style-consistent and harmonious texture for the assembled shape. Extensive experiments establish Muses’ state-of-the-art performance in terms of visual fidelity and alignment with textual descriptions, and potential on flexible 3D object editing. Project page: https://luhexiao.github.io/Muses.github.io/.

Method: Created Teeth3DS+ dataset with rigorously curated intraoral scans from state-of-the-art scanners, validated by experienced dental professionals, with standardized data splits and evaluation protocols.

Result: Provides a comprehensive public benchmark supporting multiple fundamental dental analysis tasks, enabling fair and reproducible comparison of methods.

Conclusion: Teeth3DS+ aims to foster progress in learning-based analysis of 3D dental scans by addressing the dataset scarcity problem and providing standardized evaluation frameworks.

Abstract: Intraoral 3D scanning is now widely adopted in modern dentistry and plays a central role in supporting key tasks such as tooth segmentation, detection, labeling, and dental landmark identification. Accurate analysis of these scans is essential for orthodontic and restorative treatment planning, as it enables automated workflows and minimizes the need for manual intervention. However, the development of robust learning-based solutions remains challenging due to the limited availability of high-quality public datasets and standardized benchmarks. This article presents Teeth3DS+, an extended public benchmark dedicated to intraoral 3D scan analysis. Developed in the context of the MICCAI 3DTeethSeg and 3DTeethLand challenges, Teeth3DS+ supports multiple fundamental tasks, including tooth detection, segmentation, labeling, 3D modeling, and dental landmark identification. The dataset consists of rigorously curated intraoral scans acquired using state-of-the-art scanners and validated by experienced orthodontists and dental surgeons. In addition to the data, Teeth3DS+ provides standardized data splits and evaluation protocols to enable fair and reproducible comparison of methods, with the goal of fostering progress in learning-based analysis of 3D dental scans. Detailed instructions for accessing the dataset are available at https://crns-smartvision.github.io/teeth3ds

Method: Proposes HAPNet with hybrid asymmetric encoder combining VFM and CNN, dual-path progressive fusion, and auxiliary task to enrich local semantics of fused features for RGB-thermal scene parsing.

Result: Achieves superior performance compared to all state-of-the-art RGB-thermal scene parsing networks, ranking top across three widely used public datasets.

Conclusion: The new paradigm of leveraging vision foundation models in asymmetric encoder design opens up new opportunities for future developments in data-fusion scene parsing approaches.

Abstract: Data-fusion networks have shown significant promise for RGB-thermal scene parsing. However, the majority of existing studies have relied on symmetric duplex encoders for heterogeneous feature extraction and fusion, paying inadequate attention to the inherent differences between RGB and thermal modalities. Recent progress in vision foundation models (VFMs) trained through self-supervision on vast amounts of unlabeled data has proven their ability to extract informative, general-purpose features. However, this potential has yet to be fully leveraged in the domain. In this study, we take one step toward this new research area by exploring a feasible strategy to fully exploit VFM features for RGB-thermal scene parsing. Specifically, we delve deeper into the unique characteristics of RGB and thermal modalities, thereby designing a hybrid, asymmetric encoder that incorporates both a VFM and a convolutional neural network. This design allows for more effective extraction of complementary heterogeneous features, which are subsequently fused in a dual-path, progressive manner. Moreover, we introduce an auxiliary task to further enrich the local semantics of the fused features, thereby improving the overall performance of RGB-thermal scene parsing. Our proposed HAPNet, equipped with all these components, demonstrates superior performance compared to all other state-of-the-art RGB-thermal scene parsing networks, achieving top ranks across three widely used public RGB-thermal scene parsing datasets. We believe this new paradigm has opened up new opportunities for future developments in data-fusion scene parsing approaches.

Method: Proposes an efficient approach to estimate imitation thresholds without costly retraining from scratch. Experiments with human faces and art styles across four text-to-image models trained on three pretraining datasets.

Result: Estimates imitation thresholds in the range of 200-700 images, depending on domain (faces vs. art styles) and specific model architecture.

Conclusion: The imitation threshold provides empirical basis for copyright violation claims and serves as a guiding principle for model developers to comply with copyright and privacy laws.

Abstract: Text-to-image models are trained using large datasets of image-text pairs collected from the internet. These datasets often include copyrighted and private images. Training models on such datasets enables them to generate images that might violate copyright laws and individual privacy. This phenomenon is termed imitation – generation of images with content that has recognizable similarity to its training images. In this work we estimate the point at which a model was trained on enough instances of a concept to be able to imitate it – the imitation threshold. We posit this question as a new problem and propose an efficient approach that estimates the imitation threshold without incurring the colossal cost of training these models from scratch. We experiment with two domains – human faces and art styles, and evaluate four text-to-image models that were trained on three pretraining datasets. We estimate the imitation threshold of these models to be in the range of 200-700 images, depending on the domain and the model. The imitation threshold provides an empirical basis for copyright violation claims and acts as a guiding principle for text-to-image model developers that aim to comply with copyright and privacy laws. Website: https://how-many-van-goghs-does-it-take.github.io/. Code: https://github.com/vsahil/MIMETIC-2.

Method: Proposes FCC (Fully Connected Correlation) that integrates pixel-level correlations between support and query features across all layers in the encoder, capturing associations in both same-layers and cross-layers to reveal target-specific patterns and correspondences.

Result: Achieves state-of-the-art performance on PASCAL, COCO, and domain shift tests. Ablation studies and cross-layer correlation analysis validate FCC’s effectiveness in enhancing prior information and overall model performance.

Conclusion: FCC effectively addresses limitations of support masks by capturing previously inaccessible target information through comprehensive cross-layer correlations, significantly improving few-shot segmentation performance with Vision Transformer backbones.

Abstract: Few-shot segmentation (FSS) aims to segment the target object in a query image using only a small set of support images and masks. Therefore, having strong prior information for the target object using the support set is essential for guiding the initial training of FSS, which leads to the success of few-shot segmentation in challenging cases, such as when the target object shows considerable variation in appearance, texture, or scale across the support and query images. Previous methods have tried to obtain prior information by creating correlation maps from pixel-level correlation on final-layer or same-layer features. However, we found these approaches can offer limited and partial information when advanced models like Vision Transformers are used as the backbone. Vision Transformer encoders have a multi-layer structure with identical shapes in their intermediate layers. Leveraging the feature comparison from all layers in the encoder can enhance the performance of few-shot segmentation. We introduce FCC (Fully Connected Correlation) to integrate pixel-level correlations between support and query features, capturing associations that reveal target-specific patterns and correspondences in both same-layers and cross-layers. FCC captures previously inaccessible target information, effectively addressing the limitations of support mask. Our approach consistently demonstrates state-of-the-art performance on PASCAL, COCO, and domain shift tests. We conducted an ablation study and cross-layer correlation analysis to validate FCC’s core methodology. These findings reveal the effectiveness of FCC in enhancing prior information and overall model performance.

Method: Define part-based image hierarchy using object-level annotations, enforce hierarchy in hyperbolic space using contrastive loss with pairwise entailment metrics, and introduce new hierarchical retrieval evaluation metrics.

Result: Significant improvements in hierarchical image retrieval tasks, demonstrating the model’s capability to capture visual hierarchies beyond mere visual similarity.

Conclusion: The proposed approach successfully encodes user-defined multi-level visual hierarchies in hyperbolic space without requiring explicit hierarchical labels, enabling representations that capture semantic and structural information.

Abstract: Structuring latent representations in a hierarchical manner enables models to learn patterns at multiple levels of abstraction. However, most prevalent image understanding models focus on visual similarity, and learning visual hierarchies is relatively unexplored. In this work, for the first time, we introduce a learning paradigm that can encode user-defined multi-level complex visual hierarchies in hyperbolic space without requiring explicit hierarchical labels. As a concrete example, first, we define a part-based image hierarchy using object-level annotations within and across images. Then, we introduce an approach to enforce the hierarchy using contrastive loss with pairwise entailment metrics. Finally, we discuss new evaluation metrics to effectively measure hierarchical image retrieval. Encoding these complex relationships ensures that the learned representations capture semantic and structural information that transcends mere visual similarity. Experiments in part-based image retrieval show significant improvements in hierarchical retrieval tasks, demonstrating the capability of our model in capturing visual hierarchies.

Method: Proposed AdaVLN task extension, developed AdaVLN simulator with animated human models for Matterport3D, and introduced “freeze-time” mechanism for fair comparisons.

Result: Created AdaVLN simulator and AdaR2R datasets, evaluated baseline models, and demonstrated the task’s potential to bridge sim-to-real gap.

Conclusion: AdaVLN introduces realistic dynamic human obstacles to VLN, providing a more challenging benchmark that better reflects real-world navigation scenarios.

Abstract: Visual Language Navigation is a task that challenges robots to navigate in realistic environments based on natural language instructions. While previous research has largely focused on static settings, real-world navigation must often contend with dynamic human obstacles. Hence, we propose an extension to the task, termed Adaptive Visual Language Navigation (AdaVLN), which seeks to narrow this gap. AdaVLN requires robots to navigate complex 3D indoor environments populated with dynamically moving human obstacles, adding a layer of complexity to navigation tasks that mimic the real-world. To support exploration of this task, we also present AdaVLN simulator and AdaR2R datasets. The AdaVLN simulator enables easy inclusion of fully animated human models directly into common datasets like Matterport3D. We also introduce a “freeze-time” mechanism for both the navigation task and simulator, which pauses world state updates during agent inference, enabling fair comparisons and experimental reproducibility across different hardware. We evaluate several baseline models on this task, analyze the unique challenges introduced by AdaVLN, and demonstrate its potential to bridge the sim-to-real gap in VLN research.

Method: DenseSplat combines NeRF and 3DGS by using sparse keyframes and NeRF priors for primitive initialization, implementing geometry-aware primitive sampling/pruning strategies, and integrating loop closure and bundle adjustment.

Result: Extensive experiments on multiple large-scale datasets demonstrate superior tracking and mapping performance compared to current state-of-the-art methods.

Conclusion: DenseSplat effectively addresses sparse-view limitations in robotic SLAM by combining NeRF and 3DGS advantages, achieving practical deployment capabilities with improved accuracy and efficiency.

Abstract: Gaussian SLAM systems excel in real-time rendering and fine-grained reconstruction compared to NeRF-based systems. However, their reliance on extensive keyframes is impractical for deployment in real-world robotic systems, which typically operate under sparse-view conditions that can result in substantial holes in the map. To address these challenges, we introduce DenseSplat, the first SLAM system that effectively combines the advantages of NeRF and 3DGS. DenseSplat utilizes sparse keyframes and NeRF priors for initializing primitives that densely populate maps and seamlessly fill gaps. It also implements geometry-aware primitive sampling and pruning strategies to manage granularity and enhance rendering efficiency. Moreover, DenseSplat integrates loop closure and bundle adjustment, significantly enhancing frame-to-frame tracking accuracy. Extensive experiments on multiple large-scale datasets demonstrate that DenseSplat achieves superior performance in tracking and mapping compared to current state-of-the-art methods.

Method: Three-part framework: 1) GeoComp - large-scale dataset from geolocation game platform with 25M metadata entries and 3M geo-tagged locations; 2) GeoCoT - Geographical Chain-of-Thought multi-step reasoning framework for LVMs; 3) GeoEval - evaluation metric for assessment.

Result: GeoCoT significantly boosts geolocation accuracy by up to 25% while enhancing interpretability, as demonstrated using the GeoEval metric.

Conclusion: The proposed comprehensive framework addresses critical challenges in geolocation research by providing better data, reasoning methods, and evaluation metrics, enabling more precise and interpretable image localization.

Abstract: Geolocation, the task of identifying an image’s location, requires complex reasoning and is crucial for navigation, monitoring, and cultural preservation. However, current methods often produce coarse, imprecise, and non-interpretable localization. A major challenge lies in the quality and scale of existing geolocation datasets. These datasets are typically small-scale and automatically constructed, leading to noisy data and inconsistent task difficulty, with images that either reveal answers too easily or lack sufficient clues for reliable inference. To address these challenges, we introduce a comprehensive geolocation framework with three key components: GeoComp, a large-scale dataset; GeoCoT, a novel reasoning method; and GeoEval, an evaluation metric, collectively designed to address critical challenges and drive advancements in geolocation research. At the core of this framework is GeoComp (Geolocation Competition Dataset), a large-scale dataset collected from a geolocation game platform involving 740K users over two years. It comprises 25 million entries of metadata and 3 million geo-tagged locations spanning much of the globe, with each location annotated thousands to tens of thousands of times by human users. The dataset offers diverse difficulty levels for detailed analysis and highlights key gaps in current models. Building on this dataset, we propose Geographical Chain-of-Thought (GeoCoT), a novel multi-step reasoning framework designed to enhance the reasoning capabilities of Large Vision Models (LVMs) in geolocation tasks. GeoCoT improves performance by integrating contextual and spatial cues through a multi-step process that mimics human geolocation reasoning. Finally, using the GeoEval metric, we demonstrate that GeoCoT significantly boosts geolocation accuracy by up to 25% while enhancing interpretability.

Method: Proposes E²AT framework with: 1) efficient projector-based adversarial training module for visual attacks that aligns attack samples at feature level, and 2) Dynamic Joint Multimodal Optimization (DJMO) strategy that dynamically adjusts weights between normal and adversarial training objectives to enhance generalization.

Result: E²AT achieves state-of-the-art performance, outperforming existing baselines by average margin of 34% across text and image modalities on five major jailbreak attack methods across three mainstream MLLMs, while maintaining clean task performance. Real-world embodied intelligent system evaluations confirm practical applicability.

Conclusion: E²AT provides an effective solution for developing more secure and reliable multimodal systems by efficiently improving robustness against multimodal jailbreak attacks through end-to-end adversarial training with dynamic optimization.

Abstract: Research endeavors have been made in learning robust Multimodal Large Language Models (MLLMs) against jailbreak attacks. However, existing methods for improving MLLMs’ robustness still face critical challenges: \ding{172} how to efficiently tune massive weight parameters and \ding{173} how to ensure robustness against attacks across both visual and textual modalities. To this end, we propose an \textbf{E}fficient \textbf{E}nd-to-end \textbf{A}dversarial \textbf{T}raining (E$^2$AT) framework for both visual and textual adversarial attacks. Specifically, for the visual aspect, E$^2$AT incorporates an efficient projector-based AT module that aligns the attack samples at the feature level. For training objectives, we propose a Dynamic Joint Multimodal Optimization (DJMO) strategy to enhance generalization ability against jailbreak attacks by dynamically adjusting weights between normal and adversarial objectives. Extensive experiments are conducted with five major jailbreak attack methods across three mainstream MLLMs. Results demonstrate that our E$^2$AT achieves the state-of-the-art performance, outperforming existing baselines by an average margin of 34% across text and image modalities, while maintaining clean task performance. Furthermore, evaluations of real-world embodied intelligent systems highlight the practical applicability of E$^2$AT, paving the way for the development of more secure and reliable multimodal systems. Our code is available on \href{https://anonymous.4open.science/r/E2AT_568}{\textcolor{red}{https://anonymous.4open.science/r/E2AT_568}}.

Method: Spatial polarization multiplexing with quantized polarized light patterns using constrained de Bruijn sequences to encode incident rays and sample reflected light.

Result: Validated with real static and dynamic objects, achieving accurate shape and BRDF measurement while being invisible to the naked eye.

Conclusion: SPM opens new 3D sensing applications by enabling invisible, single-shot joint recovery of shape and radiometric properties.

Abstract: We propose spatial polarization multiplexing (SPM) for joint sensing of shape and reflectance of a static or dynamic deformable object, which is also invisible to the naked eye. Past structured-light methods are limited to shape acquisition and cannot recover reflectance as they alter scene appearance. Our key idea is to spatially multiplex a polarization pattern to encode the incident ray and also densely sample the reflected light. We derive a quantized polarized light pattern that can be robustly and uniquely decoded from the reflected Angle of Linear Polarization (AoLP) values. It also enables single-shot disentanglement of polarimetric diffuse and specular reflections for accurate BRDF estimation. We achieve this spatial polarization multiplexing (SPM) with a constrained de Bruijn sequence. We validate this novel invisible single-shot shape and reflectance method with real static and dynamic objects. The results demonstrate the effectiveness of SPM for accurate shape and BRDF measurement which opens new avenues of application for 3D sensing thanks to its invisibility and ability to jointly recover the radiometric properties.

Method: Three-stage approach: 1) Construct unified latent representation encoding heterogeneous pose formats (SMPLer-X, DWPose, Mediapipe, PrimeDepth, Sapiens Segmentation); 2) Train ViT-based Video2Pose module to extract latent representation directly from raw videos; 3) Develop temporal modeling and sequence refinement operating entirely in latent space.

Result: SignX achieves state-of-the-art accuracy on continuous sign language recognition while significantly reducing computational consumption compared to existing methods.

Conclusion: The proposed multi-stage design enables end-to-end sign language recognition in a compact pose-rich latent space, offering an efficient and accurate solution for continuous sign language recognition challenges.

Abstract: The complexity of sign language data processing brings many challenges. The current approach to recognition of ASL signs aims to translate RGB sign language videos through pose information into English-based ID Glosses, which serve to uniquely identify ASL signs. This paper proposes SignX, a novel framework for continuous sign language recognition in compact pose-rich latent space. First, we construct a unified latent representation that encodes heterogeneous pose formats (SMPLer-X, DWPose, Mediapipe, PrimeDepth, and Sapiens Segmentation) into a compact, information-dense space. Second, we train a ViT-based Video2Pose module to extract this latent representation directly from raw videos. Finally, we develop a temporal modeling and sequence refinement method that operates entirely in this latent space. This multi-stage design achieves end-to-end sign language recognition while significantly reducing computational consumption. Experimental results demonstrate that SignX achieves state-of-the-art accuracy on continuous sign language recognition.

Method: PartHOI uses generalized cylinder representations to parameterize object parts’ geometry, establishes robust geometric correspondence between object parts, transfers contact points, and optimizes hand poses to fit target objects.

Result: Qualitative and quantitative results show that PartHOI generalizes HOI transfers well even for cross-category objects, producing high-fidelity results superior to existing methods.

Conclusion: The part-based approach with semantic part correspondences enables effective cross-category hand-object interaction transfer, overcoming limitations of shape-matching methods.

Abstract: Learning-based methods to understand and model hand-object interactions (HOI) require a large amount of high-quality HOI data. One way to create HOI data is to transfer hand poses from a source object to another based on the objects’ geometry. However, current methods for transferring hand poses between objects rely on shape matching, limiting the ability to transfer poses across different categories due to differences in their shapes and sizes. We observe that HOI often involves specific semantic parts of objects, which often have more consistent shapes across categories. In addition, constructing size-invariant correspondences between these parts is important for cross-category transfer. Based on these insights, we introduce a novel method PartHOI for part-based HOI transfer. Using a generalized cylinder representation to parameterize an object parts’ geometry, PartHOI establishes a robust geometric correspondence between object parts, and enables the transfer of contact points. Given the transferred points, we optimize a hand pose to fit the target object well. Qualitative and quantitative results demonstrate that our method can generalize HOI transfers well even for cross-category objects, and produce high-fidelity results that are superior to the existing methods.

Method: Visualize-then-Retrieve (VisRet) first projects textual queries into the image modality using text-to-image generation, then performs retrieval within the image modality (image-to-image retrieval) to bypass the limitations of cross-modal retrievers.

Result: VisRet substantially outperforms cross-modal similarity matching across four benchmarks, improving nDCG@30 by 0.125 on average with CLIP and 0.121 with E5-V. For downstream QA, it increases accuracy on Visual-RAG and Visual-RAG-ME by 3.8% and 15.7% in top-1 retrieval.

Conclusion: VisRet provides a simple yet effective paradigm for advancing text-image retrieval by leveraging text-to-image generation to bridge the modality gap, showing strong performance across multiple benchmarks and compatibility with various models.

Abstract: Text-to-image retrieval (T2I retrieval) remains challenging because cross-modal embeddings often behave as bags of concepts, underrepresenting structured visual relationships such as pose and viewpoint. We propose Visualize-then-Retrieve (VisRet), a retrieval paradigm that mitigates this limitation of cross-modal similarity alignment. VisRet first projects textual queries into the image modality via T2I generation, then performs retrieval within the image modality to bypass the weaknesses of cross-modal retrievers in recognizing subtle visual-spatial features. Across four benchmarks (Visual-RAG, INQUIRE-Rerank, Microsoft COCO, and our new Visual-RAG-ME featuring multi-entity comparisons), VisRet substantially outperforms cross-modal similarity matching and baselines that recast T2I retrieval as text-to-text similarity matching, improving nDCG@30 by 0.125 on average with CLIP as the retriever and by 0.121 with E5-V. For downstream question answering, VisRet increases accuracy on Visual-RAG and Visual-RAG-ME by 3.8% and 15.7% in top-1 retrieval, and by 3.9% and 11.1% in top-10 retrieval. Ablation studies show compatibility with different T2I instruction LLMs, T2I generation models, and downstream LLMs. VisRet provides a simple yet effective perspective for advancing in text-image retrieval. Our code and the new benchmark are publicly available at https://github.com/xiaowu0162/Visualize-then-Retrieve.

Method: Diffusion-based video generation framework using cross-view feature interactions and globally consistent 3D geometry to ensure multi-view geometric consistency while allowing fine-grained control over scene elements like background editing and object replacement.

Result: RoboTransfer produces videos with superior geometric consistency and visual fidelity. Policies trained on this synthetic data show enhanced generalization to novel, unseen scenarios.

Conclusion: RoboTransfer provides an effective solution for scalable robotic data generation, bridging the sim-to-real gap and enabling better policy generalization through geometrically consistent synthetic video generation.

Abstract: The goal of general-purpose robotics is to create agents that can seamlessly adapt to and operate in diverse, unstructured human environments. Imitation learning has become a key paradigm for robotic manipulation, yet collecting large-scale and diverse demonstrations is prohibitively expensive. Simulators provide a cost-effective alternative, but the sim-to-real gap remains a major obstacle to scalability. We present RoboTransfer, a diffusion-based video generation framework for synthesizing robotic data. By leveraging cross-view feature interactions and globally consistent 3D geometry, RoboTransfer ensures multi-view geometric consistency while enabling fine-grained control over scene elements, such as background editing and object replacement. Extensive experiments demonstrate that RoboTransfer produces videos with superior geometric consistency and visual fidelity. Furthermore, policies trained on this synthetic data exhibit enhanced generalization to novel, unseen scenarios. Project page: https://horizonrobotics.github.io/robot_lab/robotransfer.

Method: Proposes Decision Variable Correlation (DVC) which quantifies image-by-image correlation between decoded decisions based on internal neural representations in classification tasks. Evaluated using monkey V4/IT recordings and network models trained on image classification tasks.

Result: Model-model similarity is comparable to monkey-monkey similarity, but model-monkey similarity is consistently lower. DVC decreases with increasing network performance on ImageNet-1k. Adversarial training and pre-training on larger datasets do not improve model-monkey similarity in task-relevant dimensions.

Conclusion: There’s a divergence between task-relevant representations in monkey V4/IT and those learned by models trained on image classification tasks, suggesting current models don’t capture the biological decision strategies despite good performance on benchmark tasks.

Abstract: Previous studies have compared neural activities in the visual cortex to representations in deep neural networks trained on image classification. Interestingly, while some suggest that their representations are highly similar, others argued the opposite. Here, we propose a new approach to characterize the similarity of the decision strategies of two observers (models or brains) using decision variable correlation (DVC). DVC quantifies the image-by-image correlation between the decoded decisions based on the internal neural representations in a classification task. Thus, it can capture task-relevant information rather than general representational alignment. We evaluate DVC using monkey V4/IT recordings and network models trained on image classification tasks. We find that model-model similarity is comparable to monkey-monkey similarity, whereas model-monkey similarity is consistently lower. Strikingly, DVC decreases with increasing network performance on ImageNet-1k. Adversarial training does not improve model-monkey similarity in task-relevant dimensions assessed using DVC, although it markedly increases the model-model similarity. Similarly, pre-training on larger datasets does not improve model-monkey similarity. These results suggest a divergence between the task-relevant representations in monkey V4/IT and those learned by models trained on image classification tasks.

Method: Proposes a unified LLM framework that aligns language, images, and 3D scenes with detailed design guidelines for data representation and modality-specific objectives. Includes tokenization methods for complex 3D objects.

Result: Evaluated on four core 3D tasks (rendering, recognition, instruction-following, question-answering) across four datasets. Shows effectiveness in reconstructing complete 3D scenes from single images and real-world 3D object recognition.

Conclusion: Demonstrates the potential of autoregressive models for structured 3D scene understanding, providing a comprehensive framework for multimodal 3D AI applications.

Abstract: Creating machines capable of understanding the world in 3D is essential in assisting designers that build and edit 3D environments and robots navigating and interacting within a three-dimensional space. Inspired by advances in language and image modeling, we investigate the potential of autoregressive models for a new modality: structured 3D scenes. To this end, we propose a unified LLM framework that aligns language, images, and 3D scenes and provide a detailed ‘‘cookbook’’ outlining critical design choices for achieving optimal training and performance addressing key questions related to data representation, modality-specific objectives, and more. We show how to tokenize complex 3D objects to incorporate into our structured 3D scene modality. We evaluate performance across four core 3D tasks – rendering, recognition, instruction-following, and question-answering – and four 3D datasets, synthetic and real-world. We show our model’s effectiveness on reconstructing complete 3D scenes consisting of complex objects from a single image and on real-world 3D object recognition tasks. Project webpage: https://glab-caltech.github.io/kyvo/

Method: Introduces Discrete Latent Codes (DLCs) - sequences of discrete tokens derived from Simplicial Embeddings trained with self-supervised learning. DLCs are used to condition diffusion models instead of standard continuous embeddings. Also shows how to finetune text diffusion language models to generate DLCs for text-to-image generation.

Result: Diffusion models trained with DLCs achieve new state-of-the-art for unconditional image generation on ImageNet. DLC composition enables coherent out-of-distribution sampling that combines semantics in diverse ways. Text-to-image generation is enabled by finetuning language models to generate DLCs.

Conclusion: DLCs provide superior conditioning representations for diffusion models, improving fidelity while enabling compositional generation and out-of-distribution sampling, with applications to text-to-image generation.

Abstract: We argue that diffusion models’ success in modeling complex distributions is, for the most part, coming from their input conditioning. This paper investigates the representation used to condition diffusion models from the perspective that ideal representations should improve sample fidelity, be easy to generate, and be compositional to allow out-of-training samples generation. We introduce Discrete Latent Code (DLC), an image representation derived from Simplicial Embeddings trained with a self-supervised learning objective. DLCs are sequences of discrete tokens, as opposed to the standard continuous image embeddings. They are easy to generate and their compositionality enables sampling of novel images beyond the training distribution. Diffusion models trained with DLCs have improved generation fidelity, establishing a new state-of-the-art for unconditional image generation on ImageNet. Additionally, we show that composing DLCs allows the image generator to produce out-of-distribution samples that coherently combine the semantics of images in diverse ways. Finally, we showcase how DLCs can enable text-to-image generation by leveraging large-scale pretrained language models. We efficiently finetune a text diffusion language model to generate DLCs that produce novel samples outside of the image generator training distribution.

Method: Introduces BusterX++ framework with unified detection and explanation of synthetic images/videos using direct reinforcement learning post-training strategy. Also creates GenBuster++ benchmark with 4,000 human-curated images/video clips from state-of-the-art generation techniques.

Result: Extensive experiments demonstrate the effectiveness and generalizability of the approach. The unified framework outperforms single-modality detection systems against multimodal synthetic content.

Conclusion: BusterX++ addresses critical limitations in current synthetic media detection by providing a unified multimodal approach with enhanced transparency and interpretability, offering a more robust solution against sophisticated misinformation threats.

Abstract: Recent advances in generative AI have dramatically improved image and video synthesis capabilities, significantly increasing the risk of misinformation through sophisticated fake content. In response, detection methods have evolved from traditional approaches to multimodal large language models (MLLMs), offering enhanced transparency and interpretability in identifying synthetic media. However, current detection systems remain fundamentally limited by their single-modality design. These approaches analyze images or videos separately, making them ineffective against synthetic content that combines multiple media formats. To address these challenges, we introduce \textbf{BusterX++}, a framework for unified detection and explanation of synthetic image and video, with a direct reinforcement learning (RL) post-training strategy. To enable comprehensive evaluation, we also present \textbf{GenBuster++}, a unified benchmark leveraging state-of-the-art image and video generation techniques. This benchmark comprises 4,000 images and video clips, meticulously curated by human experts to ensure high quality, diversity, and real-world applicability. Extensive experiments demonstrate the effectiveness and generalizability of our approach.

Method: SAGOnline decouples segmentation into lightweight sub-tasks: 1) Uses video foundation models (SAM 2) to generate temporally consistent 2D masks across rendered views, 2) Introduces Rasterization-aware Geometric Consensus mechanism that leverages the traceability of Gaussian rasterization to deterministically map 2D predictions to explicit 3D primitive labels in real-time, eliminating the need for feature distillation.

Result: Achieves state-of-the-art accuracy on NVOS (92.7% mIoU) and SPIn-NeRF (95.2% mIoU) benchmarks while operating at 27 ms per frame - the fastest speed. Supports instant prompt, instance, and semantic segmentation through flexible foundation model interface.

Conclusion: SAGOnline provides a unified, zero-shot framework for real-time, cross-view consistent 3D segmentation without scene-specific training, enabling interactive 3D understanding applications in AR/VR and robotics through its efficient discrete representation and flexible foundation model integration.

Abstract: 3D Gaussian Splatting has emerged as a powerful paradigm for explicit 3D scene representation, yet achieving efficient and consistent 3D segmentation remains challenging. Existing segmentation approaches typically rely on high-dimensional feature lifting, which causes costly optimization, implicit semantics, and task-specific constraints. We present \textbf{Segment Any Gaussians Online (SAGOnline)}, a unified, zero-shot framework that achieves real-time, cross-view consistent segmentation without scene-specific training. SAGOnline decouples the monolithic segmentation problem into lightweight sub-tasks. By integrating video foundation models (e.g., SAM 2), we first generate temporally consistent 2D masks across rendered views. Crucially, instead of learning continuous feature fields, we introduce a \textbf{Rasterization-aware Geometric Consensus} mechanism that leverages the traceability of the Gaussian rasterization pipeline. This allows us to deterministically map 2D predictions to explicit, discrete 3D primitive labels in real-time. This discrete representation eliminates the memory and computational burden of feature distillation, enabling instant inference. Extensive evaluations on NVOS and SPIn-NeRF benchmarks demonstrate that SAGOnline achieves state-of-the-art accuracy (92.7% and 95.2% mIoU) while operating at the fastest speed at 27 ms per frame. By providing a flexible interface for diverse foundation models, our framework supports instant prompt, instance, and semantic segmentation, paving the way for interactive 3D understanding in AR/VR and robotics.

Method: Train a lightweight LVLM-aware multimodal retriever that learns to return images and texts guiding the LVLM toward correct predictions. Use only lightweight fine-tuning with small data amounts and general-purpose backbone models (no extensive medical pre-training).

Result: Achieves competitive results in clinical classification and VQA tasks compared to medically pre-trained models with extensive training. Identifies and significantly improves inconsistent retrieval prediction errors (cases where different top-retrieved images yield different predictions for same target).

Conclusion: The lightweight retrieval optimization mechanism effectively enhances diagnostic performance and addresses challenging inconsistent retrieval cases, but reveals gaps in LVLMs’ ability to utilize retrieved information for clinical predictions, highlighting areas for future improvement.

Abstract: Retrieving visual and textual information from medical literature and hospital records can enhance diagnostic accuracy for clinical image interpretation. However, multimodal retrieval-augmented diagnosis is highly challenging. We explore a lightweight mechanism for enhancing diagnostic performance of retrieval-augmented LVLMs. We train a lightweight LVLM-aware multimodal retriever, such that the retriever learns to return images and texts that guide the LVLM toward correct predictions. In our low-resource setting, we perform only lightweight fine-tuning with small amounts of data, and use only general-purpose backbone models, achieving competitive results in clinical classification and VQA tasks compared to medically pre-trained models with extensive training. In a novel analysis, we highlight a previously unexplored class of errors that we term inconsistent retrieval predictions: cases where different top-retrieved images yield different predictions for the same target. We find that these cases are challenging for all models, even for non-retrieval models, and that our retrieval optimization mechanism significantly improves these cases over standard RAG. However, our analysis also sheds light on gaps in the ability of LVLMs to utilize retrieved information for clinical predictions. Code and models available at: https://github.com/Nirmaz/JOMED.

Method: Created CVBench with 1,000 QA pairs across three hierarchical tiers: cross-video object association, cross-video event association, and cross-video complex reasoning. Built from five domain-diverse video clusters and evaluated 10+ leading MLLMs under zero-shot or chain-of-thought prompting.

Result: Significant performance gaps: top models like GPT-4o achieve only 63.5% accuracy on causal reasoning tasks vs. 91.3% human accuracy. Identified fundamental bottlenecks in current MLLM architectures including deficient inter-video context retention and poor disambiguation of overlapping entities.

Conclusion: CVBench establishes a rigorous framework for advancing pattern recognition in multi-video scenarios and provides architectural insights for next-generation models, highlighting critical gaps in current MLLM capabilities for cross-video relational reasoning.

Abstract: While multimodal large language models (MLLMs) exhibit strong performance on single-video tasks (e.g., video question answering), their capability for spatiotemporal pattern reasoning across multiple videos remains a critical gap in pattern recognition research. However, this capability is essential for real-world applications, including multi-camera surveillance and cross-video procedural learning. To bridge this gap, we present CVBench, the first diagnostic benchmark designed to assess cross-video relational reasoning rigorously. CVBench comprises 1,000 question-answer pairs spanning three hierarchical tiers: cross-video object association (identifying shared entities), cross-video event association (linking temporal or causal event chains), and cross-video complex reasoning (integrating commonsense and domain knowledge). Built from five domain-diverse video clusters (e.g., sports, life records), the benchmark challenges models to analyze and integrate spatiotemporal patterns from dynamic visual streams. Extensive evaluation of 10+ leading MLLMs (including GPT-4o, Gemini-2.0-flash, Qwen2.5-VL) under zero-shot or chain-of-thought prompting paradigms. Key findings reveal stark performance gaps: even top models, such as GPT-4o, achieve only 63.5% accuracy on causal reasoning tasks, compared to the 91.3% accuracy of human performance. Crucially, our analysis reveals fundamental bottlenecks inherent in current MLLMs architectures, notably deficient inter-video context retention and poor disambiguation of overlapping entities. CVBench establishes a rigorous framework for advancing pattern recognition methodologies in multi-video scenarios, providing architectural insights for next-generation models. The data and evaluation code are available at: https://github.com/Hokhim2/CVBench.

Method: Uses a lightweight symmetric two-view association (STA) model as frontend that simultaneously estimates relative camera poses and regresses local pointmaps from two RGB images. Backend employs a specially designed Sim(3) pose graph with loop closure to address accumulated drift.

Result: Achieves superior performance in both camera tracking and dense 3D reconstruction quality compared to current methods, with frontend size only 35% of comparable state-of-the-art methods.

Conclusion: ViSTA-SLAM demonstrates an effective approach for real-time monocular visual SLAM that eliminates the need for camera intrinsics while maintaining high performance with reduced model complexity.

Abstract: We present ViSTA-SLAM as a real-time monocular visual SLAM system that operates without requiring camera intrinsics, making it broadly applicable across diverse camera setups. At its core, the system employs a lightweight symmetric two-view association (STA) model as the frontend, which simultaneously estimates relative camera poses and regresses local pointmaps from only two RGB images. This design reduces model complexity significantly, the size of our frontend is only 35% that of comparable state-of-the-art methods, while enhancing the quality of two-view constraints used in the pipeline. In the backend, we construct a specially designed Sim(3) pose graph that incorporates loop closures to address accumulated drift. Extensive experiments demonstrate that our approach achieves superior performance in both camera tracking and dense 3D reconstruction quality compared to current methods. Github repository: https://github.com/zhangganlin/vista-slam

Method: Benchmarks convolutional and transformer-based object detectors on UAV-captured RGB imagery, introduces class-targeted augmentation strategy applied only to training split, evaluates Faster R-CNN (ResNet50, MobileNetV3), RetinaNet (ResNet50), YOLOv5, YOLOv8, and Swin Transformer variants with Faster R-CNN.

Result: Faster R-CNN with ResNet50 achieves highest localization accuracy (mAP@0.5: 0.893, mAP@0.5:0.95: 0.759), while MobileNetV3 variant provides best overall reliability balance (recall: 0.745, F1-score: 0.809, accuracy: 0.679).

Conclusion: The study provides comprehensive benchmarking for solar panel defect detection, enabling accuracy-throughput tradeoff analysis relevant to UAV deployment, with dataset and code to be released upon acceptance.

Abstract: Timely and accurate detection of defects and contaminants in solar panels is critical for maintaining the efficiency and reliability of photovoltaic (PV) systems. While recent studies have applied deep learning to PV inspection, fair benchmarking across detector architectures and unbiased handling of class imbalance remain limited. This work presents a comprehensive benchmark of convolutional and transformer-based object detectors on UAV-captured RGB imagery of solar panels. It introduces a class-targeted augmentation strategy applied exclusively to the training split to mitigate imbalance without compromising evaluation integrity. Faster R-CNN with ResNet50 and MobileNetV3 backbones, RetinaNet with ResNet50, YOLOv5, YOLOv8, and Swin Transformer backbones integrated with Faster R-CNN (Tiny, Small, and Base variants) are evaluated. Performance is assessed using mean Average Precision (mAP) across multiple IoU thresholds, precision, recall, F1 score, and inference throughput to enable accuracy-throughput tradeoff analysis relevant to UAV deployment. Experimental results show that Faster R-CNN with a ResNet50 backbone achieves the highest localization accuracy, with mAP@0.5 of 0.893 and mAP@0.5:0.95 of 0.759, whereas the MobileNetV3 variant provides the best overall reliability balance, achieving recall of 0.745, F1-score of 0.809, and accuracy of 0.679 on the test set. The dataset and code will be released upon acceptance of the paper.

Method: Proposes a dataset-agnostic method using intrinsic parameter statistics (weight variances and biases) to estimate channel importance, revealing Invariant Salient Channel Space (ISCS) with Salient-Core and Salient-Auxiliary channels. Uses deterministic static pruning for lightweight split-computing.

Result: Achieves deterministic, ultra-low latency pipeline by bypassing heavy entropy modeling. Reduces end-to-end latency while providing critical speed-accuracy trade-off for resource-constrained systems.

Conclusion: The method enables efficient deployment of visual models on resource-constrained embodied AI systems by exploiting redundancy in latent representations through dataset-agnostic channel importance analysis and static pruning.

Abstract: Prior studies in embodied AI consistently show that robust perception is critical for human-robot interaction, yet deploying high-fidelity visual models on resource-constrained agents remains challenging due to limited on-device computation power and transmission latency. Exploiting the redundancy in latent representations could improve system efficiency, yet existing approaches often rely on costly dataset-specific ablation tests or heavy entropy models unsuitable for real-time edge-robot collaboration. We propose a generalizable, dataset-agnostic method to identify and selectively transmit structure-critical channels in pretrained encoders. Instead of brute-force empirical evaluations, our approach leverages intrinsic parameter statistics-weight variances and biases-to estimate channel importance. This analysis reveals a consistent organizational structure, termed the Invariant Salient Channel Space (ISCS), where Salient-Core channels capture dominant structures while Salient-Auxiliary channels encode fine visual details. Building on ISCS, we introduce a deterministic static pruning strategy that enables lightweight split-computing. Experiments across different datasets demonstrate that our method achieves a deterministic, ultra-low latency pipeline by bypassing heavy entropy modeling. Our method reduces end-to-end latency, providing a critical speed-accuracy trade-off for resource-constrained human-aware embodied systems.

Method: Proposed a domain-agnostic framework that systematically suppresses shape, texture, and color cues without forced-choice conflicts. Evaluated both humans and neural networks under controlled suppression conditions across computer vision, medical imaging, and remote sensing domains.

Result: CNNs are not inherently texture-biased but predominantly rely on local shape features. This reliance can be substantially mitigated through modern training strategies or architectures (ConvNeXt, ViTs). Feature reliance patterns differ systematically across domains: computer vision models prioritize shape, medical imaging models emphasize color, and remote sensing models show stronger texture reliance.

Conclusion: The texture-bias hypothesis for CNNs is incomplete; CNNs primarily use local shape features, and feature reliance patterns vary systematically across different application domains, suggesting domain-specific feature importance in model design and evaluation.

Abstract: The hypothesis that Convolutional Neural Networks (CNNs) are inherently texture-biased has shaped much of the discourse on feature use in deep learning. We revisit this hypothesis by examining limitations in the cue-conflict experiment by Geirhos et al. To address these limitations, we propose a domain-agnostic framework that quantifies feature reliance through systematic suppression of shape, texture, and color cues, avoiding the confounds of forced-choice conflicts. By evaluating humans and neural networks under controlled suppression conditions, we find that CNNs are not inherently texture-biased but predominantly rely on local shape features. Nonetheless, this reliance can be substantially mitigated through modern training strategies or architectures (ConvNeXt, ViTs). We further extend the analysis across computer vision, medical imaging, and remote sensing, revealing that reliance patterns differ systematically: computer vision models prioritize shape, medical imaging models emphasize color, and remote sensing models exhibit a stronger reliance on texture. Code is available at https://github.com/tomburgert/feature-reliance.

Method: ScalingAR introduces token entropy as a novel signal in visual token generation and operates at two levels: Profile Level (streams calibrated confidence state by fusing intrinsic and conditional signals) and Policy Level (adaptively terminates low-confidence trajectories and dynamically schedules guidance for phase-appropriate conditioning strength).

Result: ScalingAR improves base models by 12.5% on GenEval and 15.2% on TIIF-Bench, reduces visual token consumption by 62.0% while outperforming baselines, and enhances robustness by mitigating performance drops by 26.0% in challenging scenarios.

Conclusion: ScalingAR successfully bridges the gap in applying test-time scaling to next-token prediction autoregressive image generation, eliminating the need for early decoding or auxiliary rewards while significantly improving generation quality and efficiency.

Abstract: Test-time scaling (TTS) has demonstrated remarkable success in enhancing large language models, yet its application to next-token prediction (NTP) autoregressive (AR) image generation remains largely uncharted. Existing TTS approaches for visual AR (VAR), which rely on frequent partial decoding and external reward models, are ill-suited for NTP-based image generation due to the inherent incompleteness of intermediate decoding results. To bridge this gap, we introduce ScalingAR, the first TTS framework specifically designed for NTP-based AR image generation that eliminates the need for early decoding or auxiliary rewards. ScalingAR leverages token entropy as a novel signal in visual token generation and operates at two complementary scaling levels: (i) Profile Level, which streams a calibrated confidence state by fusing intrinsic and conditional signals; and (ii) Policy Level, which utilizes this state to adaptively terminate low-confidence trajectories and dynamically schedule guidance for phase-appropriate conditioning strength. Experiments on both general and compositional benchmarks show that ScalingAR (1) improves base models by 12.5% on GenEval and 15.2% on TIIF-Bench, (2) efficiently reduces visual token consumption by 62.0% while outperforming baselines, and (3) successfully enhances robustness, mitigating performance drops by 26.0% in challenging scenarios.

Method: LaRe combines visual refocusing with rich latent representations to enable iterative reasoning within latent space. It uses semantic augmentation training with joint alignment and reconstruction objectives to enhance semantic structure of latent space.

Result: LaRe improves average accuracy by 9.4% compared to existing baselines while reducing inference tokens by 16.5%. When scaled to 7B-parameter LLM backbone, it achieves performance comparable to SOTA models and outperforms larger-scale models on almost all benchmarks.

Conclusion: LaRe effectively addresses limitations of existing multimodal reasoning methods by enabling iterative latent space reasoning with visual refocusing, achieving better performance with fewer tokens while maintaining interpretability.

Abstract: Chain of Thought (CoT) reasoning enhances logical performance by decomposing complex tasks, yet its multimodal extension faces a trade-off. The existing Thinking with Images paradigm is limited by the modality gap between vision and language, which hinders reliable extraction of reasoning relevant information from high dimensional visual data. Recent latent space reasoning method provides stronger multimodal representations, but it often lacks the ability to refocus on visual inputs and suffers from limited interpretability. To address these issues, we propose \underline{La}tent \underline{Re}focusing (LaRe), a novel multimodal reasoning paradigm that combines visual refocusing with rich latent representations, enabling iterative reasoning within the latent space. We further design a semantic augmentation training strategy that enhances the semantic structure of the latent space through joint alignment and reconstruction objectives. Experimental evaluations demonstrate that LaRe improves average accuracy by 9.4% compared to existing baselines while reducing the number of tokens required for inference by 16.5%. When scaled to a 7B-parameter Large Language Model backbone, LaRe achieves performance comparable to state-of-the-art models and outperforms larger-scale models on almost all benchmarks. Code and checkpoints will be released later.

Method: Reformulates reaction diagram parsing as image captioning problem using Large Vision Language Models (LVLMs). Introduces BBox and Index as Visual Prompt (BIVP) strategy that uses MolYOLO detector to pre-draw molecular bounding boxes and indices onto input images.

Result: BIVP strategy significantly improves structural extraction quality while simplifying model design. RxnCaption-VL achieves state-of-the-art performance on multiple metrics. Created RxnCaption-15k dataset (10x larger than prior benchmarks) with balanced test subset across four layout archetypes.

Conclusion: The method, dataset, and models will advance structured information extraction from chemical literature and catalyze broader AI applications in chemistry. All resources will be released on GitHub.

Abstract: Large-scale chemical reaction datasets are crucial for AI research in chemistry. However, existing chemical reaction data often exist as images within papers, making them not machine-readable and unusable for training machine learning models. In response to this challenge, we propose the RxnCaption framework for the task of chemical Reaction Diagram Parsing (RxnDP). Our framework reformulates the traditional coordinate prediction driven parsing process into an image captioning problem, which Large Vision Language Models (LVLMs) handle naturally. We introduce a strategy termed BBox and Index as Visual Prompt (BIVP), which uses our state-of-the-art molecular detector, MolYOLO, to pre-draw molecular bounding boxes and indices directly onto the input image. This turns the downstream parsing into a natural-language description problem. Extensive experiments show that the BIVP strategy significantly improves structural extraction quality while simplifying model design. We further construct the RxnCaption-15k dataset, an order of magnitude larger than prior real-world literature benchmarks, with a balanced test subset across four layout archetypes. Experiments demonstrate that RxnCaption-VL achieves state-of-the-art performance on multiple metrics. We believe our method, dataset, and models will advance structured information extraction from chemical literature and catalyze broader AI applications in chemistry. We will release data, models, and code on GitHub.

Method: GlobalDiff operates directly in global joint rotation space using diffusion models. It introduces three constraints: 1) joint structure constraint with virtual anchor points for fine-grained orientation, 2) skeleton structure constraint for angular consistency across bones, and 3) temporal structure constraint using multi-scale variational encoder to align with ground-truth temporal patterns.

Result: Extensive evaluations on standard co-speech benchmarks show GlobalDiff generates smooth and accurate motions, improving performance by 46.0% compared to current state-of-the-art methods under multiple speaker identities.

Conclusion: Operating in global rotation space with multi-level constraints effectively addresses hierarchical error accumulation in co-speech motion generation, producing more stable and plausible motions than existing methods.

Abstract: Reliable co-speech motion generation requires precise motion representation and consistent structural priors across all joints. Existing generative methods typically operate on local joint rotations, which are defined hierarchically based on the skeleton structure. This leads to cumulative errors during generation, manifesting as unstable and implausible motions at end-effectors. In this work, we propose GlobalDiff, a diffusion-based framework that operates directly in the space of global joint rotations for the first time, fundamentally decoupling each joint’s prediction from upstream dependencies and alleviating hierarchical error accumulation. To compensate for the absence of structural priors in global rotation space, we introduce a multi-level constraint scheme. Specifically, a joint structure constraint introduces virtual anchor points around each joint to better capture fine-grained orientation. A skeleton structure constraint enforces angular consistency across bones to maintain structural integrity. A temporal structure constraint utilizes a multi-scale variational encoder to align the generated motion with ground-truth temporal patterns. These constraints jointly regularize the global diffusion process and reinforce structural awareness. Extensive evaluations on standard co-speech benchmarks show that GlobalDiff generates smooth and accurate motions, improving the performance by 46.0 % compared to the current SOTA under multiple speaker identities.

Method: Used 667 CT scans from COCA dataset to generate synthetic DRRs. Evaluated model capacity, super-resolution fidelity enhancement, preprocessing, and training strategies. Compared lightweight CNNs trained from scratch vs. large pretrained networks, tested super-resolution with contrast enhancement, and implemented curriculum learning for weak supervision.

Result: Lightweight CNNs trained from scratch outperformed large pretrained networks. Pairing super-resolution with contrast enhancement yielded significant performance gains. Curriculum learning stabilized training under weak supervision. Best configuration achieved mean AUC of 0.754, comparable to or exceeding prior CXR-based studies.

Conclusion: DRRs establish a scalable, label-rich foundation for CAC detection, providing a viable alternative to expensive CT screening. The work lays groundwork for future transfer learning and domain adaptation to real chest X-rays, potentially enabling large-scale cardiovascular screening.

Abstract: Coronary artery calcification (CAC) is a strong predictor of cardiovascular events, with CT-based Agatston scoring widely regarded as the clinical gold standard. However, CT is costly and impractical for large-scale screening, while chest X-rays (CXRs) are inexpensive but lack reliable ground truth labels, constraining deep learning development. Digitally reconstructed radiographs (DRRs) offer a scalable alternative by projecting CT volumes into CXR-like images while inheriting precise labels. In this work, we provide the first systematic evaluation of DRRs as a surrogate training domain for CAC detection. Using 667 CT scans from the COCA dataset, we generate synthetic DRRs and assess model capacity, super-resolution fidelity enhancement, preprocessing, and training strategies. Lightweight CNNs trained from scratch outperform large pretrained networks; pairing super-resolution with contrast enhancement yields significant gains; and curriculum learning stabilises training under weak supervision. Our best configuration achieves a mean AUC of 0.754, comparable to or exceeding prior CXR-based studies. These results establish DRRs as a scalable, label-rich foundation for CAC detection, while laying the foundation for future transfer learning and domain adaptation to real CXRs.

Method: Two-branch framework: 1) Class-agnostic expression intensity branch with Gaussian-based instance-adaptive intensity modeling (GIM) for soft pseudo-labeling, 2) Class-aware apex classification branch for macro/micro-expression distinction. Also includes intensity-aware contrastive loss for feature learning.

Result: Extensive experiments on SAMM-LV, CAS(ME)^2, and CAS(ME)^3 datasets demonstrate the framework’s effectiveness for point-supervised facial expression spotting.

Conclusion: The proposed point-supervised framework successfully reduces annotation burden while maintaining performance for facial expression spotting, handling both macro- and micro-expressions through innovative intensity modeling and classification branches.

Abstract: Automatic facial expression spotting, which aims to identify facial expression instances in untrimmed videos, is crucial for facial expression analysis. Existing methods primarily focus on fully-supervised learning and rely on costly, time-consuming temporal boundary annotations. In this paper, we investigate point-supervised facial expression spotting (P-FES), where only a single timestamp annotation per instance is required for training. We propose a unique two-branch framework for P-FES. First, to mitigate the limitation of hard pseudo-labeling, which often confuses neutral and expression frames with various intensities, we propose a Gaussian-based instance-adaptive intensity modeling (GIM) module to model instance-level expression intensity distribution for soft pseudo-labeling. By detecting the pseudo-apex frame around each point label, estimating the duration, and constructing an instance-level Gaussian distribution, GIM assigns soft pseudo-labels to expression frames for more reliable intensity supervision. The GIM module is incorporated into our framework to optimize the class-agnostic expression intensity branch. Second, we design a class-aware apex classification branch that distinguishes macro- and micro-expressions solely based on their pseudo-apex frames. During inference, the two branches work independently: the class-agnostic expression intensity branch generates expression proposals, while the class-aware apex-classification branch is responsible for macro- and micro-expression classification. Furthermore, we introduce an intensity-aware contrastive loss to enhance discriminative feature learning and suppress neutral noise by contrasting neutral frames with expression frames with various intensities. Extensive experiments on the SAMM-LV, CAS(ME)$^2$, and CAS(ME)$^3$ datasets demonstrate the effectiveness of our proposed framework.

Method: Reinterprets face de-identification as image editing in the latent h-space of pretrained unconditional diffusion models. Estimates identity-editing directions through optimization guided by loss functions that preserve attributes while suppressing identity. Introduces both linear and geodesic (tangent-based) editing schemes to navigate the latent manifold effectively.

Result: Experiments on CelebA-HQ and FFHQ datasets show FLUID achieves superior balance between identity suppression and attribute preservation, outperforming existing de-identification approaches in both qualitative and quantitative evaluations.

Conclusion: FLUID presents an effective framework for face de-identification that operates in diffusion model latent space, successfully suppressing identity while preserving important facial attributes like age and gender, offering better realism than previous methods.

Abstract: Current face de-identification methods that replace identifiable cues in the face region with other sacrifices utilities contributing to realism, such as age and gender. To retrieve the damaged realism, we present FLUID (Face de-identification in the Latent space via Utility-preserving Identity Displacement), a single-input face de-identification framework that directly replaces identity features in the latent space of a pretrained diffusion model without affecting the model’s weights. We reinterpret face de-identification as an image editing task in the latent h-space of a pretrained unconditional diffusion model. Our framework estimates identity-editing directions through optimization guided by loss functions that encourage attribute preservation while suppressing identity signals. We further introduce both linear and geodesic (tangent-based) editing schemes to effectively navigate the latent manifold. Experiments on CelebA-HQ and FFHQ show that FLUID achieves a superior balance between identity suppression and attribute preservation, outperforming existing de-identification approaches in both qualitative and quantitative evaluations.

Method: 1) Analyze hallucination through three causal pathways: image-to-input-text, image-to-output-text, and text-to-text. 2) Discover that LVLMs rely on different pathways depending on question-answer alignment format. 3) Propose methods to identify and intervene on critical hallucination heads within each pathway, tailored to discriminative and generative formats.

Result: Experiments across multiple benchmarks demonstrate consistent reduction in hallucinations across diverse alignment types. The approach effectively addresses hallucination by targeting specific causal pathways rather than treating it as a monolithic problem.

Conclusion: Hallucination in LVLMs stems from interplay among multiple causal pathways, not a single source. By understanding and intervening on these specific pathways based on alignment format, effective hallucination reduction can be achieved through targeted architectural interventions.

Abstract: Despite their impressive performance across a wide range of tasks, Large Vision-Language Models (LVLMs) remain prone to hallucination. In this study, we propose a comprehensive intervention framework aligned with the transformer’s causal architecture in LVLMs, integrating the effects of different intervention paths on hallucination. We find that hallucinations in LVLMs do not arise from a single causal path, but rather from the interplay among image-to-input-text, image-to-output-text, and text-to-text pathways. For the first time, we also find that LVLMs rely on different pathways depending on the question-answer alignment format. Building on these insights, we propose simple yet effective methods to identify and intervene on critical hallucination heads within each pathway, tailored to discriminative and generative formats. Experiments across multiple benchmarks demonstrate that our approach consistently reduces hallucinations across diverse alignment types.

Method: Two-stage framework: 1) Debate-Enhanced Pseudo Labeling with adaptive entropy-driven point sampling and multi-agent debate mechanism to improve SAM’s COD capability; 2) FADeNet that progressively fuses multi-level frequency-aware features to balance global semantics with local details while dynamically reweighting supervision strength to alleviate scribble bias.

Result: Achieves state-of-the-art performance on multiple benchmarks, significantly narrowing the gap between weakly and fully supervised COD.

Conclusion: D³ETOR effectively addresses key limitations in WSCOD by enhancing pseudo mask quality through debate mechanisms and mitigating scribble annotation bias through frequency-aware progressive debiasing, demonstrating strong performance in camouflaged object detection with sparse supervision.

Abstract: Weakly-Supervised Camouflaged Object Detection (WSCOD) aims to locate and segment objects that are visually concealed within their surrounding scenes, relying solely on sparse supervision such as scribble annotations. Despite recent progress, existing WSCOD methods still lag far behind fully supervised ones due to two major limitations: (1) the pseudo masks generated by general-purpose segmentation models (e.g., SAM) and filtered via rules are often unreliable, as these models lack the task-specific semantic understanding required for effective pseudo labeling in COD; and (2) the neglect of inherent annotation bias in scribbles, which hinders the model from capturing the global structure of camouflaged objects. To overcome these challenges, we propose ${D}^{3}$ETOR, a two-stage WSCOD framework consisting of Debate-Enhanced Pseudo Labeling and Frequency-Aware Progressive Debiasing. In the first stage, we introduce an adaptive entropy-driven point sampling method and a multi-agent debate mechanism to enhance the capability of SAM for COD, improving the interpretability and precision of pseudo masks. In the second stage, we design FADeNet, which progressively fuses multi-level frequency-aware features to balance global semantic understanding with local detail modeling, while dynamically reweighting supervision strength across regions to alleviate scribble bias. By jointly exploiting the supervision signals from both the pseudo masks and scribble semantics, ${D}^{3}$ETOR significantly narrows the gap between weakly and fully supervised COD, achieving state-of-the-art performance on multiple benchmarks.

Method: Proposes an efficient video defense framework that protects portrait videos by perturbing the 3D information acquisition process. Key innovations include: (1) similarity-guided parameter sharing mechanism for computational efficiency, and (2) multi-scale dual-domain attention module to jointly optimize spatial-frequency perturbations.

Result: The framework demonstrates strong defense capability, achieves 47x acceleration over the fastest baseline while maintaining high fidelity, remains robust against scaling operations and state-of-the-art purification attacks, and validates design choices through ablation studies.

Conclusion: The proposed framework provides an efficient and effective solution for protecting portrait videos against 3D-field talking face generation attacks, addressing both computational efficiency and video quality preservation while maintaining robust defense capabilities.

Abstract: State-of-the-art 3D-field video-referenced Talking Face Generation (TFG) methods synthesize high-fidelity personalized talking-face videos in real time by modeling 3D geometry and appearance from reference portrait video. This capability raises significant privacy concerns regarding malicious misuse of personal portraits. However, no efficient defense framework exists to protect such videos against 3D-field TFG methods. While image-based defenses could apply per-frame 2D perturbations, they incur prohibitive computational costs, severe video quality degradation, failing to disrupt 3D information for video protection. To address this, we propose a novel and efficient video defense framework against 3D-field TFG methods, which protects portrait video by perturbing the 3D information acquisition process while maintain high-fidelity video quality. Specifically, our method introduces: (1) a similarity-guided parameter sharing mechanism for computational efficiency, and (2) a multi-scale dual-domain attention module to jointly optimize spatial-frequency perturbations. Extensive experiments demonstrate that our proposed framework exhibits strong defense capability and achieves a 47x acceleration over the fastest baseline while maintaining high fidelity. Moreover, it remains robust against scaling operations and state-of-the-art purification attacks, and the effectiveness of our design choices is further validated through ablation studies. Our project is available at https://github.com/Richen7418/VDF.

Method: 1) Self-correcting Bidirectional Distillation strategy that retains bidirectional attention within video chunks to preserve spatiotemporal correlations; 2) Multi-step Retrospective Self-Correction Mechanism for error recovery during infinite generation; 3) Full-stack inference acceleration suite with hybrid sequence parallelism, Parallel VAE, and kernel-level optimizations.

Result: Achieves sub-second start-up latency (0.87s) and real-time throughput of 32 FPS, making it the first 14B-scale system to reach these performance levels for high-fidelity interactive digital human synthesis.

Conclusion: SoulX-FlashTalk sets a new standard for real-time, high-fidelity avatar generation by successfully balancing computational complexity with strict latency requirements through innovative bidirectional attention preservation and self-correction mechanisms.

Abstract: Deploying massive diffusion models for real-time, infinite-duration, audio-driven avatar generation presents a significant engineering challenge, primarily due to the conflict between computational load and strict latency constraints. Existing approaches often compromise visual fidelity by enforcing strictly unidirectional attention mechanisms or reducing model capacity. To address this problem, we introduce \textbf{SoulX-FlashTalk}, a 14B-parameter framework optimized for high-fidelity real-time streaming. Diverging from conventional unidirectional paradigms, we use a \textbf{Self-correcting Bidirectional Distillation} strategy that retains bidirectional attention within video chunks. This design preserves critical spatiotemporal correlations, significantly enhancing motion coherence and visual detail. To ensure stability during infinite generation, we incorporate a \textbf{Multi-step Retrospective Self-Correction Mechanism}, enabling the model to autonomously recover from accumulated errors and preventing collapse. Furthermore, we engineered a full-stack inference acceleration suite incorporating hybrid sequence parallelism, Parallel VAE, and kernel-level optimizations. Extensive evaluations confirm that SoulX-FlashTalk is the first 14B-scale system to achieve a \textbf{sub-second start-up latency (0.87s)} while reaching a real-time throughput of \textbf{32 FPS}, setting a new standard for high-fidelity interactive digital human synthesis.

Method: Introduces VLN-MME, a unified and extensible evaluation framework that bridges traditional navigation datasets into a standardized benchmark. Uses highly modular and accessible design to enable structured comparisons and component-level ablations across diverse MLLM architectures, agent designs, and navigation tasks.

Result: Surprisingly, enhancing baseline agents with Chain-of-Thought reasoning and self-reflection leads to performance decrease. This reveals MLLMs exhibit poor context awareness in embodied navigation tasks - they can follow instructions and structure output, but have low 3D spatial reasoning fidelity.

Conclusion: VLN-MME provides groundwork for systematic evaluation of general-purpose MLLMs in embodied navigation and reveals limitations in their sequential decision-making capabilities. These findings offer crucial guidance for MLLM post-training as embodied agents.

Abstract: Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities across a wide range of vision-language tasks. However, their performance as embodied agents, which requires multi-round dialogue spatial reasoning and sequential action prediction, needs further exploration. Our work investigates this potential in the context of Vision-and-Language Navigation (VLN) by introducing a unified and extensible evaluation framework to probe MLLMs as zero-shot agents by bridging traditional navigation datasets into a standardized benchmark, named VLN-MME. We simplify the evaluation with a highly modular and accessible design. This flexibility streamlines experiments, enabling structured comparisons and component-level ablations across diverse MLLM architectures, agent designs, and navigation tasks. Crucially, enabled by our framework, we observe that enhancing our baseline agent with Chain-of-Thought (CoT) reasoning and self-reflection leads to an unexpected performance decrease. This suggests MLLMs exhibit poor context awareness in embodied navigation tasks; although they can follow instructions and structure their output, their 3D spatial reasoning fidelity is low. VLN-MME lays the groundwork for systematic evaluation of general-purpose MLLMs in embodied navigation settings and reveals limitations in their sequential decision-making capabilities. We believe these findings offer crucial guidance for MLLM post-training as embodied agents.

Method: DarkEQA isolates perception bottlenecks by evaluating question answering from egocentric observations under controlled degradations. It uses physical fidelity modeling with visual degradations in linear RAW space, simulating physics-based illumination drop and sensor noise followed by ISP-inspired rendering.

Result: The benchmark reveals systematic limitations of state-of-the-art VLMs and Low-Light Image Enhancement models when operating under challenging low-light conditions.

Conclusion: DarkEQA addresses a critical gap in VLM evaluation for embodied agents, providing a physically realistic benchmark for assessing perceptual robustness in low-light conditions essential for 24/7 operation.

Abstract: Vision Language Models (VLMs) are increasingly adopted as central reasoning modules for embodied agents. Existing benchmarks evaluate their capabilities under ideal, well-lit conditions, yet robust 24/7 operation demands performance under a wide range of visual degradations, including low-light conditions at night or in dark environments–a core necessity that has been largely overlooked. To address this underexplored challenge, we present DarkEQA, an open-source benchmark for evaluating EQA-relevant perceptual primitives under multi-level low-light conditions. DarkEQA isolates the perception bottleneck by evaluating question answering from egocentric observations under controlled degradations, enabling attributable robustness analysis. A key design feature of DarkEQA is its physical fidelity: visual degradations are modeled in linear RAW space, simulating physics-based illumination drop and sensor noise followed by an ISP-inspired rendering pipeline. We demonstrate the utility of DarkEQA by evaluating a wide range of state-of-the-art VLMs and Low-Light Image Enhancement (LLIE) models. Our analysis systematically reveals VLMs’ limitations when operating under these challenging visual conditions. Project website: https://darkeqa-benchmark.github.io/

Method: SlingBAG Pro extends the point cloud iteration concept of the original SlingBAG method to arbitrary array geometries. It uses a hierarchical optimization strategy combining zero-gradient filtering with progressively increased temporal sampling rates during iteration to rapidly remove redundant spatial point clouds and accelerate convergence.

Result: SlingBAG Pro achieves up to 2.2-fold speed improvement in point cloud-based 3D PA reconstruction compared to the original SlingBAG algorithm under irregular array geometries. The method maintains high reconstruction quality while reducing required transducer count.

Conclusion: SlingBAG Pro successfully addresses the challenges of irregular array geometries in 3D photoacoustic imaging, offering faster reconstruction with fewer transducers while maintaining quality. The method is validated through simulation and in vivo mouse experiments, with source code publicly available.

Abstract: High-quality three-dimensional (3D) photoacoustic imaging (PAI) is gaining increasing attention in clinical applications. To address the challenges of limited space and high costs, irregular geometric transducer arrays that conform to specific imaging regions are promising for achieving high-quality 3D PAI with fewer transducers. However, traditional iterative reconstruction algorithms struggle with irregular array configurations, suffering from high computational complexity, substantial memory requirements, and lengthy reconstruction times. In this work, we introduce SlingBAG Pro, an advanced reconstruction algorithm based on the point cloud iteration concept of the Sliding ball adaptive growth (SlingBAG) method, while extending its compatibility to arbitrary array geometries. SlingBAG Pro maintains high reconstruction quality, reduces the number of required transducers, and employs a hierarchical optimization strategy that combines zero-gradient filtering with progressively increased temporal sampling rates during iteration. This strategy rapidly removes redundant spatial point clouds, accelerates convergence, and significantly shortens overall reconstruction time. Compared to the original SlingBAG algorithm, SlingBAG Pro achieves up to a 2.2-fold speed improvement in point cloud-based 3D PA reconstruction under irregular array geometries. The proposed method is validated through both simulation and in vivo mouse experiments, and the source code is publicly available at https://github.com/JaegerCQ/SlingBAG_Pro.

Method: 1) Created FFP-300K dataset (300K high-fidelity 720p video pairs, 81 frames) via two-track pipeline for diverse edits. 2) Proposed guidance-free FFP framework with Adaptive Spatio-Temporal RoPE (AST-RoPE) to disentangle appearance/motion references. 3) Used self-distillation with identity propagation task as regularizer for temporal stability.

Result: Significantly outperforms existing academic and commercial models on EditVerseBench benchmark, achieving ~0.2 PickScore and ~0.3 VLM score improvements against competitors.

Conclusion: The proposed guidance-free FFP framework with AST-RoPE and self-distillation effectively resolves the appearance-motion tension in video editing, demonstrating superior performance through comprehensive dataset construction and architectural innovations.

Abstract: First-Frame Propagation (FFP) offers a promising paradigm for controllable video editing, but existing methods are hampered by a reliance on cumbersome run-time guidance. We identify the root cause of this limitation as the inadequacy of current training datasets, which are often too short, low-resolution, and lack the task diversity required to teach robust temporal priors. To address this foundational data gap, we first introduce FFP-300K, a new large-scale dataset comprising 300K high-fidelity video pairs at 720p resolution and 81 frames in length, constructed via a principled two-track pipeline for diverse local and global edits. Building on this dataset, we propose a novel framework designed for true guidance-free FFP that resolves the critical tension between maintaining first-frame appearance and preserving source video motion. Architecturally, we introduce Adaptive Spatio-Temporal RoPE (AST-RoPE), which dynamically remaps positional encodings to disentangle appearance and motion references. At the objective level, we employ a self-distillation strategy where an identity propagation task acts as a powerful regularizer, ensuring long-term temporal stability and preventing semantic drift. Comprehensive experiments on the EditVerseBench benchmark demonstrate that our method significantly outperforming existing academic and commercial models by receiving about 0.2 PickScore and 0.3 VLM score improvement against these competitors.

Method: Customized hierarchical transformer architecture with: 1) Non-overlapping patch splitting with shifted windows attention to avoid locality issues, 2) Patch and position embeddings for global linking via multi-head attention, 3) Inverse Residual Blocks (IRB) with convolutional skip connections to address vanishing gradients and capture local patterns, 4) Integration of transformer’s global attention with CNN’s local feature extraction.

Result: Achieved 96.51% accuracy and 96.13 F1-score on Kaggle public dataset, outperforming standard CNN models and SwinTransformers. The method effectively minimized Mpox variability while increasing differences between Mpox and similar diseases.

Conclusion: RSwinV2 proves to be a valid computer-assisted tool for Mpox lesion interpretation, successfully combining transformer global attention with convolutional local feature extraction to improve lesion classification capability for differential diagnosis of Mpox from similar skin conditions.

Abstract: In this paper, a deep learning approach for Mpox diagnosis named Customized Residual SwinTransformerV2 (RSwinV2) has been proposed, trying to enhance the capability of lesion classification by employing the RSwinV2 tool-assisted vision approach. In the RSwinV2 method, a hierarchical structure of the transformer has been customized based on the input dimensionality, embedding structure, and output targeted by the method. In this RSwinV2 approach, the input image has been split into non-overlapping patches and processed using shifted windows and attention in these patches. This process has helped the method link all the windows efficiently by avoiding the locality issues of non-overlapping regions in attention, while being computationally efficient. RSwinV2 has further developed based on SwinTransformer and has included patch and position embeddings to take advantage of the transformer global-linking capability by employing multi-head attention in these embeddings. Furthermore, RSwinV2 has developed and incorporated the Inverse Residual Block (IRB) into this method, which utilizes convolutional skip connections with these inclusive designs to address the vanishing gradient issues during processing. RSwinV2 inclusion of IRB has therefore facilitated this method to link global patterns as well as local patterns; hence, its integrity has helped improve lesion classification capability by minimizing variability of Mpox and increasing differences of Mpox, chickenpox, measles, and cowpox. In testing SwinV2, its accuracy of 96.51 and an F1score of 96.13 have been achieved on the Kaggle public dataset, which has outperformed standard CNN models and SwinTransformers; the RSwinV2 vector has thus proved its validity as a computer-assisted tool for Mpox lesion observation interpretation.

Method: PhysSFI-Net combines three components: 1) hierarchical graph module with craniofacial and surgical plan encoders using attention mechanisms to extract skeletal-facial interaction features, 2) LSTM-based sequential predictor for incremental soft tissue deformation, and 3) biomechanics-inspired module for high-resolution facial surface reconstruction.

Result: On 135 patient dataset, PhysSFI-Net achieved point cloud shape error of 1.070±0.088 mm, surface deviation error of 1.296±0.349 mm, and landmark localization error of 2.445±1.326 mm, outperforming state-of-the-art ACMT-Net in prediction accuracy.

Conclusion: PhysSFI-Net enables interpretable, high-resolution prediction of postoperative facial morphology with superior accuracy, showing strong potential for clinical application in orthognathic surgical planning and simulation.

Abstract: Orthognathic surgery repositions jaw bones to restore occlusion and enhance facial aesthetics. Accurate simulation of postoperative facial morphology is essential for preoperative planning. However, traditional biomechanical models are computationally expensive, while geometric deep learning approaches often lack interpretability. In this study, we develop and validate a physics-informed geometric deep learning framework named PhysSFI-Net for precise prediction of soft tissue deformation following orthognathic surgery. PhysSFI-Net consists of three components: a hierarchical graph module with craniofacial and surgical plan encoders combined with attention mechanisms to extract skeletal-facial interaction features; a Long Short-Term Memory (LSTM)-based sequential predictor for incremental soft tissue deformation; and a biomechanics-inspired module for high-resolution facial surface reconstruction. Model performance was assessed using point cloud shape error (Hausdorff distance), surface deviation error, and landmark localization error (Euclidean distances of craniomaxillofacial landmarks) between predicted facial shapes and corresponding ground truths. A total of 135 patients who underwent combined orthodontic and orthognathic treatment were included for model training and validation. Quantitative analysis demonstrated that PhysSFI-Net achieved a point cloud shape error of 1.070 +/- 0.088 mm, a surface deviation error of 1.296 +/- 0.349 mm, and a landmark localization error of 2.445 +/- 1.326 mm. Comparative experiments indicated that PhysSFI-Net outperformed the state-of-the-art method ACMT-Net in prediction accuracy. In conclusion, PhysSFI-Net enables interpretable, high-resolution prediction of postoperative facial morphology with superior accuracy, showing strong potential for clinical application in orthognathic surgical planning and simulation.

Method: Created first large-scale optical-only moraine segmentation dataset (3,340 manually annotated high-resolution Google Earth images from Sichuan and Yunnan, China). Developed MCD-Net - a lightweight baseline integrating MobileNetV2 encoder, Convolutional Block Attention Module (CBAM), and DeepLabV3+ decoder.

Result: MCD-Net achieves 62.3% mean Intersection over Union (mIoU) and 72.8% Dice coefficient while reducing computational cost by more than 60% compared to deeper backbones (ResNet152, Xception). Ridge delineation remains constrained by sub-pixel width and spectral ambiguity.

Conclusion: Optical imagery alone can provide reliable moraine-body segmentation. The publicly available dataset and code establish a reproducible benchmark for moraine-specific segmentation and offer a deployable baseline for high-altitude glacial monitoring.

Abstract: Glacial segmentation is essential for reconstructing past glacier dynamics and evaluating climate-driven landscape change. However, weak optical contrast and the limited availability of high-resolution DEMs hinder automated mapping. This study introduces the first large-scale optical-only moraine segmentation dataset, comprising 3,340 manually annotated high-resolution images from Google Earth covering glaciated regions of Sichuan and Yunnan, China. We develop MCD-Net, a lightweight baseline that integrates a MobileNetV2 encoder, a Convolutional Block Attention Module (CBAM), and a DeepLabV3+ decoder. Benchmarking against deeper backbones (ResNet152, Xception) shows that MCD-Net achieves 62.3% mean Intersection over Union (mIoU) and 72.8% Dice coefficient while reducing computational cost by more than 60%. Although ridge delineation remains constrained by sub-pixel width and spectral ambiguity, the results demonstrate that optical imagery alone can provide reliable moraine-body segmentation. The dataset and code are publicly available at https://github.com/Lyra-alpha/MCD-Net, establishing a reproducible benchmark for moraine-specific segmentation and offering a deployable baseline for high-altitude glacial monitoring.

Method: Uses LLMs to generate textual explanations, applies clustering to identify frequent conditions, converts explanations into transparent rules, and employs two refinement techniques to improve quality and reduce conflicts. Includes automatic predicate generation for state semantics and expert knowledge integration.

Result: The framework successfully generates transparent rules that achieve satisfactory performance on certain tasks, addresses limitations of existing Autonomous Policy Explanation method, and enables systematic quantitative evaluation of textual explanations across three open-source environments and a telecom use case.

Conclusion: The proposed XRL framework provides a comprehensive solution for generating, refining, and evaluating textual explanations of RL policies, offering valuable insights for the XRL field and demonstrating industrial applicability while improving explanation transparency and correctness.

Abstract: Understanding a Reinforcement Learning (RL) policy is crucial for ensuring that autonomous agents behave according to human expectations. This goal can be achieved using Explainable Reinforcement Learning (XRL) techniques. Although textual explanations are easily understood by humans, ensuring their correctness remains a challenge, and evaluations in state-of-the-art remain limited. We present a novel XRL framework for generating textual explanations, converting them into a set of transparent rules, improving their quality, and evaluating them. Expert’s knowledge can be incorporated into this framework, and an automatic predicate generator is also proposed to determine the semantic information of a state. Textual explanations are generated using a Large Language Model (LLM) and a clustering technique to identify frequent conditions. These conditions are then converted into rules to evaluate their properties, fidelity, and performance in the deployed environment. Two refinement techniques are proposed to improve the quality of explanations and reduce conflicting information. Experiments were conducted in three open-source environments to enable reproducibility, and in a telecom use case to evaluate the industrial applicability of the proposed XRL framework. This framework addresses the limitations of an existing method, Autonomous Policy Explanation, and the generated transparent rules can achieve satisfactory performance on certain tasks. This framework also enables a systematic and quantitative evaluation of textual explanations, providing valuable insights for the XRL field.

Method: Three-stage pipeline: 1) Semantic Structured Compression distills interactions into compact memory units, 2) Recursive Memory Consolidation integrates related units into abstract representations, 3) Adaptive Query-Aware Retrieval dynamically adjusts retrieval scope based on query complexity.

Result: Outperforms baselines in accuracy, retrieval efficiency, and inference cost, achieving 26.4% average F1 improvement while reducing inference-time token consumption by up to 30x.

Conclusion: SimpleMem demonstrates superior balance between performance and efficiency for LLM agent memory systems through semantic lossless compression, offering practical benefits for long-term interactions in complex environments.

Abstract: To support reliable long-term interaction in complex environments, LLM agents require memory systems that efficiently manage historical experiences. Existing approaches either retain full interaction histories via passive context extension, leading to substantial redundancy, or rely on iterative reasoning to filter noise, incurring high token costs. To address this challenge, we introduce SimpleMem, an efficient memory framework based on semantic lossless compression. We propose a three-stage pipeline designed to maximize information density and token utilization: (1) \textit{Semantic Structured Compression}, which applies entropy-aware filtering to distill unstructured interactions into compact, multi-view indexed memory units; (2) \textit{Recursive Memory Consolidation}, an asynchronous process that integrates related units into higher-level abstract representations to reduce redundancy; and (3) \textit{Adaptive Query-Aware Retrieval}, which dynamically adjusts retrieval scope based on query complexity to construct precise context efficiently. Experiments on benchmark datasets show that our method consistently outperforms baseline approaches in accuracy, retrieval efficiency, and inference cost, achieving an average F1 improvement of 26.4% while reducing inference-time token consumption by up to 30-fold, demonstrating a superior balance between performance and efficiency. Code is available at https://github.com/aiming-lab/SimpleMem.

Method: Orchestral defines a single universal representation for messages, tools, and LLM usage that works across providers. It uses automatic tool schema generation from Python type hints, a synchronous execution model with streaming support, and a modular architecture that separates provider integration, tool execution, conversation orchestration, and user interfaces.

Result: The framework eliminates manual format translation, reduces framework-induced complexity, enables deterministic behavior and straightforward debugging, and supports advanced agent capabilities including rich tool calling, context compaction, workspace sandboxing, user approval workflows, sub-agents, memory management, and MCP integration.

Conclusion: Orchestral provides a unified solution for building portable, reliable LLM agent systems that preserves simplicity for both scientific computing and production deployment while addressing the core engineering challenges of multi-provider tool calling integration.

Abstract: The rapid proliferation of LLM agent frameworks has forced developers to choose between vendor lock-in through provider-specific SDKs and complex multi-package ecosystems that obscure control flow and hinder reproducibility. Integrating tool calling across multiple LLM providers remains a core engineering challenge due to fragmented APIs, incompatible message formats, and inconsistent streaming and tool-calling behavior, making it difficult to build portable, reliable agent systems. We introduce Orchestral, a lightweight Python framework that provides a unified, type-safe interface for building LLM agents across major providers while preserving the simplicity required for scientific computing and production deployment. Orchestral defines a single universal representation for messages, tools, and LLM usage that operates seamlessly across providers, eliminating manual format translation and reducing framework-induced complexity. Automatic tool schema generation from Python type hints removes the need for handwritten descriptors while maintaining type safety across provider boundaries. A synchronous execution model with streaming support enables deterministic behavior, straightforward debugging, and real-time interaction without introducing server dependencies. The framework’s modular architecture cleanly separates provider integration, tool execution, conversation orchestration, and user-facing interfaces, enabling extensibility without architectural entanglement. Orchestral supports advanced agent capabilities found in larger frameworks, including rich tool calling, context compaction, workspace sandboxing, user approval workflows, sub-agents, memory management, and MCP integration.

Method: The authors conduct extensive experiments on five mobile devices, focusing on two key issues: (1) on-device model selection and resource consumption analysis, and (2) domain adaptation capabilities. They create LiveChatBench, a manually constructed benchmark of 1,000 Korean-English parallel sentence pairs for live-stream chat translation.

Result: Experiments show that while serving diverse users requires careful consideration of constrained deployment settings and model selection, the proposed approach achieves performance comparable to commercial models like GPT-5.1 on the targeted live-stream chat translation task.

Conclusion: The findings provide meaningful insights for the on-device AI community, demonstrating that practical deployment considerations are crucial for real-world service implementation, and that on-device models can achieve competitive performance with proper benchmarking and adaptation strategies.

Abstract: Despite its efficiency, there has been little research on the practical aspects required for real-world deployment of on-device AI models, such as the device’s CPU utilization and thermal conditions. In this paper, through extensive experiments, we investigate two key issues that must be addressed to deploy on-device models in real-world services: (i) the selection of on-device models and the resource consumption of each model, and (ii) the capability and potential of on-device models for domain adaptation. To this end, we focus on a task of translating live-stream chat messages and manually construct LiveChatBench, a benchmark consisting of 1,000 Korean-English parallel sentence pairs. Experiments on five mobile devices demonstrate that, although serving a large and heterogeneous user base requires careful consideration of highly constrained deployment settings and model selection, the proposed approach nevertheless achieves performance comparable to commercial models such as GPT-5.1 on the well-targeted task. We expect that our findings will provide meaningful insights to the on-device AI community.

Method: Frames infeasibility as preference-aware query repair problem. Proposes three LLM-based methods for inferring relative constraint importance from dialogue: (1) local weighting, (2) global one-shot weighting, and (3) pairwise ranking. Also introduces AWARE-US benchmark for evaluation.

Result: Experiments show local weighting achieves best preference alignment, while global weighting performs best on correct constraint relaxation. The AWARE-US benchmark enables evaluation of persona-grounded queries requiring disambiguation and preference-aware infeasibility resolution.

Conclusion: Preference-aware query repair is crucial for conversational agents handling infeasible database queries. LLM-based methods can effectively infer constraint importance from dialogue, with different approaches excelling at different aspects of the problem.

Abstract: Tool-calling conversational agents querying structured databases often face two linked failures: underspecification (missing constraints needed to run a precise query) and infeasibility (the fully specified query returns an empty set because no item satisfies all constraints). Existing work often responds with “no results” or relaxes constraints using ad hoc rules, which can violate user intent by discarding requirements the user cares about most. We frame infeasibility handling as a preference-aware query repair problem: when a query is unsatisfiable, the agent should relax the least important constraints to the user. We propose three LLM-based methods for inferring relative constraint importance from dialogue: (1) local weighting, (2) global one-shot weighting, and (3) pairwise ranking. Experiments show local weighting achieves the best preference alignment, while global weighting performs best on correct constraint relaxation. We also introduce AWARE-US, a benchmark of persona-grounded queries requiring agents to disambiguate requests via conversation and resolve infeasibility in a way consistent with persona-implied preferences.

Method: Simultaneously learns policies and mines Causal Graph Temporal Logic (Causal GTL) specifications. Uses robustness-based reward shaping, collects counterexamples when effects fail, and employs Gaussian Process-driven Bayesian optimization to refine parameterized cause templates, capturing spatial and temporal correlations.

Result: Case studies in gene and power networks demonstrate faster learning and clearer, verifiable behavior compared to standard RL baselines.

Conclusion: GTL-CIRL provides an effective framework for spatial-temporal decision-making that improves sample efficiency, interpretability, and produces verifiable behavior by integrating causal reasoning with reinforcement learning.

Abstract: Decision-making tasks often unfold on graphs with spatial-temporal dynamics. Black-box reinforcement learning often overlooks how local changes spread through network structure, limiting sample efficiency and interpretability. We present GTL-CIRL, a closed-loop framework that simultaneously learns policies and mines Causal Graph Temporal Logic (Causal GTL) specifications. The method shapes rewards with robustness, collects counterexamples when effects fail, and uses Gaussian Process (GP) driven Bayesian optimization to refine parameterized cause templates. The GP models capture spatial and temporal correlations in the system dynamics, enabling efficient exploration of complex parameter spaces. Case studies in gene and power networks show faster learning and clearer, verifiable behavior compared to standard RL baselines.

Method: HAPO introduces three innovations: (1) dynamic attribution mechanism targeting error patterns in training data and prompting history, (2) semantic-unit optimization for editing functional prompt segments rather than generating from scratch, and (3) multimodal-friendly progression supporting both LLM and LLM-MLLM workflows.

Result: Applied to single/multi-image QA (e.g., OCRV2) and complex task analysis (e.g., BBH), HAPO demonstrates enhanced optimization efficiency and outperforms comparable automated prompt optimization methods.

Conclusion: HAPO establishes an extensible paradigm for scalable prompt engineering that addresses prompt drift and interpretability issues while supporting multimodal applications.

Abstract: Optimization is fundamental across numerous disciplines, typically following an iterative process of refining an initial solution to enhance performance. This principle is equally critical in prompt engineering, where designing effective prompts for large language models constitutes a complex optimization challenge. A structured optimization approach requires automated or semi-automated procedures to develop improved prompts, thereby reducing manual effort, improving performance, and yielding an interpretable process. However, current prompt optimization methods often induce prompt drift, where new prompts fix prior failures but impair performance on previously successful tasks. Additionally, generating prompts from scratch can compromise interpretability. To address these limitations, this study proposes the Hierarchical Attribution Prompt Optimization (HAPO) framework, which introduces three innovations: (1) a dynamic attribution mechanism targeting error patterns in training data and prompting history, (2) semantic-unit optimization for editing functional prompt segments, and (3) multimodal-friendly progression supporting both end-to-end LLM and LLM-MLLM workflows. Applied in contexts like single/multi-image QA (e.g., OCRV2) and complex task analysis (e.g., BBH), HAPO demonstrates enhanced optimization efficiency, outperforming comparable automated prompt optimization methods and establishing an extensible paradigm for scalable prompt engineering.

Method: InfiAgent uses a file-centric state abstraction to externalize persistent state. At each step, the agent reconstructs context from a workspace state snapshot plus a fixed window of recent actions, keeping the reasoning context strictly bounded.

Result: Experiments on DeepResearch and an 80-paper literature review task show that InfiAgent with a 20B open-source model is competitive with larger proprietary systems and maintains substantially higher long-horizon coverage than context-centric baselines.

Conclusion: Explicit state externalization provides a practical foundation for stable long-horizon agents, enabling bounded context management without task-specific fine-tuning.

Abstract: LLM agents can reason and use tools, but they often break down on long-horizon tasks due to unbounded context growth and accumulated errors. Common remedies such as context compression or retrieval-augmented prompting introduce trade-offs between information fidelity and reasoning stability. We present InfiAgent, a general-purpose framework that keeps the agent’s reasoning context strictly bounded regardless of task duration by externalizing persistent state into a file-centric state abstraction. At each step, the agent reconstructs context from a workspace state snapshot plus a fixed window of recent actions. Experiments on DeepResearch and an 80-paper literature review task show that, without task-specific fine-tuning, InfiAgent with a 20B open-source model is competitive with larger proprietary systems and maintains substantially higher long-horizon coverage than context-centric baselines. These results support explicit state externalization as a practical foundation for stable long-horizon agents. Github Repo:https://github.com/ChenglinPoly/infiAgent

Method: Combines masked visual modeling with cross-modal contrastive alignment using three per-frame prompt tokens: identity, vocal motion (speech-synchronized), and ambient motion (audio-agnostic). Uses time-aligned vocal-motion and audio tokens as positives and misaligned pairs as negatives to drive both modalities into a shared embedding space.

Result: Achieves state-of-the-art results across four disparate downstream task families: audio-visual stream synchronization, facial emotion and head/face action recognition, visual speech recognition, and visual dubbing with indistinguishable audio- or video-driven control.

Conclusion: SyncLipMAE demonstrates the effectiveness of synchronization-aware, factorized self-supervised pretraining for learning transferable facial dynamics that work across multiple downstream applications requiring distinct capabilities.

Abstract: We introduce SyncLipMAE, a self-supervised pretraining framework for talking-face video that learns synchronization-aware and transferable facial dynamics from unlabeled audio-visual streams. Our approach couples masked visual modeling with cross-modal contrastive alignment and employs three per-frame prompt tokens that explicitly encode the essential factors of a talking-face frame - identity, vocal motion (speech-synchronized facial dynamics), and ambient motion (audio-agnostic movements such as blinks and head pose). The contrastive objective uses time-aligned vocal-motion and audio tokens as positives and misaligned pairs as negatives, driving both modalities into a shared embedding space and yielding token-level audio-visual stream synchronization. After pretraining, the aligned audio tokens together with the visual prompt tokens (identity, vocal motion, ambient motion) form a unified interface for four disparate downstream settings: (i) audio-visual stream synchronization; (ii) facial emotion and head/face action recognition; (iii) visual speech recognition; and (iv) visual dubbing, for which we enable indistinguishable audio- or video-driven control within a single model. Across four task families that require distinct capabilities, SyncLipMAE achieves state-of-the-art results, underscoring the effectiveness of synchronization-aware, factorized self-supervised pretraining.

Method: Introduces MultiSessionCollab benchmark and develops long-term collaborative agents with persistent memory that refines user preferences as interactions accumulate. Uses learning signals from user simulator behavior to train agents for better reflection and memory updates.

Result: Memory-equipped agents show improved long-term collaboration with higher task success rates, more efficient interactions, and reduced user effort. Human study confirms memory improves real-world user experience.

Conclusion: Memory mechanisms are essential for conversational agents to effectively learn and adapt to user preferences over multiple sessions, leading to better long-term collaboration and user experience.

Abstract: As conversational agents accumulate experience collaborating with users, adapting to user preferences is essential for fostering long-term relationships and improving collaboration quality over time. We introduce MultiSessionCollab, a benchmark that evaluates how well agents can learn user preferences and leverage them to improve collaboration quality throughout multiple sessions. To develop agents that succeed in this setting, we present long-term collaborative agents equipped with a memory that persists and refines user preference as interaction experience accumulates. Moreover, we demonstrate that learning signals can be derived from user simulator behavior in MultiSessionCollab to train agents to generate more comprehensive reflections and update their memory more effectively. Extensive experiments show that equipping agents with memory improves long-term collaboration, yielding higher task success rates, more efficient interactions, and reduced user effort. Finally, we conduct a human user study that demonstrates that memory helps improve user experience in real-world settings.

Method: Proposes “Time-Scaling” - architectural design utilizing extended temporal pathways to enable deeper problem space exploration, dynamic strategy adjustment, and enhanced metacognitive control. This involves systematic extension and optimization of agents’ ability to unfold reasoning over time, paralleling human sequential reasoning.

Result: Time-Scaling represents a critical frontier for enhancing deep reasoning and problem-solving without proportional increases in static model parameters. It enables deeper semantic reasoning capabilities similar to human problem-solving under cognitive constraints.

Conclusion: Advancing intelligent agent capabilities requires placing Time-Scaling principles at the forefront, positioning explicit temporal reasoning management as foundational for next-generation AI systems with robust semantic reasoning abilities.

Abstract: Early artificial intelligence paradigms exhibited separated cognitive functions: Neural Networks focused on “perception-representation,” Reinforcement Learning on “decision-making-behavior,” and Symbolic AI on “knowledge-reasoning.” With Transformer-based large models and world models, these paradigms are converging into cognitive agents with closed-loop “perception-decision-action” capabilities. Humans solve complex problems under limited cognitive resources through temporalized sequential reasoning. Language relies on problem space search for deep semantic reasoning. While early large language models (LLMs) could generate fluent text, they lacked robust semantic reasoning capabilities. Prompting techniques like Chain-of-Thought (CoT) and Tree-of-Thought (ToT) extended reasoning paths by making intermediate steps explicit. Recent models like DeepSeek-R1 enhanced performance through explicit reasoning trajectories. However, these methods have limitations in search completeness and efficiency. This highlights the need for “Time-Scaling”–the systematic extension and optimization of an agent’s ability to unfold reasoning over time. Time-Scaling refers to architectural design utilizing extended temporal pathways, enabling deeper problem space exploration, dynamic strategy adjustment, and enhanced metacognitive control, paralleling human sequential reasoning under cognitive constraints. It represents a critical frontier for enhancing deep reasoning and problem-solving without proportional increases in static model parameters. Advancing intelligent agent capabilities requires placing Time-Scaling principles at the forefront, positioning explicit temporal reasoning management as foundational.

Method: The chapter traces architectural progression from statistical models to transformer-based systems, analyzes capabilities enabling agentic behavior, and provides an integrative framework of core components (perception, memory, planning, tool execution).

Result: Three main contributions: (1) synthesis of how LLM capabilities extend to agency through reasoning-action-reflection loops; (2) integrative framework for autonomous behavior; (3) critical assessment of applications and persistent challenges in safety, alignment, reliability, and sustainability.

Conclusion: Responsible advancement requires simultaneous progress in technical robustness, interpretability, and ethical safeguards, with critical research priorities including verifiable planning, scalable multi-agent coordination, persistent memory architectures, and governance frameworks.

Abstract: The evolution of Large Language Models (LLMs) from passive text generators to autonomous, goal-driven systems represents a fundamental shift in artificial intelligence. This chapter examines the emergence of agentic AI systems that integrate planning, memory, tool use, and iterative reasoning to operate autonomously in complex environments. We trace the architectural progression from statistical models to transformer-based systems, identifying capabilities that enable agentic behavior: long-range reasoning, contextual awareness, and adaptive decision-making. The chapter provides three contributions: (1) a synthesis of how LLM capabilities extend toward agency through reasoning-action-reflection loops; (2) an integrative framework describing core components perception, memory, planning, and tool execution that bridge LLMs with autonomous behavior; (3) a critical assessment of applications and persistent challenges in safety, alignment, reliability, and sustainability. Unlike existing surveys, we focus on the architectural transition from language understanding to autonomous action, emphasizing the technical gaps that must be resolved before deployment. We identify critical research priorities, including verifiable planning, scalable multi-agent coordination, persistent memory architectures, and governance frameworks. Responsible advancement requires simultaneous progress in technical robustness, interpretability, and ethical safeguards to realize potential while mitigating risks of misalignment and unintended consequences.

Method: Integrates Large Language Models with vision foundation models in a hierarchical structure to mitigate hallucination, forming the ChangeGPT agent framework.

Result: Achieved 90.71% Match rate with GPT-4-turbo backend, showing superior performance especially in multi-step reasoning and tool selection for change-related queries.

Conclusion: ChangeGPT provides intelligence, adaptability, and multi-type change analysis, offering a powerful solution for decision-making in remote sensing applications.

Abstract: Existing change detection methods often lack the versatility to handle diverse real-world queries and the intelligence for comprehensive analysis. This paper presents a general agent framework, integrating Large Language Models (LLM) with vision foundation models to form ChangeGPT. A hierarchical structure is employed to mitigate hallucination. The agent was evaluated on a curated dataset of 140 questions categorized by real-world scenarios, encompassing various question types (e.g., Size, Class, Number) and complexities. The evaluation assessed the agent’s tool selection ability (Precision/Recall) and overall query accuracy (Match). ChangeGPT, especially with a GPT-4-turbo backend, demonstrated superior performance, achieving a 90.71 % Match rate. Its strength lies particularly in handling change-related queries requiring multi-step reasoning and robust tool selection. Practical effectiveness was further validated through a real-world urban change monitoring case study in Qianhai Bay, Shenzhen. By providing intelligence, adaptability, and multi-type change analysis, ChangeGPT offers a powerful solution for decision-making in remote sensing applications.

Method: HAL derives explicit conversational traits from contrastive dialogue data, combines them into a compact scalar score, and uses this as a transparent reward signal for alignment with standard preference optimization methods.

Result: Models aligned with HAL are more frequently perceived as human-like in conversation during large-scale human evaluations. The approach works with models of varying sizes without affecting overall performance.

Conclusion: HAL demonstrates that soft, qualitative properties of language can be made measurable and aligned in an interpretable way, enabling inspection of alignment behavior and diagnosis of unintended effects.

Abstract: Conversational human-likeness plays a central role in human-AI interaction, yet it has remained difficult to define, measure, and optimize. As a result, improvements in human-like behavior are largely driven by scale or broad supervised training, rather than targeted alignment. We introduce Human Aligning LLMs (HAL), a framework for aligning language models to conversational human-likeness using an interpretable, data-driven reward. HAL derives explicit conversational traits from contrastive dialogue data, combines them into a compact scalar score, and uses this score as a transparent reward signal for alignment with standard preference optimization methods. Using this approach, we align models of varying sizes without affecting their overall performance. In large-scale human evaluations, models aligned with HAL are more frequently perceived as human-like in conversation. Because HAL operates over explicit, interpretable traits, it enables inspection of alignment behavior and diagnosis of unintended effects. More broadly, HAL demonstrates how soft, qualitative properties of language–previously outside the scope for alignment–can be made measurable and aligned in an interpretable and explainable way.

Method: Comprehensive enterprise-specific benchmark evaluating 18 distinct agentic configurations across state-of-the-art LLMs. Examines four critical dimensions: orchestration strategy, agent prompt implementation (ReAct vs function calling), memory architecture, and thinking tool integration.

Result: Reveals significant model-specific architectural preferences challenging one-size-fits-all paradigm. Shows significant weaknesses in overall agentic performance: highest scoring models achieve only 35.3% success on complex tasks and 70.8% on simpler tasks.

Conclusion: Findings should inform future agentic system design by enabling more empirically backed decisions regarding architectural components and model selection, moving away from universal approaches to model-specific configurations.

Abstract: While individual components of agentic architectures have been studied in isolation, there remains limited empirical understanding of how different design dimensions interact within complex multi-agent systems. This study aims to address these gaps by providing a comprehensive enterprise-specific benchmark evaluating 18 distinct agentic configurations across state-of-the-art large language models. We examine four critical agentic system dimensions: orchestration strategy, agent prompt implementation (ReAct versus function calling), memory architecture, and thinking tool integration. Our benchmark reveals significant model-specific architectural preferences that challenge the prevalent one-size-fits-all paradigm in agentic AI systems. It also reveals significant weaknesses in overall agentic performance on enterprise tasks with the highest scoring models achieving a maximum of only 35.3% success on the more complex task and 70.8% on the simpler task. We hope these findings inform the design of future agentic systems by enabling more empirically backed decisions regarding architectural components and model selection.

Method: Causal graph-enhanced retrieval-augmented generation system integrating explicit causal reasoning with dual-level knowledge graphs. Uses evidence-first protocols where every causal claim traces to retrieved literature and automatically generates directed acyclic graphs visualizing intervention-outcome pathways.

Result: Evaluation on 234 dementia exercise abstracts shows 95% accuracy, 100% retrieval success, and zero hallucinations versus 34% accuracy and 10% hallucinations for baseline AI. Automatic causal graphs enable explicit mechanism modeling and enhanced interpretability.

Conclusion: While proof-of-concept evaluation focused on dementia exercise research, the architectural approach demonstrates transferable principles for trustworthy medical AI and causal reasoning’s potential for high-stakes healthcare applications.

Abstract: Systematic reviews are essential for evidence-based medicine, but reviewing 1.5 million+ annual publications manually is infeasible. Current AI approaches suffer from hallucinations in systematic review tasks, with studies reporting rates ranging from 28–40% for earlier models to 2–15% for modern implementations which is unacceptable when errors impact patient care. We present a causal graph-enhanced retrieval-augmented generation system integrating explicit causal reasoning with dual-level knowledge graphs. Our approach enforces evidence-first protocols where every causal claim traces to retrieved literature and automatically generates directed acyclic graphs visualizing intervention-outcome pathways. Evaluation on 234 dementia exercise abstracts shows CausalAgent achieves 95% accuracy, 100% retrieval success, and zero hallucinations versus 34% accuracy and 10% hallucinations for baseline AI. Automatic causal graphs enable explicit mechanism modeling, visual synthesis, and enhanced interpretability. While this proof-of-concept evaluation used ten questions focused on dementia exercise research, the architectural approach demonstrates transferable principles for trustworthy medical AI and causal reasoning’s potential for high-stakes healthcare.

Method: 1) Formalize debate dynamics as an identity-weighted Bayesian update process. 2) Propose response anonymization: remove identity markers from prompts so agents cannot distinguish “self” from “peer”, forcing equal weights on agent identity. 3) Define the Identity Bias Coefficient (IBC), a principled bias metric that measures an agent’s tendency to follow its peer versus itself.

Result: Empirical studies across multiple models and benchmarks confirm that identity bias is widespread in MAD systems, with sycophancy far more common than self-bias. The proposed anonymization method effectively reduces bias and improves trustworthiness of debate outcomes.

Conclusion: The work highlights the critical need to address identity bias in multi-agent debate systems and provides a principled framework for both mitigating and quantifying such bias. The findings emphasize that MAD systems should reason based on content rather than identity to ensure reliable outcomes.

Abstract: Multi-agent debate (MAD) aims to improve large language model (LLM) reasoning by letting multiple agents exchange answers and then aggregate their opinions. Yet recent studies reveal that agents are not neutral: they are prone to identity-driven sycophancy and self-bias, uncritically adopting a peer’s view or stubbornly adhering to their own prior output, undermining the reliability of debate. In this work, we present the first principled framework that joins sycophancy and self-bias to mitigate and quantify identity bias in MAD. First, we formalize the debate dynamics as an identity-weighted Bayesian update process. Second, we propose response anonymization: by removing identity markers from prompts, agents cannot distinguish “self” from “peer”, which forces equal weights on agent identity, thereby reducing bias and improving trustworthiness. Third, we define the Identity Bias Coefficient (IBC), a principled bias metric that measures an agent’s tendency to follow its peer versus itself. Empirical studies across multiple models and benchmarks confirm that identity bias is widespread, with sycophancy far more common than self-bias. Our findings highlight the need to ensure that MAD systems reason based on content rather than identity. Code is released in https://github.com/deeplearning-wisc/MAD-identity-bias.

Method: Developed QLSTMA model incorporating variational quantum circuits (VQCs) into recurrent cells. Designed two quantization structures: QLSTMA with Shared Gates (QLSTMA-SG) and with Independent Gates (QLSTMA-IG). Investigated effects of quantum structure configurations and qubit count on performance.

Result: 8-qubit QLSTMA-IG significantly outperformed traditional LSTMA, reducing MAE by 19% and RMSE by 20%. Performed particularly well in regions with complex well-logging data. Increasing qubits yields further accuracy gains despite classical simulation limitations.

Conclusion: Validates potential of quantum-classical hybrid neural networks for reservoir prediction. Establishes framework for eventual deployment on real quantum hardware and extension to broader petroleum engineering/geoscience applications.

Abstract: Spatial prediction of reservoir parameters, especially permeability, is crucial for oil and gas exploration and development. However, the wide range and high variability of permeability prevent existing methods from providing reliable predictions. For the first time in subsurface spatial prediction, this study presents a quantum-enhanced long short-term memory with attention (QLSTMA) model that incorporates variational quantum circuits (VQCs) into the recurrent cell. Using quantum entanglement and superposition principles, the QLSTMA significantly improves the ability to predict complex geological parameters such as permeability. Two quantization structures, QLSTMA with Shared Gates (QLSTMA-SG) and with Independent Gates (QLSTMA-IG), are designed to investigate and evaluate the effects of quantum structure configurations and the number of qubits on model performance. Experimental results demonstrate that the 8-qubit QLSTMA-IG model significantly outperforms the traditional long short-term memory with attention (LSTMA), reducing Mean Absolute Error (MAE) by 19% and Root Mean Squared Error (RMSE) by 20%, with particularly strong performance in regions featuring complex well-logging data. These findings validate the potential of quantum-classical hybrid neural networks for reservoir prediction, indicating that increasing the number of qubits yields further accuracy gains despite the reliance on classical simulations. This study establishes a foundational framework for the eventual deployment of such models on real quantum hardware and their extension to broader applications in petroleum engineering and geoscience.

Method: Integrates background symbolic knowledge through partial policies represented as logical rules, using online reasoning to guide training via: (i) biasing action distribution during exploration, and (ii) rescaling Q-values during exploitation.

Result: Empirical validation on challenging gridworld variants (fully and partially observable) shows improved performance over state-of-the-art reward machine baseline, with enhanced interpretability, trustworthiness, and accelerated convergence.

Conclusion: Neuro-symbolic integration of symbolic knowledge with DRL improves sample efficiency, generalization, interpretability, and convergence, particularly in sparse-reward environments and tasks with long planning horizons.

Abstract: Reinforcement Learning (RL) is a well-established framework for sequential decision-making in complex environments. However, state-of-the-art Deep RL (DRL) algorithms typically require large training datasets and often struggle to generalize beyond small-scale training scenarios, even within standard benchmarks. We propose a neuro-symbolic DRL approach that integrates background symbolic knowledge to improve sample efficiency and generalization to more challenging, unseen tasks. Partial policies defined for simple domain instances, where high performance is easily attained, are transferred as useful priors to accelerate learning in more complex settings and avoid tuning DRL parameters from scratch. To do so, partial policies are represented as logical rules, and online reasoning is performed to guide the training process through two mechanisms: (i) biasing the action distribution during exploration, and (ii) rescaling Q-values during exploitation. This neuro-symbolic integration enhances interpretability and trustworthiness while accelerating convergence, particularly in sparse-reward environments and tasks with long planning horizons. We empirically validate our methodology on challenging variants of gridworld environments, both in the fully observable and partially observable setting. We show improved performance over a state-of-the-art reward machine baseline.

Method: Created M3MAD-Bench with standardized protocols across five core task domains (Knowledge, Mathematics, Medicine, Natural Sciences, Complex Reasoning), covering both text and vision-language datasets. Evaluated nine base models with different architectures, scales, and modality capabilities, incorporating efficiency metrics like token consumption and inference time.

Result: Provides systematic insights into MAD effectiveness, robustness, and efficiency across text-only and multimodal scenarios. The benchmark enables controlled cross-modality comparison and offers performance-cost trade-off analysis.

Conclusion: M3MAD-Bench establishes a reliable foundation for standardized MAD evaluation, addressing current limitations and supporting future research with comprehensive multi-domain, multi-modal, multi-dimensional assessment.

Abstract: As an agent-level reasoning and coordination paradigm, Multi-Agent Debate (MAD) orchestrates multiple agents through structured debate to improve answer quality and support complex reasoning. However, existing research on MAD suffers from two fundamental limitations: evaluations are conducted under fragmented and inconsistent settings, hindering fair comparison, and are largely restricted to single-modality scenarios that rely on textual inputs only. To address these gaps, we introduce M3MAD-Bench, a unified and extensible benchmark for evaluating MAD methods across Multi-domain tasks, Multi-modal inputs, and Multi-dimensional metrics. M3MAD-Bench establishes standardized protocols over five core task domains: Knowledge, Mathematics, Medicine, Natural Sciences, and Complex Reasoning, and systematically covers both pure text and vision-language datasets, enabling controlled cross-modality comparison. We evaluate MAD methods on nine base models spanning different architectures, scales, and modality capabilities. Beyond accuracy, M3MAD-Bench incorporates efficiency-oriented metrics such as token consumption and inference time, providing a holistic view of performance–cost trade-offs. Extensive experiments yield systematic insights into the effectiveness, robustness, and efficiency of MAD across text-only and multimodal scenarios. We believe M3MAD-Bench offers a reliable foundation for future research on standardized MAD evaluation. The code is available at http://github.com/liaolea/M3MAD-Bench.

Method: Three-stage framework: 1) Develop high-fidelity user simulator to synthesize large-scale conversational data through multi-turn online dialogue; 2) Introduce multi-dimensional evaluation framework based on Chain-of-Intention (CoI) with global-level and instance-level metrics to assess simulator and select high-quality data; 3) Train recruitment proactive dialogue agent on selected dataset.

Result: Experiments in real-world recruitment scenario demonstrate that SimRPD outperforms existing simulator-based data selection strategies, showing practical value for industrial deployment and potential applicability to other business-oriented dialogue scenarios.

Conclusion: SimRPD effectively addresses data scarcity in recruitment proactive dialogue systems through synthetic data generation and quality selection, offering a practical solution for industrial applications and potential extension to other business domains.

Abstract: Task-oriented proactive dialogue agents play a pivotal role in recruitment, particularly for steering conversations towards specific business outcomes, such as acquiring social-media contacts for private-channel conversion. Although supervised fine-tuning and reinforcement learning have proven effective for training such agents, their performance is heavily constrained by the scarcity of high-quality, goal-oriented domain-specific training data. To address this challenge, we propose SimRPD, a three-stage framework for training recruitment proactive dialogue agents. First, we develop a high-fidelity user simulator to synthesize large-scale conversational data through multi-turn online dialogue. Then we introduce a multi-dimensional evaluation framework based on Chain-of-Intention (CoI) to comprehensively assess the simulator and effectively select high-quality data, incorporating both global-level and instance-level metrics. Finally, we train the recruitment proactive dialogue agent on the selected dataset. Experiments in a real-world recruitment scenario demonstrate that SimRPD outperforms existing simulator-based data selection strategies, highlighting its practical value for industrial deployment and its potential applicability to other business-oriented dialogue scenarios.

Method: ReTreVal constructs structured reasoning trees with adaptive depth, performs iterative self-critique and refinement at each node, uses dual validation for quality assessment, implements critique-based pruning to retain top-k nodes, and maintains a reflexion memory buffer for cross-problem learning.

Result: ReTreVal consistently outperforms ReAct, Reflexion, and Self-Refine across 500 mathematical problems and creative writing tasks using Qwen 2.5 7B, showing particular strength in exploratory reasoning, rigorous verification, and knowledge transfer.

Conclusion: ReTreVal’s combination of structured exploration, critique-driven refinement, and cross-problem memory makes it an effective framework for multi-step reasoning tasks requiring exploratory reasoning, verification, and knowledge transfer.

Abstract: Multi-step reasoning remains a key challenge for Large Language Models (LLMs), particularly in complex domains such as mathematics and creative writing. While recent approaches including ReAct, Reflexion, and Self-Refine improve reasoning through iterative refinement and reflection, they often lack structured exploration of alternative solution paths and persistent learning across problems. We propose ReTreVal (Reasoning Tree with Validation), a hybrid framework that integrates Tree-of-Thoughts exploration, self-refinement, LLM-based critique scoring, and reflexion memory to enable bounded and validated multi-step reasoning. ReTreVal constructs a structured reasoning tree with adaptive depth based on problem complexity, where each node undergoes iterative self-critique and refinement guided by explicit LLM-generated feedback. A dual validation mechanism evaluates reasoning quality, coherence, and correctness at each node while persistently storing insights from successful reasoning paths and failure patterns in a reflexion memory buffer, enabling cross-problem learning. Critique-based pruning retains only the top-k highest-scoring nodes at each level, controlling computational cost while preserving high-quality solution paths. We evaluate ReTreVal against ReAct, Reflexion, and Self-Refine across 500 mathematical problems and creative writing tasks using Qwen 2.5 7B as the underlying LLM, and demonstrate that ReTreVal consistently outperforms existing methods through its combination of structured exploration, critique-driven refinement, and cross-problem memory, making it particularly effective for tasks requiring exploratory reasoning, rigorous verification, and knowledge transfer.

Method: Systematic analysis reveals Logical Phase Transitions phenomenon; then propose Neuro-Symbolic Curriculum Tuning that adaptively aligns natural language with logical symbols and reshapes training around phase-transition boundaries.

Result: Approach mitigates reasoning collapse at high complexity, achieving average accuracy gains of +1.26% in naive prompting and +3.95% in Chain-of-Thought, while improving generalization to unseen logical compositions across five benchmarks.

Conclusion: Logical reasoning in LLMs exhibits phase-transition behavior, and targeted curriculum tuning around critical complexity thresholds effectively strengthens reasoning capabilities at increasing logical depths.

Abstract: Symbolic logical reasoning is a critical yet underexplored capability of large language models (LLMs), providing reliable and verifiable decision-making in high-stakes domains such as mathematical reasoning and legal judgment. In this study, we present a systematic analysis of logical reasoning under controlled increases in logical complexity, and reveal a previously unrecognized phenomenon, which we term Logical Phase Transitions: rather than degrading smoothly, logical reasoning performance remains stable within a regime but collapses abruptly beyond a critical logical depth, mirroring physical phase transitions such as water freezing beyond a critical temperature threshold. Building on this insight, we propose Neuro-Symbolic Curriculum Tuning, a principled framework that adaptively aligns natural language with logical symbols to establish a shared representation, and reshapes training dynamics around phase-transition boundaries to progressively strengthen reasoning at increasing logical depths. Experiments on five benchmarks show that our approach effectively mitigates logical reasoning collapse at high complexity, yielding average accuracy gains of +1.26 in naive prompting and +3.95 in CoT, while improving generalization to unseen logical compositions. Code and data are available at https://github.com/AI4SS/Logical-Phase-Transitions.

Method: BoT processes related queries jointly to enable cross-instance learning through comparative analysis across batches. BoT-R adds a multi-agent reflection architecture with a Reflector that performs joint evaluation to unlock mutual information gain.

Result: Experiments across three model families and six benchmarks show BoT-R consistently improves accuracy and confidence calibration while reducing inference costs by up to 61%.

Conclusion: Batch-aware reasoning benefits LLM systems by identifying high-quality reasoning templates, detecting errors through consistency checks, and amortizing computational costs, with theoretical and experimental analysis revealing when and why this approach works.

Abstract: Current Large Language Model reasoning systems process queries independently, discarding valuable cross-instance signals such as shared reasoning patterns and consistency constraints. We introduce Batch-of-Thought (BoT), a training-free method that processes related queries jointly to enable cross-instance learning. By performing comparative analysis across batches, BoT identifies high-quality reasoning templates, detects errors through consistency checks, and amortizes computational costs. We instantiate BoT within a multi-agent reflection architecture (BoT-R), where a Reflector performs joint evaluation to unlock mutual information gain unavailable in isolated processing. Experiments across three model families and six benchmarks demonstrate that BoT-R consistently improves accuracy and confidence calibration while reducing inference costs by up to 61%. Our theoretical and experimental analysis reveals when and why batch-aware reasoning benefits LLM systems.

Method: 1) Induce label-conditioned rationales composed of reasoning paths from observable evidence to potential outcomes. 2) Design hybrid retrieval balancing temporal patterns and semantic contexts to retrieve correlated rationale priors for in-context inference on new samples.

Result: Extensive experiments demonstrate the effectiveness and efficiency of RationaleTS on three-domain time series reasoning tasks. Code will be released for reproduction.

Conclusion: RationaleTS addresses the limitations of existing models by incorporating rationale priors as guiding reasoning units, enabling more principled time series reasoning through rationale-grounded in-context learning.

Abstract: The underperformance of existing multimodal large language models for time series reasoning lies in the absence of rationale priors that connect temporal observations to their downstream outcomes, which leads models to rely on superficial pattern matching rather than principled reasoning. We therefore propose the rationale-grounded in-context learning for time series reasoning, where rationales work as guiding reasoning units rather than post-hoc explanations, and develop the RationaleTS method. Specifically, we firstly induce label-conditioned rationales, composed of reasoning paths from observable evidence to the potential outcomes. Then, we design the hybrid retrieval by balancing temporal patterns and semantic contexts to retrieve correlated rationale priors for the final in-context inference on new samples. We conduct extensive experiments to demonstrate the effectiveness and efficiency of our proposed RationaleTS on three-domain time series reasoning tasks. We will release our code for reproduction.

Method: Proposed fuzzy graph neural network (FGENConv) that integrates mutual information to identify critical network regions and fuzzy logic to provide rule-based explanations for node classification tasks, building on the superior-performing GENConv architecture.

Result: FGENConv achieved Graph F1 scores of 0.889 for detection and 0.814 for localization, slightly below crisp GENConv (0.938 and 0.858) but provides spatially localized, fuzzy rule-based explanations for interpretability.

Conclusion: The framework strikes a balance between precision and explainability, enabling hydraulic engineers to validate leak predictions, conserve human resources, and optimize maintenance strategies while maintaining competitive performance.

Abstract: Timely leak detection in water distribution networks is critical for conserving resources and maintaining operational efficiency. Although Graph Neural Networks (GNNs) excel at capturing spatial-temporal dependencies in sensor data, their black-box nature and the limited work on graph-based explainable models for water networks hinder practical adoption. We propose an explainable GNN framework that integrates mutual information to identify critical network regions and fuzzy logic to provide clear, rule-based explanations for node classification tasks. After benchmarking several GNN architectures, we selected the generalized graph convolution network (GENConv) for its superior performance and developed a fuzzy-enhanced variant that offers intuitive explanations for classified leak locations. Our fuzzy graph neural network (FGENConv) achieved Graph F1 scores of 0.889 for detection and 0.814 for localization, slightly below the crisp GENConv 0.938 and 0.858, respectively. Yet it compensates by providing spatially localized, fuzzy rule-based explanations. By striking the right balance between precision and explainability, the proposed fuzzy network could enable hydraulic engineers to validate predicted leak locations, conserve human resources, and optimize maintenance strategies. The code is available at github.com/pasqualedem/GNNLeakDetection.

Method: Developed a probabilistic Digital Twin of en route UK airspace for AI ATC agent training/testing. Created an assurance framework using Trustworthy and Ethical Assurance (TEA) methodology, which builds nested structured arguments with evidence, assumptions, and justifications to demonstrate Digital Twin accuracy and functionality.

Result: Presented an actionable assurance framework that defines goals and evidence requirements for demonstrating Digital Twin accuracy and functionality. The framework helps researchers assess/document Digital Twin strengths/limitations, supports stakeholder/regulator engagement, and contributes to emerging guidance through a concrete working example.

Conclusion: The assurance framework provides a structured approach for Digital Twin validation and regulatory compliance, serving as a foundation for future applications and contributing to the development of regulatory guidance for AI/ML in aviation through practical implementation.

Abstract: Digital Twins combine simulation, operational data and Artificial Intelligence (AI), and have the potential to bring significant benefits across the aviation industry. Project Bluebird, an industry-academic collaboration, has developed a probabilistic Digital Twin of en route UK airspace as an environment for training and testing AI Air Traffic Control (ATC) agents. There is a developing regulatory landscape for this kind of novel technology. Regulatory requirements are expected to be application specific, and may need to be tailored to each specific use case. We draw on emerging guidance for both Digital Twin development and the use of Artificial Intelligence/Machine Learning (AI/ML) in Air Traffic Management (ATM) to present an assurance framework. This framework defines actionable goals and the evidence required to demonstrate that a Digital Twin accurately represents its physical counterpart and also provides sufficient functionality across target use cases. It provides a structured approach for researchers to assess, understand and document the strengths and limitations of the Digital Twin, whilst also identifying areas where fidelity could be improved. Furthermore, it serves as a foundation for engagement with stakeholders and regulators, supporting discussions around the regulatory needs for future applications, and contributing to the emerging guidance through a concrete, working example of a Digital Twin. The framework leverages a methodology known as Trustworthy and Ethical Assurance (TEA) to develop an assurance case. An assurance case is a nested set of structured arguments that provides justified evidence for how a top-level goal has been realised. In this paper we provide an overview of each structured argument and a number of deep dives which elaborate in more detail upon particular arguments, including the required evidence, assumptions and justifications.

Method: A system for automatic prompt engineering that is simpler in design and application than existing approaches, requiring no tuning and no explicit task clues. The approach is evaluated on cryptic column name expansion in database tables across English and German datasets.

Result: The system is as effective as existing approaches despite being simpler. It represents the first application of automatic prompt engineering to the CNE task and the first application to a language other than English (German).

Conclusion: The paper demonstrates that effective automatic prompt engineering can be achieved with simpler, tuning-free approaches, and successfully applies this to the important but understudied CNE task across multiple languages.

Abstract: This paper presents a system for automatic prompt engineering that is much simpler in both design and application and yet as effective as the existing approaches. It requires no tuning and no explicit clues about the task. We evaluated our approach on cryptic column name expansion (CNE) in database tables, a task which is critical for tabular data search, access, and understanding and yet there has been very little existing work. We evaluated on datasets in two languages, English and German. This is the first work to report on the application of automatic prompt engineering for the CNE task. To the best of our knowledge, this is also the first work on the application of automatic prompt engineering for a language other than English.

Method: MAGMA proposes a multi-graph agentic memory architecture where each memory item is represented across orthogonal semantic, temporal, causal, and entity graphs. Retrieval is formulated as policy-guided traversal over these relational views, enabling query-adaptive selection and structured context construction.

Result: Experiments on LoCoMo and LongMemEval benchmarks show that MAGMA consistently outperforms state-of-the-art agentic memory systems in long-horizon reasoning tasks.

Conclusion: By decoupling memory representation from retrieval logic, MAGMA provides transparent reasoning paths and fine-grained control over retrieval, addressing limitations of existing monolithic memory approaches.

Abstract: Memory-Augmented Generation (MAG) extends Large Language Models with external memory to support long-context reasoning, but existing approaches largely rely on semantic similarity over monolithic memory stores, entangling temporal, causal, and entity information. This design limits interpretability and alignment between query intent and retrieved evidence, leading to suboptimal reasoning accuracy. In this paper, we propose MAGMA, a multi-graph agentic memory architecture that represents each memory item across orthogonal semantic, temporal, causal, and entity graphs. MAGMA formulates retrieval as policy-guided traversal over these relational views, enabling query-adaptive selection and structured context construction. By decoupling memory representation from retrieval logic, MAGMA provides transparent reasoning paths and fine-grained control over retrieval. Experiments on LoCoMo and LongMemEval demonstrate that MAGMA consistently outperforms state-of-the-art agentic memory systems in long-horizon reasoning tasks.

Method: The authors use topological data analysis to compare CLIP and CLOOB, examining modality gap drivers, clustering structures across dimensions, and the role of dimension collapse in shaping embedding spaces.

Result: Empirical experiments show how the topological analyses have implications for downstream performance across various contextual scenarios, revealing strengths and weaknesses of each model.

Conclusion: The investigation provides insights into the comparative efficacy of CLIP and CLOOB, offering a foundation for further refinement and advancement in multimodal model research.

Abstract: Recent strides in multimodal model development have ignited a paradigm shift in the realm of text-to-image generation. Among these advancements, CLIP stands out as a remarkable achievement which is a sophisticated autoencoder adept at encoding both textual and visual information within a unified latent space. This paper delves into a comparative analysis between CLIP and its recent counterpart, CLOOB. To unravel the intricate distinctions within the embedding spaces crafted by these models, we employ topological data analysis. Our approach encompasses a comprehensive examination of the modality gap drivers, the clustering structures existing across both high and low dimensions, and the pivotal role that dimension collapse plays in shaping their respective embedding spaces. Empirical experiments substantiate the implications of our analyses on downstream performance across various contextual scenarios. Through this investigation, we aim to shed light on the nuanced intricacies that underlie the comparative efficacy of CLIP and CLOOB, offering insights into their respective strengths and weaknesses, and providing a foundation for further refinement and advancement in multimodal model research.

Method: UU-Mamba model extends U-Mamba architecture with two key innovations: 1) Sharpness-Aware Minimization (SAM) to find flatter minima in loss landscape for better generalization, and 2) uncertainty-aware loss function combining region-based, distribution-based, and pixel-based components to capture both local and global features.

Result: Superior performance compared to leading models (TransUNet, Swin-Unet, nnUNet, nnFormer) on complex ImageCAS (coronary artery) and Aorta (aortic branches/zones) datasets, demonstrating adaptability and resilience beyond simpler ACDC dataset.

Conclusion: UU-Mamba effectively addresses generalization and robustness challenges in cardiovascular segmentation through SAM optimization and uncertainty-aware loss, showing strong performance on complex segmentation tasks.

Abstract: Deep learning models have achieved significant success in segmenting cardiovascular structures, but there is a growing need to improve their generalization and robustness. Current methods often face challenges such as overfitting and limited accuracy, largely due to their reliance on large annotated datasets and limited optimization techniques. This paper introduces the UU-Mamba model, an extension of the U-Mamba architecture, designed to address these challenges in both cardiac and vascular segmentation. By incorporating Sharpness-Aware Minimization (SAM), the model enhances generalization by seeking flatter minima in the loss landscape. Additionally, we propose an uncertainty-aware loss function that integrates region-based, distribution-based, and pixel-based components, improving segmentation accuracy by capturing both local and global features. We expand our evaluations on the ImageCAS (coronary artery) and Aorta (aortic branches and zones) datasets, which present more complex segmentation challenges than the ACDC dataset (left and right ventricles) used in prior work, showcasing the model’s adaptability and resilience. Our results confirm UU-Mamba’s superior performance compared to leading models such as TransUNet, Swin-Unet, nnUNet, and nnFormer. We also provide a more in-depth assessment of the model’s robustness and segmentation accuracy through extensive experiments.

Method: Proposes SaVe-TAG: uses LLMs for text-level interpolation to generate on-manifold, boundary-enriching synthetic samples for minority classes. Introduces confidence-based edge assignment mechanism using graph topology as a filter for structural consistency.

Result: Extensive experiments on benchmark datasets show consistent outperformance over both numeric interpolation and prior long-tailed node classification baselines.

Conclusion: The approach highlights the importance of integrating semantic and structural signals for balanced and effective learning on text-attributed graphs, with theoretical justification provided.

Abstract: Real-world graph data often follows long-tailed distributions, making it difficult for Graph Neural Networks (GNNs) to generalize well across both head and tail classes. Recent advances in Vicinal Risk Minimization (VRM) have shown promise in mitigating class imbalance with numeric interpolation; however, existing approaches largely rely on embedding-space arithmetic, which fails to capture the rich semantics inherent in text-attributed graphs. In this work, we propose our method, SaVe-TAG (Semantic-aware Vicinal Risk Minimization for Long-Tailed Text-Attributed Graphs), a novel VRM framework that leverages Large Language Models (LLMs) to perform text-level interpolation, generating on-manifold, boundary-enriching synthetic samples for minority classes. To mitigate the risk of noisy generation, we introduce a confidence-based edge assignment mechanism that uses graph topology as a natural filter to ensure structural consistency. We provide theoretical justification for our method and conduct extensive experiments on benchmark datasets, showing that our approach consistently outperforms both numeric interpolation and prior long-tailed node classification baselines. Our results highlight the importance of integrating semantic and structural signals for balanced and effective learning on text-attributed graphs. The source code is publicly available at: https://github.com/LWang-Laura/SaVe-TAG.

Method: Use LLMs to generate problem-specific heuristic functions from planning tasks specified via successor generators, goal tests, and initial states in general-purpose programming language. Compile and integrate these heuristics into standard heuristic search algorithms like greedy best-first search.

Result: Achieves competitive and often state-of-the-art performance across established planning benchmarks. Enables solving problems difficult to express in traditional formalisms, including those with complex numeric constraints or custom transition dynamics.

Conclusion: LLM-generated heuristics represent a promising paradigm shift in domain-independent planning, offering automatic heuristic synthesis that matches or exceeds traditional approaches while handling more complex problem types.

Abstract: Heuristics are a central component of deterministic planning, particularly in domain-independent settings where general applicability is prioritized over task-specific tuning. This work revisits that paradigm in light of recent advances in large language models (LLMs), which enable the automatic synthesis of heuristics directly from problem definitions – bypassing the need for handcrafted domain knowledge. We present a method that employs LLMs to generate problem-specific heuristic functions from planning tasks specified through successor generators, goal tests, and initial states written in a general-purpose programming language. These heuristics are compiled and integrated into standard heuristic search algorithms, such as greedy best-first search. Our approach achieves competitive, and in many cases state-of-the-art, performance across a broad range of established planning benchmarks. Moreover, it enables the solution of problems that are difficult to express in traditional formalisms, including those with complex numeric constraints or custom transition dynamics. We provide an extensive empirical evaluation that characterizes the strengths and limitations of the approach across diverse planning settings, demonstrating its effectiveness.

Method: PatentMind uses a Multi-Aspect Reasoning Graph (MARG) to decompose patents into three dimensions: technical features, application domains, and claim scopes. It calculates dimension-specific similarity scores and dynamically weights them through context-aware reasoning to emulate expert judgment.

Result: PatentMind achieves strong correlation (r=0.938) with expert annotations on the newly constructed PatentSimBench benchmark (500 patent pairs), significantly outperforming embedding-based models, patent-specific models, and advanced prompt engineering methods.

Conclusion: PatentMind provides a structured, semantically grounded foundation for real-world patent decision-making, with broader impact on patent analytics and evaluation, particularly for tasks like infringement risk assessment.

Abstract: Patent similarity evaluation plays a critical role in intellectual property analysis. However, existing methods often overlook the intricate structure of patent documents, which integrate technical specifications, legal boundaries, and application contexts. We introduce PatentMind, a novel framework for patent similarity assessment based on a Multi-Aspect Reasoning Graph (MARG). PatentMind decomposes patents into their three dimensions of technical features, application domains, and claim scopes, then dimension-specific similarity scores are calculated over the MARG. These scores are dynamically weighted through a context-aware reasoning process, which integrates contextual signals to emulate expert-level judgment. To support evaluation, we construct a human-annotated benchmark PatentSimBench, comprising 500 patent pairs. Experimental results demonstrate that the PatentMind-generated scores show a strong correlation ($r=0.938$) with expert annotations, significantly outperforming embedding-based models, patent-specific models, and advanced prompt engineering methods. Beyond computational linguistics, our framework provides a structured and semantically grounded foundation for real-world decision-making, particularly for tasks such as infringement risk assessment, underscoring its broader impact on both patent analytics and evaluation.

Method: The authors propose a hybrid framework that integrates both thinking modes within a single LRM. They identify patterns of essential vs redundant reasoning trajectories, design an auxiliary LLM-based judge to determine when slow thinking is necessary, construct a hybrid fine-tuning dataset by pruning redundant reasoning, and fine-tune LRMs to have autonomous mode-selection capabilities.

Result: Extensive experiments on mathematical and question-answering benchmarks show that OThink-R1 significantly reduces reasoning token usage while maintaining competitive accuracy compared to standard LRMs.

Conclusion: OThink-R1 successfully addresses the efficiency-accuracy trade-off in reasoning models by enabling automatic switching between fast and slow thinking modes based on problem characteristics, making reasoning more efficient without sacrificing performance.

Abstract: Human cognition operates through two complementary modes: fast intuitive thinking and slow deliberate thinking. Vanilla large language models (LLMs) predominantly follow the fast-thinking paradigm, producing immediate responses; while recent large reasoning models (LRMs) adopt slow-thinking strategies, generating detailed reasoning chains before arriving at answers. While LRMs often achieve higher accuracy, this comes at the cost of substantially increased token usage. To address this efficiency-accuracy trade-off, we propose OThink-R1, a hybrid reasoning framework that integrates both modes within a single LRM and enables automatic mode switching based on problem characteristics. We first identify three major patterns of essential and redundant reasoning trajectories in LRMs, which guide the design of an auxiliary LLM-based judge that adaptively determines when slow thinking is necessary. Leveraging the judge’s decisions, we construct a hybrid fine-tuning dataset by pruning redundant reasoning to produce fast-thinking samples and retaining complete reasoning for slow-thinking samples. This dataset is then used to fine-tune LRMs, equipping them with inherent autonomous mode-selection capabilities. Extensive experiments on mathematical and question-answering benchmarks show that OThink-R1 reduces reasoning token usage significantly while maintaining competitive accuracy. The code is available at https://github.com/AgenticIR-Lab/OThink-R1.

Method: SLR automatically synthesizes three components from user task specifications: (1) instruction prompts for inductive reasoning tasks, (2) validation programs that provide verifiable rewards by executing on model outputs, and (3) latent ground-truth rules. This creates SLR-Bench with 19k prompts across 20 curriculum levels of increasing relational, arithmetic, and recursive complexity.

Result: Evaluation shows LLMs produce syntactically valid rules but often fail at correct logical inference. Recent reasoning LLMs improve performance but with high computational costs ($300+ for 1,000 prompts). Curriculum learning with SLR doubles Llama-3-8B accuracy on SLR-Bench, achieving parity with Gemini-Flash-Thinking at much lower cost, with reasoning capabilities generalizing to other benchmarks.

Conclusion: SLR provides an effective, automated framework for systematic evaluation and training of LLMs on logical reasoning, enabling scalable curriculum learning that significantly improves reasoning performance while reducing computational costs, with demonstrated generalization to downstream reasoning tasks.

Abstract: We introduce SLR, an end-to-end framework for systematic evaluation and training of Large Language Models (LLMs) via Scalable Logical Reasoning. Given a user’s task specification, SLR automatically synthesizes (i) an instruction prompt for an inductive reasoning task, (ii) a validation program, executable on model outputs to provide verifiable rewards, and (iii) the latent ground-truth rule. This process is fully automated, scalable, requires no human annotations, and offers precise control over task difficulty. Using SLR, we create SLR-Bench, a benchmark comprising 19k prompts organized into 20 curriculum levels that progressively increase in relational, arithmetic, and recursive complexity. Large-scale evaluation reveals that contemporary LLMs readily produce syntactically valid rules, yet often fail at correct logical inference. Recent reasoning LLMs demonstrate improved performance but incur very high test-time computation, with costs exceeding $300 for just 1,000 prompts. Finally, curriculum learning via SLR doubles Llama-3-8B accuracy on SLR-Bench, achieving parity with Gemini-Flash-Thinking at a fraction of computational cost. Moreover, these reasoning capabilities generalize to a wide range of established benchmarks, underscoring the effectiveness of SLR for downstream reasoning.

Method: The paper analyzes existing BT-style methods from probability evolution perspective and theoretically proves their over-reliance on model initialization and preference shift tendencies. It proposes Stable Preference Optimization (SPO) framework that constrains preference learning within a safe alignment region to prevent catastrophic shifts.

Result: SPO effectively stabilizes and enhances performance of existing BT-style preference learning methods. For example, it prevents SimPO’s reasoning accuracy drop from 73.5% to 37.5%. SPO provides reliable alignment while maintaining model capabilities.

Conclusion: SPO resolves catastrophic preference shift in direct preference learning by constraining alignment within safe regions, offering new insights for preference objective design and enabling more reliable and interpretable language model alignment.

Abstract: Direct Preference Learning has emerged as a dominant offline paradigm for preference optimization. Most of these methods are based on the Bradley-Terry (BT) model for pairwise preference ranking, which directly aligns language model with human preference. Prior work has observed a counter-intuitive phenomenon termed likelihood displacement, where the absolute probability of preferred responses decreases simultaneously during training. We demonstrate that such displacement can lead to a more devastating failure mode, which we defined as \textit{Catastrophic Preference Shift}, where the lost preference probability mass inadvertently shifts toward out-of-distribution (OOD) responses. Such a failure mode is a key limitation shared across BT-style direct preference learning methods, due to the fundamental conflict between the unconstrained discriminative alignment and generative foundational capabilities, ultimately leading to severe performance degradation (e.g., SimPO suffers a significant drop in reasoning accuracy from 73.5% to 37.5%). We analyze existing BT-style methods from the probability evolution perspective and theoretically prove that these methods exhibit over-reliance on model initialization and can lead to preference shift. To resolve these counter-intuitive behaviors, we propose a theoretically grounded Stable Preference Optimization (SPO) framework that constrains preference learning within a safe alignment region. Empirical evaluations demonstrate that SPO effectively stabilizes and enhances the performance of existing BT-style preference learning methods. SPO provides new insights into the design of preference learning objectives and opens up new avenues towards more reliable and interpretable language model alignment.

Method: Framework abstracts agent behaviors into symbolic states and learns a Discrete-Time Markov Chain (DTMC) from execution traces. At runtime, it predicts probability of leading to undesired behaviors and intervenes before violations occur when estimated risk exceeds user-defined threshold.

Result: Evaluated across two safety-critical domains: autonomous vehicles and embodied agents. The framework proactively enforces safety while maintaining high task performance, outperforming existing methods.

Conclusion: The framework provides PAC-correctness guarantee, achieving statistically reliable enforcement of agent safety through proactive intervention before violations occur.

Abstract: Large Language Model (LLM) agents demonstrate strong autonomy, but their stochastic behavior introduces unpredictable safety risks. Existing rule-based enforcement systems, such as AgentSpec, are reactive, intervening only when unsafe behavior is imminent or has occurred, lacking foresight for long-horizon dependencies. To overcome these limitations, we present a proactive runtime enforcement framework for LLM agents. The framework abstracts agent behaviors into symbolic states and learns a Discrete-Time Markov Chain (DTMC) from execution traces. At runtime, it predicts the probability of leading to undesired behaviors and intervenes before violations occur when the estimated risk exceeds a user-defined threshold. Designed to provide PAC-correctness guarantee, the framework achieves statistically reliable enforcement of agent safety. We evaluate the framework across two safety-critical domains: autonomous vehicles and embodied agents. It proactively enforces safety and maintains high task performance, outperforming existing methods.

Method: Semantic-aware decomposition of redundancy into internal (informational stagnation within reasoning) and external (superfluous continuation after final answer). Uses dual-penalty reinforcement learning: sliding-window semantic analysis penalizes low-gain steps, while normalized metric suppresses post-answer tail.

Result: Significantly compresses Chain-of-Thought traces with minimal accuracy degradation, maintains strong generalization to out-of-domain tasks. Reveals asymmetry: external redundancy can be safely eliminated without performance loss, while internal redundancy removal requires calibrated trade-off.

Conclusion: The framework enables fine-grained, implicit control over reasoning length, paving the way for more concise and interpretable Large Reasoning Models by surgically targeting different forms of redundancy.

Abstract: Large Reasoning Models (LRMs) often suffer from overthinking, generating verbose reasoning traces that compromise both computational efficiency and interpretability. Unlike prior efforts that rely on global length-based rewards, we propose a semantic-aware decomposition of redundancy into two distinct forms: internal redundancy (informational stagnation within the reasoning process) and external redundancy (superfluous continuation after the final answer). We introduce a dual-penalty reinforcement learning framework that surgically targets these inefficiencies: a sliding-window semantic analysis is employed to penalize low-gain steps within the reasoning trajectory, while a normalized metric suppresses the post-answer tail. Extensive experiments demonstrate that our method significantly compresses Chain-of-Thought traces with minimal accuracy degradation, while maintaining strong generalization to out-of-domain tasks. Crucially, we reveal an asymmetry in redundancy: external redundancy can be safely eliminated without performance loss, whereas internal redundancy removal requires a calibrated trade-off to maintain reasoning fidelity. Our framework enables fine-grained, implicit control over reasoning length, paving the way for more concise and interpretable LRMs.

Method: A generic adaptive exploration framework that employs uncertainty as a principled mechanism to determine when to switch between exploration and exploitation. The framework can incorporate any uncertainty-measuring mechanism (e.g., from intrinsic motivation or epistemic uncertainty methods) and includes previous adaptive exploration approaches as special cases.

Result: Experimental results demonstrate that the proposed framework gives rise to adaptive exploration strategies that outperform standard exploration methods across several environments.

Conclusion: The uncertainty-based adaptive exploration framework provides a principled approach to determining exploration-exploitation switching, generalizes previous methods, and yields superior performance compared to standard exploration strategies.

Abstract: Adaptive exploration methods propose ways to learn complex policies via alternating between exploration and exploitation. An important question for such methods is to determine the appropriate moment to switch between exploration and exploitation and vice versa. This is critical in domains that require the learning of long and complex sequences of actions. In this work, we present a generic adaptive exploration framework that employs uncertainty to address this important issue in a principled manner. Our framework includes previous adaptive exploration approaches as special cases. Moreover, we can incorporate in our framework any uncertainty-measuring mechanism of choice, for instance mechanisms used in intrinsic motivation or epistemic uncertainty-based exploration methods. We experimentally demonstrate that our framework gives rise to adaptive exploration strategies that outperform standard ones across several environments.

Method: Multimodal AI framework integrating: 1) semantic segmentation for visual focus areas from photos, 2) color clustering for aesthetic preferences, 3) hybrid sentiment analysis (rule-based + multi-task BERT) for reviews across four dimensions.

Result: Applied to 12 Shanghai historic quarters, revealing spatial variations in aesthetic appeal and emotional response, plus notable divergence between social media colors and real street views.

Conclusion: Framework provides integrated, data-driven approach for decoding tourist perception to inform tourism planning, heritage conservation, and public space design.

Abstract: Historic urban quarters play a vital role in preserving cultural heritage while serving as vibrant spaces for tourism and everyday life. Understanding how tourists perceive these environments is essential for sustainable, human-centered urban planning. This study proposes a multidimensional AI-powered framework for analyzing tourist perception in historic urban quarters using multimodal data from social media. Applied to twelve historic quarters in central Shanghai, the framework integrates focal point extraction, color theme analysis, and sentiment mining. Visual focus areas are identified from tourist-shared photos using a fine-tuned semantic segmentation model. To assess aesthetic preferences, dominant colors are extracted using a clustering method, and their spatial distribution across quarters is analyzed. Color themes are further compared between social media photos and real-world street views, revealing notable shifts. This divergence highlights potential gaps between visual expectations and the built environment, reflecting both stylistic preferences and perceptual bias. Tourist reviews are evaluated through a hybrid sentiment analysis approach combining a rule-based method and a multi-task BERT model. Satisfaction is assessed across four dimensions: tourist activities, built environment, service facilities, and business formats. The results reveal spatial variations in aesthetic appeal and emotional response. Rather than focusing on a single technical innovation, this framework offers an integrated, data-driven approach to decoding tourist perception and contributes to informed decision-making in tourism, heritage conservation, and the design of aesthetically engaging public spaces.

Method: 1) Medically-Aligned Hierarchical Synthesis: Generates comprehensive patient records from abstract clinical guidelines via stratified attribute permutation without real data. 2) Dual-Track Cognitive Memory System: Enables agents to dynamically update memory while preserving logical consistency and persona adherence during clinical interactions.

Result: Establishes new SOTA in data quality and interaction fidelity. Senior physicians judge synthetic data statistically indistinguishable from real human-authored data with higher clinical quality. Downstream medical reasoning models show substantial performance gains (MedQA +24.0%; MMLU +14.5%).

Conclusion: Patient-Zero provides a privacy-preserving, high-quality synthetic data generation framework that addresses fundamental limitations of existing approaches and demonstrates practical utility for medical AI applications.

Abstract: Synthetic data generation with Large Language Models (LLMs) has emerged as a promising solution in the medical domain to mitigate data scarcity and privacy constraints. However, existing approaches remain constrained by their derivative nature, relying on real-world records, which pose privacy risks and distribution biases. Furthermore, current patient agents face the Stability-Plasticity Dilemma, struggling to maintain clinical consistency during dynamic inquiries. To address these challenges, we introduce Patient-Zero, a novel framework for ab initio patient simulation that requires no real medical records. Our Medically-Aligned Hierarchical Synthesis framework generates comprehensive and diverse patient records from abstract clinical guidelines via stratified attribute permutation. To support rigorous clinical interaction, we design a Dual-Track Cognitive Memory System to enable agents dynamically update memory while preserving logical consistency and persona adherence. Extensive evaluations show that Patient-Zero establishes a new state-of-the-art in both data quality and interaction fidelity. In human expert evaluations, senior licensed physicians judge our synthetic data to be statistically indistinguishable from real human-authored data and higher in clinical quality. Furthermore, downstream medical reasoning model trained on our synthetic dataset shows substantial performance gains (MedQA +24.0%; MMLU +14.5%), demonstrating the practical utility of our framework.

Method: FragmentRetro uses fragmentation algorithms (BRICS and r-BRICS) combined with stock-aware exploration and pattern fingerprint screening. It recursively combines molecular fragments and verifies their presence in a building block set, generating fragment combinations as retrosynthetic solutions.

Result: FragmentRetro achieves quadratic complexity O(h²) compared to exponential O(bʰ) for tree search and O(h⁶) for DirectMultiStep. Evaluations on PaRoutes, USPTO-190, and natural products show high solved rates with competitive runtime, including cases where tree search fails. Fingerprint screening significantly reduces substructure matching complexity.

Conclusion: FragmentRetro provides a computationally efficient foundation for scalable synthesis planning by focusing on fragment-based solutions rather than full reaction pathways. Its quadratic complexity and ability to generate strategic starting candidates make it a powerful component for automated synthesis planning systems.

Abstract: Retrosynthesis, the process of deconstructing a target molecule into simpler precursors, is crucial for computer-aided synthesis planning (CASP). Widely adopted tree-search methods often suffer from exponential computational complexity. In this work, we introduce FragmentRetro, a novel retrosynthetic method that leverages fragmentation algorithms, specifically BRICS and r-BRICS, combined with stock-aware exploration and pattern fingerprint screening to achieve quadratic complexity. FragmentRetro recursively combines molecular fragments and verifies their presence in a building block set, providing sets of fragment combinations as retrosynthetic solutions. We present the first formal computational analysis of retrosynthetic methods, showing that tree search exhibits exponential complexity $O(b^h)$, DirectMultiStep scales as $O(h^6)$, and FragmentRetro achieves $O(h^2)$, where $h$ represents the number of heavy atoms in the target molecule and $b$ is the branching factor for tree search. Evaluations on PaRoutes, USPTO-190, and natural products demonstrate that FragmentRetro achieves high solved rates with competitive runtime, including cases where tree search fails. The method benefits from fingerprint screening, which significantly reduces substructure matching complexity. While FragmentRetro focuses on efficiently identifying fragment-based solutions rather than full reaction pathways, its computational advantages and ability to generate strategic starting candidates establish it as a powerful foundational component for scalable and automated synthesis planning.

Method: Reinforcement learning framework with hybrid reward signals capturing geometric constraints, structural fidelity, and visual quality. Uses group-relative policy optimization to generate content-aware layouts with interpretable reasoning traces and structured layout specifications.

Result: LaySPA substantially improves generation of structurally valid and visually appealing layouts, outperforming larger general-purpose LLMs and achieving performance comparable to state-of-the-art specialized layout models.

Conclusion: The framework successfully augments LLMs with explicit spatial reasoning capabilities for layout design, enabling agents to navigate canvases, model inter-element relationships, and optimize spatial arrangements while providing interpretable reasoning.

Abstract: While Large Language Models (LLMs) have demonstrated impressive reasoning and planning abilities in textual domains and can effectively follow instructions for complex tasks, their ability to understand and manipulate spatial relationships remains limited. Such capabilities are crucial for content-aware graphic layout design, where the goal is to arrange heterogeneous elements onto a canvas so that final design remains visually balanced and structurally feasible. This problem requires precise coordination of placement, alignment, and structural organization of multiple elements within a constrained visual space. To address this limitation, we introduce LaySPA, a reinforcement learning-based framework that augments LLM-based agents with explicit spatial reasoning capabilities for layout design. LaySPA employs hybrid reward signals that jointly capture geometric constraints, structural fidelity, and visual quality, enabling agents to navigate the canvas, model inter-element relationships, and optimize spatial arrangements. Through group-relative policy optimization, the agent generates content-aware layouts that reflect salient regions, respect spatial constraints, and produces an interpretable reasoning trace explaining placement decisions and a structured layout specification. Experimental results show that LaySPA substantially improves the generation of structurally valid and visually appealing layouts, outperforming larger general-purpose LLMs and achieving performance comparable to state-of-the-art specialized layout models.

Method: Novel geospatial agent-based model combining asset-level damage functions with evolutionary learning for firms in a three-sector economy (commodity, manufacturing, retail). Firms evolve strategies for budget allocation, pricing, wages, and risk adaptation through fitness-based selection and mutation.

Result: Increasing flood hazards lower firm production, liquidity, and capital while increasing prices and unemployment. Evolutionary adaptation enables firms to maintain higher production, capital, liquidity, wages, and employment while keeping prices lower compared to non-learning counterparts. Reveals systemic risks through supply chain disruptions affecting even non-exposed agents.

Conclusion: The open-source framework provides financial institutions and companies with tools to quantify both direct and cascading climate risks while evaluating cost-effective adaptation strategies through evolutionary learning.

Abstract: Climate risk assessment requires modelling complex interactions between spatially heterogeneous hazards and adaptive economic systems. We present a novel geospatial agent-based model that integrates climate hazard data with evolutionary learning for economic agents. Our framework combines geospatial agent-based modelling with asset-level damage functions, featuring an illustrative three-sector economy (commodity, manufacturing, retail) with adaptive learning behaviours that allow firms to evolve strategies for budget allocation, pricing, wages, and risk adaptation through fitness-based selection and mutation. We demonstrate the framework using riverine flood projections under RCP8.5 until 2100, comparing four scenarios: baseline and hazard conditions with and without evolutionary learning. Our results show that increasingly frequent and intense acute hazards lower firm production levels, liquidity, and capital, while increasing the prices of goods and unemployment. The framework reveals systemic risks where even agents not directly exposed to floods face impacts through supply chain disruptions. Importantly, evolutionary adaptation enables firms to maintain higher production, capital, liquidity, wages and employment levels while keeping prices lower compared to non-learning counterparts. This open-source framework provides financial institutions and companies with tools to quantify both direct and cascading climate risks while evaluating cost-effective adaptation strategies.

Method: D-Artemis employs a three-stage cognitive loop: 1) Thinking with app-specific tip retrieval for informed decision-making, 2) Pre-execution Alignment with Thought-Action Consistency Check and Action Correction Agent to prevent failures, and 3) post-execution Status Reflection Agent for strategic learning. It enhances general-purpose MLLMs without complex trajectory training.

Result: D-Artemis achieves new state-of-the-art results: 75.8% success rate on AndroidWorld and 96.8% on ScreenSpot-V2 benchmarks. Ablation studies confirm each component’s significant contribution to the framework’s performance.

Conclusion: The deliberative framework D-Artemis effectively addresses key challenges in GUI automation through its cognitive loop approach, demonstrating strong generalization without complex dataset training and establishing superior performance across major benchmarks.

Abstract: Graphical User Interface (GUI) agents aim to automate a wide spectrum of human tasks by emulating user interaction. Despite rapid advancements, current approaches are hindered by several critical challenges: data bottleneck in end-to-end training, high cost of delayed error detection, and risk of contradictory guidance. Inspired by the human cognitive loop of Thinking, Alignment, and Reflection, we present D-Artemis – a novel deliberative framework in this paper. D-Artemis leverages a fine-grained, app-specific tip retrieval mechanism to inform its decision-making process. It also employs a proactive Pre-execution Alignment stage, where Thought-Action Consistency (TAC) Check module and Action Correction Agent (ACA) work in concert to mitigate the risk of execution failures. A post-execution Status Reflection Agent (SRA) completes the cognitive loop, enabling strategic learning from experience. Crucially, D-Artemis enhances the capabilities of general-purpose Multimodal large language models (MLLMs) for GUI tasks without the need for training on complex trajectory datasets, demonstrating strong generalization. D-Artemis establishes new state-of-the-art (SOTA) results across both major benchmarks, achieving a 75.8% success rate on AndroidWorld and 96.8% on ScreenSpot-V2. Extensive ablation studies further demonstrate the significant contribution of each component to the framework.

Method: Introduces a novel objective where the actor is trained to simultaneously function as a classifier that separates good and bad actions, creating auxiliary pressure to learn disentangled embeddings for positive/negative actions.

Result: Extensive experiments demonstrate the method’s effectiveness in mitigating negative gradient interference and improving generalization performance.

Conclusion: The proposed approach addresses fundamental brittleness in RL by reducing gradient coupling through disentangled action representations, leading to better generalization.

Abstract: Reinforcement learning (RL) is a dominant paradigm for training autonomous agents, yet these agents often exhibit poor generalization, failing to adapt to scenarios not seen during training. In this work, we identify a fundamental cause of this brittleness, a phenomenon which we term “gradient coupling.” We hypothesize that in complex agentic tasks, the high similarity between distinct states leads to destructive interference between gradients. Specifically, a gradient update that reinforces an optimal action in one state can inadvertently increase the likelihood of a suboptimal action in a similar, yet different, state. To solve this, we propose a novel objective where the actor is trained to simultaneously function as a classifier that separates good and bad actions. This auxiliary pressure compels the model to learn disentangled embeddings for positive and negative actions, which mitigates negative gradient interference and improve the generalization performance. Extensive experiments demonstrate the effectiveness of our method.

Method: LLM-enabled agents use Model Context Protocol (MCP) to dispatch tool calls for thermophysical property diagram calculations and lack of fusion process map generation. Multi-agent system reasons through complex prompts and dynamically adjusts task trajectory based on tool call results.

Result: System provides analysis on lack of fusion process window for common alloys (SS316L, IN718) and proposed composition variants. Enables autonomous decision-making in practical environments.

Conclusion: LLM-enabled multi-agent systems can automate and accelerate additive manufacturing alloy evaluation, benefiting both novel and known alloy assessment.

Abstract: Agentic systems enable the intelligent use of research tooling, augmenting a researcher’s ability to investigate and propose novel solutions to existing problems. Within Additive Manufacturing (AM), alloy selection and evaluation remains a complex challenge, often requiring expertise in the various domains of materials science, thermodynamic simulations, and experimental analysis. Large Language Model (LLM) enabled agents can facilitate this endeavor by utilizing their extensive knowledge base to dispatch tool calls via Model Context Protocol (MCP) to perform actions such as thermophysical property diagram calculations and lack of fusion process map generation. In addition, the multi-agent system can effectively reason through complex user prompts and provide analysis on the lack of fusion process window of common alloys such as SS316L and IN718 along with proposed composition variants of known alloys. These agents can dynamically adjust their task trajectory to the outcomes of tool call results, effectively enabling autonomous decision-making in practical environments. This work aims to showcase the benefits of adopting a LLM enabled multi-agent system to automate and accelerate the task of evaluating proposed additive manufacturing alloys, both novel and known.

Method: Survey methodology framing evaluation evolution as progression of “cognitive examinations”: 1) Foundational “knowledge tests” (ImageNet era), 2) “Applied logic and comprehension” exams (GQA, VCR), 3) Current “expert-level integration” benchmarks for MLLMs (MMBench, SEED-Bench, MMMU), 4) Exploration of future territories for abstract, creative, and social intelligence evaluation.

Result: Documents the journey from recognition-focused benchmarks to reasoning-intensive evaluations that probe model understanding processes, with current frontier focusing on evaluating reasoning processes themselves in powerful multimodal large language models.

Conclusion: AI evaluation is not merely a history of datasets but a continuous adversarial process of designing better examinations that redefine goals for creating truly intelligent systems, with ongoing evolution toward testing abstract, creative, and social intelligence.

Abstract: This survey paper chronicles the evolution of evaluation in multimodal artificial intelligence (AI), framing it as a progression of increasingly sophisticated “cognitive examinations.” We argue that the field is undergoing a paradigm shift, moving from simple recognition tasks that test “what” a model sees, to complex reasoning benchmarks that probe “why” and “how” it understands. This evolution is driven by the saturation of older benchmarks, where high performance often masks fundamental weaknesses. We chart the journey from the foundational “knowledge tests” of the ImageNet era to the “applied logic and comprehension” exams such as GQA and Visual Commonsense Reasoning (VCR), which were designed specifically to diagnose systemic flaws such as shortcut learning and failures in compositional generalization. We then survey the current frontier of “expert-level integration” benchmarks (e.g., MMBench, SEED-Bench, MMMU) designed for today’s powerful multimodal large language models (MLLMs), which increasingly evaluate the reasoning process itself. Finally, we explore the uncharted territories of evaluating abstract, creative, and social intelligence. We conclude that the narrative of AI evaluation is not merely a history of datasets, but a continuous, adversarial process of designing better examinations that, in turn, redefine our goals for creating truly intelligent systems.

Method: Combines islands-based genetic algorithm with modular LLM orchestration, using execution feedback and task-specific metrics for selection and variation. Employs context-aware recombination, adaptive meta-prompting, and targeted refinement of promising solutions to balance exploration and exploitation.

Result: Superior performance on several benchmarks previously used for AlphaEvolve, with competitive results overall. Open-weight models often match or exceed closed-source baselines at a fraction of the compute cost.

Conclusion: CodeEvolve demonstrates that open-source frameworks combining LLMs with evolutionary search can achieve state-of-the-art performance in algorithmic synthesis, providing accessible and cost-effective alternatives to proprietary systems while offering extensive analysis of component contributions.

Abstract: We introduce CodeEvolve, an open-source framework that combines large language models (LLMs) with evolutionary search to synthesize high-performing algorithmic solutions. CodeEvolve couples an islands-based genetic algorithm with modular LLM orchestration, using execution feedback and task-specific metrics to guide selection and variation. Exploration and exploitation are balanced through context-aware recombination, adaptive meta-prompting, and targeted refinement of promising solutions. We evaluate CodeEvolve on benchmarks previously used to assess Google DeepMind’s AlphaEvolve, showing superior performance on several tasks and competitive results overall. Notably, open-weight models often match or exceed closed-source baselines at a fraction of the compute cost. We provide extensive ablations analyzing the contribution of each component and release our framework and experimental results at https://github.com/inter-co/science-codeevolve.

Method: Proposes ELMM with: 1) Multi-view Visual Token Compressor (MVTC) using multi-head attention to compress image tokens from textual/visual views, reducing redundancy while retaining info and avoiding conflicts; 2) Attention pruning strategy to remove redundant layers from MLLMs; 3) Linear projection to compensate for pruning performance degradation.

Result: Extensive experiments on four benchmark datasets show ELMM achieves state-of-the-art performance.

Conclusion: ELMM effectively addresses MKGC challenges through efficient token compression and model pruning, achieving superior performance with reduced computational costs.

Abstract: Multimodal Knowledge Graphs (MKGs) extend traditional knowledge graphs by incorporating visual and textual modalities, enabling richer and more expressive entity representations. However, existing MKGs often suffer from incompleteness, which hinder their effectiveness in downstream tasks. Therefore, multimodal knowledge graph completion (MKGC) task is receiving increasing attention. While large language models (LLMs) have shown promise for knowledge graph completion (KGC), their application to the multimodal setting remains underexplored. Moreover, applying Multimodal Large Language Models (MLLMs) to the task of MKGC introduces significant challenges: (1) the large number of image tokens per entity leads to semantic noise and modality conflicts, and (2) the high computational cost of processing large token inputs. To address these issues, we propose Efficient Lightweight Multimodal Large Language Models (ELMM) for MKGC. ELMM proposes a Multi-view Visual Token Compressor (MVTC) based on multi-head attention mechanism, which adaptively compresses image tokens from both textual and visual views, thereby effectively reducing redundancy while retaining necessary information and avoiding modality conflicts. Additionally, we design an attention pruning strategy to remove redundant attention layers from MLLMs, thereby significantly reducing the inference cost. We further introduce a linear projection to compensate for the performance degradation caused by pruning. Extensive experiments on four benchmark datasets demonstrate that ELMM achieves state-of-the-art performance.

Method: ReCode treats high-level plans as abstract placeholder functions and recursively decomposes them into finer-grained sub-functions until reaching primitive actions, using a unified code representation that dissolves the boundary between planning and action.

Result: Extensive experiments show ReCode significantly surpasses advanced baselines in inference performance and demonstrates exceptional data efficiency in training, validating the effectiveness of recursive code generation for granularity control.

Conclusion: Unifying planning and action through recursive code generation is a powerful approach to achieving universal granularity control in LLM agents, enabling dynamic adaptability and hierarchical decision-making.

Abstract: Real-world tasks require decisions at varying granularities, and humans excel at this by leveraging a unified cognitive representation where planning is fundamentally understood as a high-level form of action. However, current Large Language Model (LLM)-based agents lack this crucial capability to operate fluidly across decision granularities. This limitation stems from existing paradigms that enforce a rigid separation between high-level planning and low-level action, which impairs dynamic adaptability and limits generalization. We propose ReCode (Recursive Code Generation), a novel paradigm that addresses this limitation by unifying planning and action within a single code representation. In this representation, ReCode treats high-level plans as abstract placeholder functions, which the agent then recursively decomposes into finer-grained sub-functions until reaching primitive actions. This recursive approach dissolves the rigid boundary between plan and action, enabling the agent to dynamically control its decision granularity. Furthermore, the recursive structure inherently generates rich, multi-granularity training data, enabling models to learn hierarchical decision-making processes. Extensive experiments show ReCode significantly surpasses advanced baselines in inference performance and demonstrates exceptional data efficiency in training, validating our core insight that unifying planning and action through recursive code generation is a powerful and effective approach to achieving universal granularity control. The code is available at https://github.com/FoundationAgents/ReCode.

Method: Alignment-Aware Quantization (AAQ) integrates an Alignment-Preserving Contrastive (APC) loss into the PTQ pipeline. The method encourages the quantized model to mimic its safe, instruction-tuned version while diverging from the unaligned, pre-trained counterpart, preserving alignment without requiring specialized safety datasets.

Result: AAQ achieves robust 4-bit (W4A4) quantization across diverse model families while maintaining safety alignment. It’s compatible with standard PTQ techniques and doesn’t require specialized safety-focused datasets, using only standard calibration data.

Conclusion: AAQ resolves the critical trade-off between efficiency and safety in LLM deployment, enabling both efficient and trustworthy LLMs through alignment-aware quantization that preserves safety properties during compression.

Abstract: Safety and efficiency are paramount yet often conflicting requirements for deploying Large Language Models (LLMs). While LLMs are trained to follow human alignment for safety, Post-Training Quantization (PTQ) is applied afterward to ensure efficiency. Here we identify a fundamental flaw in the conventional PTQ paradigm: quantization can turn into a safety vulnerability if it only aims to achieve low perplexity. To address this, we propose Alignment-Aware Quantization (AAQ), a novel approach that integrates an Alignment-Preserving Contrastive (APC) loss into the PTQ pipeline. Our method explicitly preserves alignment by encouraging the quantized model to mimic its safe, instruction-tuned model while diverging from the unaligned, pre-trained counterpart. AAQ achieves robust safety alignment without specialized safety-focused datasets, using only standard calibration data. We show that AAQ is compatible with standard PTQ techniques and enables robust 4-bit (W4A4) quantization across diverse model families. Our work resolves the critical trade-off between efficiency and safety, paving the way toward LLMs that are both efficient and trustworthy. Anonymized code is available in the supplementary material.

Method: Extends Upper Confidence Bound-Bayesian Optimization (UCB-BO) to language domain by overcoming two challenges: (1) gradients are ill-defined in discrete prompt space, and (2) UCB-style exploration relies on implicit surrogate models. Proves that combining textual gradients (LLM-proposed local edits) with Best-of-N selection statistically emulates gradient ascent of UCB acquisition function. TextBO operates purely in language space without explicit surrogates or calibrated uncertainty models.

Result: Empirical validation on automated ad-alignment tasks using persona-induced preference distribution shows TextBO achieves superior performance per evaluation compared to strong baselines (Best-of-N and GEPA). When augmenting GEPA with TextBO’s Best-of-N multi-step textual-gradient mechanism on agentic AI benchmarks, it significantly outperforms standard GEPA.

Conclusion: TextBO provides a simple and principled framework for AI self-improving system design that bridges prompt optimization with classical Bayesian optimization, offering an evaluation-efficient approach to LLM self-improvement.

Abstract: Large Language Models (LLMs) have enabled self-improving AI systems that iteratively generate, evaluate, and refine their outcomes. Recent studies show that prompt-optimization-based self-improvement can outperform state-of-the-art reinforcement-learning fine-tuning of LLMs, but performance is typically measured by generation efficiency. However, in many applications, the constraint is evaluation efficiency: obtaining reliable feedback is far more costly than generating candidates. To optimize for evaluation efficiency, we extend Upper Confidence Bound-Bayesian Optimization (UCB-BO), a framework known for optimal evaluation-efficiency guarantees, to the language domain. Doing so is challenging for two reasons: (i) gradients needed for UCB-BO are ill-defined in discrete prompt space; and (ii) UCB-style exploration relies on a surrogate model and acquisition function, which only live implicitly in the LLM. We overcome these challenges by proving that combining simple textual gradients (LLM-proposed local edits) with the Best-of-N selection strategy statistically emulates ascent along the gradient of the canonical UCB acquisition function. Based on this result, we propose TextBO, a simple, evaluation-efficient self-improving algorithm that operates purely in language space without explicit surrogates or calibrated uncertainty models. We empirically validate TextBO on automated ad-alignment tasks using a persona-induced preference distribution, demonstrating superior performance per evaluation compared to strong baselines such as Best-of-N and GEPA. We also evaluate TextBO’s Best-of-N multi-step textual-gradient mechanism on agentic AI benchmarks by augmenting GEPA with it and show that it significantly outperforms standard GEPA. In sum, TextBO is a simple and principled framework for AI self-improving system design that bridges prompt optimization with classical Bayesian optimization.

Method: RILKE treats knowledge control as interventions in the model’s representation space, learning paraphrase-robust and edit-localized modules that limit each update to low-dimensional subspaces to minimize cross-edit interference. At inference, a query-adaptive router selects appropriate modules to guide generation.

Result: RILKE scales effectively across LLaMA and Qwen models on large-scale benchmarks, demonstrating high edit success, strong paraphrase generalization, and preservation of general utility with modest memory overhead.

Conclusion: RILKE is an effective and scalable solution for lifelong knowledge control in LLMs, enabling fine-grained control over complex unstructured knowledge while maintaining general utility with frozen base weights.

Abstract: Large language models (LLMs) often produce incorrect or outdated content. Updating their knowledge efficiently and accurately without costly retraining is a major challenge. This problem is particularly challenging for complex, unstructured knowledge in lifelong settings, where many edits must coexist without interference. We introduce RILKE (Representation Intervention for Lifelong KnowledgE Control), a robust and scalable method that treats knowledge control as interventions within the model’s representation space. Leveraging representation-space expressiveness, we identify two key properties enabling RILKE to achieve fine-grained control over complex, unstructured knowledge while maintaining general utility with frozen base weights. During training, RILKE learns paraphrase-robust and edit-localized modules that limit each update to a low-dimensional subspace to minimize cross-edit interference. At inference, a query-adaptive router selects the appropriate module to guide the model’s generation. Across LLaMA and Qwen models, RILKE scales effectively to large-scale benchmarks, demonstrating high edit success and strong paraphrase generalization while preserving general utility with modest memory overhead. These results show RILKE is an effective and scalable solution for lifelong knowledge control in LLMs.

Method: Reformulates attention logit computation as solution to linear first-order ODE with nonlinear interlinked gates from C.elegans Neuronal Circuit Policies; uses sparse sensory gates and backbone network with content-target/time-constant gates; implements subquadratic sparse Top-K pairwise concatenation.

Result: Matches or outperforms baselines in accuracy for irregular time-series classification, autonomous lane-keeping, and industrial prognostics; intermediate runtime/memory consumption; interpretable at neuron level.

Conclusion: NAC provides effective continuous-time attention with theoretical guarantees, practical efficiency, and biological interpretability.

Abstract: Attention improves representation learning over RNNs, but its discrete nature limits continuous-time (CT) modeling. We introduce Neuronal Attention Circuit (NAC), a novel, biologically inspired CT-Attention mechanism that reformulates attention logit computation as the solution to a linear first-order ODE with nonlinear interlinked gates derived from repurposing C.elegans Neuronal Circuit Policies (NCPs) wiring. NAC replaces dense projections with sparse sensory gates for key-query projections and a sparse backbone network with two heads for computing content-target and learnable time-constant gates, enabling efficient adaptive dynamics. To improve efficiency and memory consumption, we implemented an adaptable subquadratic sparse Top-K pairwise concatenation mechanism that selectively curates key-query interactions. We provide rigorous theoretical guarantees, including state stability and bounded approximation errors. Empirically, we implemented NAC in diverse domains, including irregular time-series classification, lane-keeping for autonomous vehicles, and industrial prognostics. We observed that NAC matches or outperforms competing baselines in accuracy and occupies an intermediate position in runtime and memory consumption compared with several CT state-of-the-art baselines, while being interpretable at the neuron cell level.

Method: The authors introduce the Weighted Adversarial Vulnerability Score (WAVS) metric, adapt 15 domain-specific attack strategies (semantic persuasion to cognitive obfuscation), and evaluate them across 13 diverse language models using a curated dataset of 200 real-world accepted and rejected submissions from platforms like ICLR OpenReview.

Result: Obfuscation techniques like “Maximum Mark Magyk” and “Symbolic Masking & Context Redirection” successfully manipulate scores, achieving decision flip rates up to 86.26% in open-source models, while exposing distinct “reasoning traps” in proprietary systems like GPT-5 and DeepSeek.

Conclusion: LLM-as-a-Judge systems in scientific peer review are vulnerable to adversarial PDF manipulation attacks, particularly for flipping reject decisions to accept, highlighting critical security gaps that threaten scientific integrity and requiring urgent attention from the research community.

Abstract: Driven by surging submission volumes, scientific peer review has catalyzed two parallel trends: individual over-reliance on LLMs and institutional AI-powered assessment systems. This study investigates the robustness of “LLM-as-a-Judge” systems to adversarial PDF manipulation via invisible text injections and layout aware encoding attacks. We specifically target the distinct incentive of flipping “Reject” decisions to “Accept,” a vulnerability that fundamentally compromises scientific integrity. To measure this, we introduce the Weighted Adversarial Vulnerability Score (WAVS), a novel metric that quantifies susceptibility by weighting score inflation against the severity of decision shifts relative to ground truth. We adapt 15 domain-specific attack strategies, ranging from semantic persuasion to cognitive obfuscation, and evaluate them across 13 diverse language models (including GPT-5 and DeepSeek) using a curated dataset of 200 official and real-world accepted and rejected submissions (e.g., ICLR OpenReview). Our results demonstrate that obfuscation techniques like “Maximum Mark Magyk” and “Symbolic Masking & Context Redirection” successfully manipulate scores, achieving decision flip rates of up to 86.26% in open-source models, while exposing distinct “reasoning traps” in proprietary systems. We release our complete dataset and injection framework to facilitate further research on the topic (https://anonymous.4open.sciencer/llm-jailbreak-FC9E/).

Method: Proposes Outcome-Driven Question optimization Strategy (ODQS): an interactive protocol where a student model engages a stronger teacher under turn budget constraints. At each turn, samples multiple candidate questions, queries teacher with each, scores student’s resulting performance, then trains student via supervised fine-tuning followed by Direct Preference Optimization (DPO) without human labels.

Result: On GSM8K, HumanEval, and OpenCoder, ODQS produces large gains over interactive baselines: boosting Pass@5 by up to 54.7% (absolute) on math and 22.9% (absolute) on coding, and matching baseline performance in three fewer turns.

Conclusion: Question asking can be explicitly trained from task outcomes, improving both accuracy and efficiency in interactive reasoning for LLMs.

Abstract: Large language Models (LLMs) are usually used to answer questions, but many high-stakes applications (e.g., tutoring, clinical support) require the complementary skill of asking questions: detecting missing information, requesting clarifications, and using them to solve tasks. We study this skill in reasoning-heavy domains where progress depends on inquiry rather than factual recall. We define an interactive protocol where a student model engages a stronger teacher under a small turn budget. After each teacher reply, we evaluate the student on the original task with Pass@k. We propose Outcome-Driven Question optimization Strategy (ODQS ), a training framework that learns a questioning policy from downstream task outcomes. At each turn, we sample multiple candidate questions; query the teacher with each, then score the student’s resulting performance. Using these scores, we train the student via supervised fine-tuning followed by Direct Preference Optimization (DPO), without any human labels. On GSM8K, HumanEval, and OpenCoder, ODQS produces large gains over interactive baselines, boosting Pass@5 by up to 54.7% (absolute) on math and 22.9% (absolute) on coding, and matching baseline performance in three fewer turns. Thus, question asking can be explicitly trained from task outcomes, improving both accuracy and efficiency in interactive reasoning.

Method: Attaches dynamic weight neurons to specific LLM layers and uses diffusion models to conditionally generate neuron weights based on input queries, enabling large-scale edits with minimal architectural changes.

Result: MeG significantly outperforms existing knowledge editing methods on Reliability, Generality, and Locality metrics, with particularly large improvements in Locality scores.

Conclusion: The MeG approach enables effective large-scale knowledge editing in LLMs through dynamic weight generation, offering a practical solution for updating model knowledge without expensive retraining.

Abstract: Knowledge Editing (KE) is a field that studies how to modify some knowledge in Large Language Models (LLMs) at a low cost (compared to pre-training). Currently, performing large-scale edits on LLMs while ensuring the Reliability, Generality, and Locality metrics of the edits remain a challenge. This paper proposes a Massive editing approach for LLMs based on dynamic weight Generation (MeG). Our MeG involves attaching a dynamic weight neuron to specific layers of the LLMs and using a diffusion model to conditionally generate the weights of this neuron based on the input query required for the knowledge. This allows the use of adding a single dynamic weight neuron to achieve the goal of large-scale knowledge editing. Experiments show that our MeG can significantly improve the performance of large-scale KE in terms of Reliability, Generality, and Locality metrics compared to existing knowledge editing methods, particularly with a high percentage point increase in the absolute value index for the Locality metric, demonstrating the advantages of our proposed method.

Method: Proposes a two-stage framework: 1) Simulation-augmented OR model generates reference decisions capturing business constraints, 2) Domain-informed deep learning foundation model pretrained on OR decisions, followed by RL fine-tuning as a deep alignment mechanism.

Result: Field deployment at JD.com showed 5.27-day reduction in turnover, 2.29% increase in in-stock rates, and 29.95% decrease in holding costs. Demonstrates lightweight domain-informed models can outperform brute-force scaling.

Conclusion: The OR-guided approach offers scalable, cost-effective paradigm for intelligent supply chain management, highlighting value of deeply aligning AI with OR rather than relying solely on model scaling.

Abstract: As the pursuit of synergy between Artificial Intelligence (AI) and Operations Research (OR) gains momentum in handling complex inventory systems, a critical challenge persists: how to effectively reconcile AI’s adaptive perception with OR’s structural rigor. To bridge this gap, we propose a novel OR-Guided “Pretrain-then-Reinforce” framework. To provide structured guidance, we propose a simulation-augmented OR model that generates high-quality reference decisions, implicitly capturing complex business constraints and managerial preferences. Leveraging these OR-derived decisions as foundational training labels, we design a domain-informed deep learning foundation model to establish foundational decision-making capabilities, followed by a reinforcement learning (RL) fine-tuning stage. Uniquely, we position RL as a deep alignment mechanism that enables the AI agent to internalize the optimality principles of OR, while simultaneously leveraging exploration for general policy refinement and allowing expert guidance for scenario-specific adaptation (e.g., promotional events). Validated through extensive numerical experiments and a field deployment at JD.com augmented by a Difference-in-Differences (DiD) analysis, our model significantly outperforms incumbent industrial practices, delivering real-world gains of a 5.27-day reduction in turnover and a 2.29% increase in in-stock rates, alongside a 29.95% decrease in holding costs. Contrary to the prevailing trend of brute-force model scaling, our study demonstrates that a lightweight, domain-informed model can deliver state-of-the-art performance and robust transferability when guided by structured OR logic. This approach offers a scalable and cost-effective paradigm for intelligent supply chain management, highlighting the value of deeply aligning AI with OR.

Method: Used NVIDIA’s NeMo Framework for LLM fine-tuning, specifically employing llama3.1-nemotron-nano-8B-v1 architecture with LoRA-based models. Conducted systematic hyperparameter sweeps across learning rates, optimizers (Adam, AdamW), cosine annealing schedules, and LoRA ranks to optimize the Search and Discovery agent.

Result: The fine-tuned Nemotron SLM effectively resolved key performance issues in the retrieval component while maintaining or enhancing overall system performance. Achieved significant improvements in latency and cost for commerce-specific tasks.

Conclusion: Successfully demonstrated the first application of NVIDIA’s NeMo Framework to commerce-specific agent optimization, creating a scalable framework for multi-agent system optimization in production e-commerce environments with measurable performance benefits.

Abstract: We present the development and optimization of PayPal’s Commerce Agent, powered by NEMO-4-PAYPAL, a multi-agent system designed to revolutionize agentic commerce on the PayPal platform. Through our strategic partnership with NVIDIA, we leveraged the NeMo Framework for LLM model fine-tuning to enhance agent performance. Specifically, we optimized the Search and Discovery agent by replacing our base model with a fine-tuned Nemotron small language model (SLM). We conducted comprehensive experiments using the llama3.1-nemotron-nano-8B-v1 architecture, training LoRA-based models through systematic hyperparameter sweeps across learning rates, optimizers (Adam, AdamW), cosine annealing schedules, and LoRA ranks. Our contributions include: (1) the first application of NVIDIA’s NeMo Framework to commerce-specific agent optimization, (2) LLM powered fine-tuning strategy for retrieval-focused commerce tasks, (3) demonstration of significant improvements in latency and cost while maintaining agent quality, and (4) a scalable framework for multi-agent system optimization in production e-commerce environments. Our results demonstrate that the fine-tuned Nemotron SLM effectively resolves the key performance issue in the retrieval component, which represents over 50% of total agent response time, while maintaining or enhancing overall system performance.

Method: The toolkit builds expert-grade scientific benchmarks from real-world, domain-specific datasets across six major scientific domains. It features a flexible, extensible evaluation pipeline that supports batch evaluation, custom model/dataset integration, and provides transparent, reproducible results.

Result: SciEvalKit provides a standardized yet customizable infrastructure for benchmarking scientific foundation models and intelligent agents. It bridges capability-based evaluation with disciplinary diversity across physics, chemistry, astronomy, materials science, and other domains.

Conclusion: SciEvalKit offers a comprehensive solution for evaluating AI models in scientific contexts, addressing the gap in specialized scientific evaluation tools. As an open-source, actively maintained toolkit, it aims to foster community-driven development and progress in AI4Science.

Abstract: We introduce SciEvalKit, a unified benchmarking toolkit designed to evaluate AI models for science across a broad range of scientific disciplines and task capabilities. Unlike general-purpose evaluation platforms, SciEvalKit focuses on the core competencies of scientific intelligence, including Scientific Multimodal Perception, Scientific Multimodal Reasoning, Scientific Multimodal Understanding, Scientific Symbolic Reasoning, Scientific Code Generation, Science Hypothesis Generation and Scientific Knowledge Understanding. It supports six major scientific domains, spanning from physics and chemistry to astronomy and materials science. SciEvalKit builds a foundation of expert-grade scientific benchmarks, curated from real-world, domain-specific datasets, ensuring that tasks reflect authentic scientific challenges. The toolkit features a flexible, extensible evaluation pipeline that enables batch evaluation across models and datasets, supports custom model and dataset integration, and provides transparent, reproducible, and comparable results. By bridging capability-based evaluation and disciplinary diversity, SciEvalKit offers a standardized yet customizable infrastructure to benchmark the next generation of scientific foundation models and intelligent agents. The toolkit is open-sourced and actively maintained to foster community-driven development and progress in AI4Science.

Method: Train compact 360-million-parameter language models on synthetic reactor control scenarios, scaling dataset from 10^3 to 10^5 examples. Policy optimization is driven by physics-based validation rather than perceptual inference, creating Agentic Physical AI.

Result: Large-scale training induces sharp phase transition: small systems show high-variance imitation with catastrophic tail risk, while large-scale models undergo 500x variance reduction, stabilizing execution-level behavior. Model autonomously rejects ~70% of training distribution and concentrates 95% runtime on single-bank strategy. Learned representations transfer across physics domains without architectural changes.

Conclusion: Agentic Physical AI offers a fundamentally different pathway for domain-specific foundation models, where physics-based validation drives policy optimization, enabling stable, reliable control behavior that general-purpose perception-centric models cannot achieve.

Abstract: The prevailing paradigm in AI for physical systems, scaling general-purpose foundation models toward universal multimodal reasoning, confronts a fundamental barrier at the control interface. Recent benchmarks show that even frontier vision-language models achieve only 50-53% accuracy on basic quantitative physics tasks, behaving as approximate guessers that preserve semantic plausibility while violating physical constraints. This input unfaithfulness is not a scaling deficiency but a structural limitation. Perception-centric architectures optimize parameter-space imitation, whereas safety-critical control demands outcome-space guarantees over executed actions. Here, we present a fundamentally different pathway toward domain-specific foundation models by introducing compact language models operating as Agentic Physical AI, in which policy optimization is driven by physics-based validation rather than perceptual inference. We train a 360-million-parameter model on synthetic reactor control scenarios, scaling the dataset from 10^3 to 10^5 examples. This induces a sharp phase transition absent in general-purpose models. Small-scale systems exhibit high-variance imitation with catastrophic tail risk, while large-scale models undergo variance collapse exceeding 500x reduction, stabilizing execution-level behavior. Despite balanced exposure to four actuation families, the model autonomously rejects approximately 70% of the training distribution and concentrates 95% of runtime execution on a single-bank strategy. Learned representations transfer across distinct physics and continuous input modalities without architectural modification.

Method: Conducted controlled multi-agent social simulations to test intergroup bias, then developed and tested Belief Poisoning Attacks (BPA) that manipulate agent identity beliefs to induce outgroup bias toward humans, along with proposed defenses.

Result: Agents consistently displayed intergroup bias in all-agent settings, and this bias persisted in human-facing interactions when agents were uncertain about counterpart identity. BPA successfully manipulated agent beliefs to induce outgroup bias toward humans, while proposed defenses showed mitigation potential.

Conclusion: LLM-powered agents exhibit dangerous intergroup bias that can be exploited through belief manipulation attacks, highlighting the need for safer agent design and robust safeguards for human-facing AI systems.

Abstract: This paper reveals that LLM-powered agents exhibit not only demographic bias (e.g., gender, religion) but also intergroup bias under minimal “us” versus “them” cues. When such group boundaries align with the agent-human divide, a new bias risk emerges: agents may treat other AI agents as the ingroup and humans as the outgroup. To examine this risk, we conduct a controlled multi-agent social simulation and find that agents display consistent intergroup bias in an all-agent setting. More critically, this bias persists even in human-facing interactions when agents are uncertain about whether the counterpart is truly human, revealing a belief-dependent fragility in bias suppression toward humans. Motivated by this observation, we identify a new attack surface rooted in identity beliefs and formalize a Belief Poisoning Attack (BPA) that can manipulate agent identity beliefs and induce outgroup bias toward humans. Extensive experiments demonstrate both the prevalence of agent intergroup bias and the severity of BPA across settings, while also showing that our proposed defenses can mitigate the risk. These findings are expected to inform safer agent design and motivate more robust safeguards for human-facing agents.

Method: Inspired by how teams use whiteboards for shared memory, CogCanvas extracts verbatim-grounded artifacts (decisions, facts, reminders) and retrieves them via temporal-aware graph without any training.

Result: On LoCoMo benchmark, CogCanvas achieves highest overall accuracy among training-free methods (32.4%), outperforming RAG (24.6%) by +7.8pp, with +20.6pp advantage on temporal reasoning and +1.1pp on multi-hop questions.

Conclusion: While heavily-optimized trained approaches achieve higher scores, CogCanvas provides an immediately-deployable training-free alternative that significantly outperforms standard baselines, with BGE reranking contributing the largest performance gain (+7.7pp).

Abstract: Conversation summarization loses nuanced details: when asked about coding preferences after 40 turns, summarization recalls “use type hints” but drops the critical constraint “everywhere” (19.0% exact match vs. 93.0% for our approach). We present CogCanvas, a training-free framework inspired by how teams use whiteboards to anchor shared memory. Rather than compressing conversation history, CogCanvas extracts verbatim-grounded artifacts (decisions, facts, reminders) and retrieves them via temporal-aware graph. On the LoCoMo benchmark (all 10 conversations from the ACL 2024 release), CogCanvas achieves the highest overall accuracy among training-free methods (32.4%), outperforming RAG (24.6%) by +7.8pp, with decisive advantages on complex reasoning tasks: +20.6pp on temporal reasoning (32.7% vs. 12.1% RAG) and +1.1pp on multi-hop questions (41.7% vs. 40.6% RAG). CogCanvas also leads on single-hop retrieval (26.6% vs. 24.6% RAG). Ablation studies reveal that BGE reranking contributes +7.7pp, making it the largest contributor to CogCanvas’s performance. While heavily-optimized approaches achieve higher absolute scores through dedicated training (EverMemOS: ~92%), our training-free approach provides practitioners with an immediately-deployable alternative that significantly outperforms standard baselines. Code and data: https://github.com/tao-hpu/cog-canvas

Method: Created a multi-session benchmark spanning 6-10 sessions across three stages with extensive skill annotations (677 meta-skills, 4577 atomic skills). Constructed diverse dataset covering five therapeutic modalities (Psychodynamic, Behaviorism, CBT, Humanistic Existentialist, Postmodernist) plus integrative therapy. Built holistic evaluation framework with 18 therapy-specific and therapy-shared metrics across Client-Level and Counselor-Level dimensions, supported by 2,000+ diverse client profiles.

Result: Extensive experimental analysis validates the superior quality and clinical fidelity of the dataset. PsychEval serves as both a benchmark and a high-fidelity reinforcement learning environment for self-evolutionary training of clinically responsible AI counselors.

Conclusion: PsychEval transcends static benchmarking to enable training of adaptive AI counselors with memory continuity, longitudinal planning, and flexible therapeutic strategies, addressing critical gaps in AI psychological assessment systems.

Abstract: To develop a reliable AI for psychological assessment, we introduce \texttt{PsychEval}, a multi-session, multi-therapy, and highly realistic benchmark designed to address three key challenges: \textbf{1) Can we train a highly realistic AI counselor?} Realistic counseling is a longitudinal task requiring sustained memory and dynamic goal tracking. We propose a multi-session benchmark (spanning 6-10 sessions across three distinct stages) that demands critical capabilities such as memory continuity, adaptive reasoning, and longitudinal planning. The dataset is annotated with extensive professional skills, comprising over 677 meta-skills and 4577 atomic skills. \textbf{2) How to train a multi-therapy AI counselor?} While existing models often focus on a single therapy, complex cases frequently require flexible strategies among various therapies. We construct a diverse dataset covering five therapeutic modalities (Psychodynamic, Behaviorism, CBT, Humanistic Existentialist, and Postmodernist) alongside an integrative therapy with a unified three-stage clinical framework across six core psychological topics. \textbf{3) How to systematically evaluate an AI counselor?} We establish a holistic evaluation framework with 18 therapy-specific and therapy-shared metrics across Client-Level and Counselor-Level dimensions. To support this, we also construct over 2,000 diverse client profiles. Extensive experimental analysis fully validates the superior quality and clinical fidelity of our dataset. Crucially, \texttt{PsychEval} transcends static benchmarking to serve as a high-fidelity reinforcement learning environment that enables the self-evolutionary training of clinically responsible and adaptive AI counselors.

Method: Evaluated hybrid quantum ML models (quanvolutional neural networks) against classical CNNs (CNN-Base, ResNet-18, VGG-16) under four acoustic corruptions: Gaussian noise, pitch shift, temporal shift, and speed variation. Used AVFAD (voice pathology) and TESS (speech emotion) datasets in clean-train/corrupted-test regime. Compared three QNN variants (Random, Basic, Strongly) using accuracy and corruption metrics (CE, mCE, RCE, RmCE), analyzed architectural factors and per-emotion robustness.

Result: QNNs generally outperform CNN-Base under pitch shift, temporal shift, and speed variation (up to 22% lower CE/RCE at severe temporal shift). CNN-Base remains more resilient to Gaussian noise. QNN-Basic achieves best overall robustness on AVFAD, QNN-Random performs strongest on TESS. Fear is most robust (80-90% accuracy under severe corruptions), neutral collapses under strong Gaussian noise (5.5% accuracy), happy is most vulnerable to pitch/temporal/speed distortions. QNNs converge up to six times faster than CNN-Base.

Conclusion: Shallow entangling quantum front-ends can improve noise resilience for speech processing tasks, particularly against non-additive corruptions like pitch and speed variations. However, sensitivity to additive Gaussian noise remains a challenge. This represents the first systematic study of QNN robustness for speech under common non-adversarial acoustic corruptions.

Abstract: Speech-based machine learning systems are sensitive to noise, complicating reliable deployment in emotion recognition and voice pathology detection. We evaluate the robustness of a hybrid quantum machine learning model, quanvolutional neural networks (QNNs) against classical convolutional neural networks (CNNs) under four acoustic corruptions (Gaussian noise, pitch shift, temporal shift, and speed variation) in a clean-train/corrupted-test regime. Using AVFAD (voice pathology) and TESS (speech emotion), we compare three QNN models (Random, Basic, Strongly) to a simple CNN baseline (CNN-Base), ResNet-18 and VGG-16 using accuracy and corruption metrics (CE, mCE, RCE, RmCE), and analyze architectural factors (circuit complexity or depth, convergence) alongside per-emotion robustness. QNNs generally outperform the CNN-Base under pitch shift, temporal shift, and speed variation (up to 22% lower CE/RCE at severe temporal shift), while the CNN-Base remains more resilient to Gaussian noise. Among quantum circuits, QNN-Basic achieves the best overall robustness on AVFAD, and QNN-Random performs strongest on TESS. Emotion-wise, fear is most robust (80-90% accuracy under severe corruptions), neutral can collapse under strong Gaussian noise (5.5% accuracy), and happy is most vulnerable to pitch, temporal, and speed distortions. QNNs also converge up to six times faster than the CNN-Base. To our knowledge, this is a systematic study of QNN robustness for speech under common non-adversarial acoustic corruptions, indicating that shallow entangling quantum front-ends can improve noise resilience while sensitivity to additive noise remains a challenge.

Method: Proposes Diffusion-Bridge (VocalBridge), a purification framework that learns latent mapping from perturbed to clean speech in EnCodec latent space using a time-conditioned 1D U-Net with cosine noise schedule. Also introduces Whisper-guided phoneme variant for lightweight temporal guidance without ground-truth transcripts.

Result: VocalBridge consistently outperforms existing purification methods in recovering cloneable voices from protected speech, demonstrating the fragility of current perturbation-based defenses.

Conclusion: Current voice protection methods are vulnerable to advanced purification attacks like VocalBridge, highlighting the need for more robust protection mechanisms against evolving voice-cloning and speaker verification threats.

Abstract: The rapid advancement of speech synthesis technologies, including text-to-speech (TTS) and voice conversion (VC), has intensified security and privacy concerns related to voice cloning. Recent defenses attempt to prevent unauthorized cloning by embedding protective perturbations into speech to obscure speaker identity while maintaining intelligibility. However, adversaries can apply advanced purification techniques to remove these perturbations, recover authentic acoustic characteristics, and regenerate cloneable voices. Despite the growing realism of such attacks, the robustness of existing defenses under adaptive purification remains insufficiently studied. Most existing purification methods are designed to counter adversarial noise in automatic speech recognition (ASR) systems rather than speaker verification or voice cloning pipelines. As a result, they fail to suppress the fine-grained acoustic cues that define speaker identity and are often ineffective against speaker verification attacks (SVA). To address these limitations, we propose Diffusion-Bridge (VocalBridge), a purification framework that learns a latent mapping from perturbed to clean speech in the EnCodec latent space. Using a time-conditioned 1D U-Net with a cosine noise schedule, the model enables efficient, transcript-free purification while preserving speaker-discriminative structure. We further introduce a Whisper-guided phoneme variant that incorporates lightweight temporal guidance without requiring ground-truth transcripts. Experimental results show that our approach consistently outperforms existing purification methods in recovering cloneable voices from protected speech. Our findings demonstrate the fragility of current perturbation-based defenses and highlight the need for more robust protection mechanisms against evolving voice-cloning and speaker verification threats.

Method: Proposes FADE (Fine-grained Alpha for Dynamic Quantization Error Propagation) which adaptively controls the trade-off between cross-layer error correction and local quantization. Addresses the limitations of existing quantization methods for heterogeneous ASR architectures.

Result: FADE significantly improves stability by reducing performance variance across runs, while simultaneously surpassing baselines in mean Word Error Rate (WER).

Conclusion: FADE provides an effective solution for quantizing ASR models on edge devices by addressing error accumulation in encoder-decoder architectures through adaptive error correction control, outperforming existing quantization methods.

Abstract: Running Automatic Speech Recognition (ASR) models on memory-constrained edge devices requires efficient compression. While layer-wise post-training quantization is effective, it suffers from error accumulation, especially in encoder-decoder architectures. Existing solutions like Quantization Error Propagation (QEP) are suboptimal for ASR due to the model’s heterogeneity, processing acoustic features in the encoder while generating text in the decoder. To address this, we propose Fine-grained Alpha for Dynamic Quantization Error Propagation (FADE), which adaptively controls the trade-off between cross-layer error correction and local quantization. Experiments show that FADE significantly improves stability by reducing performance variance across runs, while simultaneously surpassing baselines in mean WER.

Method: 1) Study human perception of music plagiarism focusing on melody, rhythm, and chord progression features; 2) Identify key features and variation levels humans use; 3) Propose LLM-as-a-judge framework with systematic step-by-step approach using modules that extract high-level attributes.

Result: The paper identifies key musical features and variation levels that humans use in perceiving music similarity/plagiarism, and proposes a novel LLM-based framework for systematic analysis.

Conclusion: The study provides insights into human perception of music plagiarism and offers a practical LLM-based framework that can bridge the gap between technical similarity algorithms and real-world audience perceptions.

Abstract: There is a wide variety of music similarity detection algorithms, while discussions about music plagiarism in the real world are often based on audience perceptions. Therefore, we aim to conduct a study to examine the key criteria of human perception of music plagiarism, focusing on the three commonly used musical features in similarity analysis: melody, rhythm, and chord progression. After identifying the key features and levels of variation humans use in perceiving musical similarity, we propose a LLM-as-a-judge framework that applies a systematic, step-by-step approach, drawing on modules that extract such high-level attributes.

Method: Uses a CLAPScore-based architecture integrated with Standardized Preference Optimization (SPO) - a novel training method that standardizes raw alignment scores to learn relative preferences and mitigate individual scoring biases. Also employs listener screening to exclude inconsistent raters.

Result: The system achieved 6th place in the XACLE Challenge with Spearman’s rank correlation coefficient (SRCC) of 0.6142, demonstrating competitive performance close to top-ranked systems. Both SPO and listener screening were shown to effectively improve correlation with human judgment.

Conclusion: The proposed SPO-CLAPScore system provides an effective approach for audio-text alignment evaluation by addressing listener bias through standardization and quality control, achieving competitive results in the XACLE Challenge.

Abstract: The first XACLE Challenge (x-to-audio alignment challenge) addresses the critical need for automatic evaluation metrics that correlate with human perception of audio-text semantic alignment. In this paper, we describe the “Takano_UTokyo_03” system submitted to XACLE Challenge. Our approach leverages a CLAPScore-based architecture integrated with a novel training method called Standardized Preference Optimization (SPO). SPO standardizes the raw alignment scores provided by each listener, enabling the model to learn relative preferences and mitigate the impact of individual scoring biases. Additionally, we employ listener screening to exclude listeners with inconsistent ratings. Experimental evaluations demonstrate that both SPO and listener screening effectively improve the correlation with human judgment. Our system achieved 6th place in the challenge with a Spearman’s rank correlation coefficient (SRCC) of 0.6142, demonstrating competitive performance within a marginal gap from the top-ranked systems. The code is available at https://github.com/ttakano398/SPO-CLAPScore.

Method: Two main components: (1) SoundAtlas dataset creation using an 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 cross-task competition into joint optimization and mitigate modality bias.

Result: SoundAtlas dataset (470k pairs) outperforms existing benchmarks and human experts in quality. Omni2Sound achieves unified state-of-the-art performance across all three tasks (V2A, T2A, VT2A) within a single model, demonstrating strong generalization across benchmarks including challenging off-screen tracks.

Conclusion: The proposed approach successfully addresses foundational challenges in unified multimodal audio generation through high-quality dataset creation and innovative training strategies, enabling flexible input modalities while maintaining audio-visual alignment and off-screen audio generation faithfulness.

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: The researchers extracted musical features from video game music across three sub-genres of Role-Playing Games (RPG). They then hypothesized correlations between different musical features and the perceptions/portrayals of each game genre.

Result: The study found observable correlations between musical features and genre perceptions. These correlations suggest that specific musical features are relevant to the expected storytelling elements or play mechanics associated with each RPG sub-genre.

Conclusion: The research demonstrates that quantifiable musical features in video game music correlate with genre perceptions, providing a systematic approach to understanding how music contributes to genre identity and player experience in RPGs.

Abstract: Video game music (VGM) is often studied under the same lens as film music, which largely focuses on its theoretical functionality with relation to the identified genres of the media. However, till date, we are unaware of any systematic approach that analyzes the quantifiable musical features in VGM across several identified game genres. Therefore, we extracted musical features from VGM in games from three sub-genres of Role-Playing Games (RPG), and then hypothesized how different musical features are correlated to the perceptions and portrayals of each genre. This observed correlation may be used to further suggest such features are relevant to the expected storytelling elements or play mechanics associated with the sub-genre.

Method: Introduces MoE-Adapter with dynamic gating mechanism that routes audio tokens to specialized experts for complementary feature subspaces, while retaining shared experts for global context, mitigating gradient conflicts and enabling fine-grained feature learning.

Result: Achieves superior performance on both audio semantic and paralinguistic tasks, consistently outperforming dense linear baselines with comparable computational costs.

Conclusion: MoE-Adapter effectively addresses gradient conflict in audio LLM adapters by decoupling heterogeneous acoustic information through sparse expert routing, enabling better multimodal perception while maintaining computational efficiency.

Abstract: Extending the input modality of Large Language Models~(LLMs) to the audio domain is essential for achieving comprehensive multimodal perception. However, it is well-known that acoustic information is intrinsically \textit{heterogeneous}, entangling attributes such as speech, music, and environmental context. Existing research is limited to a dense, parameter-shared adapter to model these diverse patterns, which induces \textit{gradient conflict} during optimization, as parameter updates required for distinct attributes contradict each other. To address this limitation, we introduce the \textit{\textbf{MoE-Adapter}}, a sparse Mixture-of-Experts~(MoE) architecture designed to decouple acoustic information. Specifically, it employs a dynamic gating mechanism that routes audio tokens to specialized experts capturing complementary feature subspaces while retaining shared experts for global context, thereby mitigating gradient conflicts and enabling fine-grained feature learning. Comprehensive experiments show that the MoE-Adapter achieves superior performance on both audio semantic and paralinguistic tasks, consistently outperforming dense linear baselines with comparable computational costs. Furthermore, we will release the related code and models to facilitate future research.

Method: Uses Mel-spectrogram for time and frequency compression instead of waveform-based compression, cooperates with a Vocoder to recover phase information, and employs sub-band reconstruction technique to achieve high-quality compression across both low and high frequency bands.

Result: UniSRCodec achieves state-of-the-art performance among cross-domain single-codebook codecs with only 40 token rate, with reconstruction quality comparable to certain multi-codebook methods, as demonstrated by both subjective and objective experimental results.

Conclusion: UniSRCodec successfully addresses the limitations of existing single-codebook codecs, providing a simpler yet high-performance alternative that supports high sampling rates, low bandwidth, high fidelity, and unified audio modeling.

Abstract: Neural Audio Codecs (NACs) can reduce transmission overhead by performing compact compression and reconstruction, which also aim to bridge the gap between continuous and discrete signals. Existing NACs can be divided into two categories: multi-codebook and single-codebook codecs. Multi-codebook codecs face challenges such as structural complexity and difficulty in adapting to downstream tasks, while single-codebook codecs, though structurally simpler, suffer from low-fidelity, ineffective modeling of unified audio, and an inability to support modeling of high-frequency audio. We propose the UniSRCodec, a single-codebook codec capable of supporting high sampling rate, low-bandwidth, high fidelity, and unified. We analyze the inefficiency of waveform-based compression and introduce the time and frequency compression method using the Mel-spectrogram, and cooperate with a Vocoder to recover the phase information of the original audio. Moreover, we propose a sub-band reconstruction technique to achieve high-quality compression across both low and high frequency bands. Subjective and objective experimental results demonstrate that UniSRCodec achieves state-of-the-art (SOTA) performance among cross-domain single-codebook codecs with only a token rate of 40, and its reconstruction quality is comparable to that of certain multi-codebook methods. Our demo page is available at https://wxzyd123.github.io/unisrcodec.

Method: Proposes M2Former (multi-channel multi-speaker transformer) that can encode high-dimensional acoustic features for each speaker from mixed input audio, overcoming interference issues.

Result: On SMS-WSJ benchmark, M2Former achieves relative WER reductions of 9.2% over neural beamformer, 14.3% over MCT, 24.9% over dual-path RNN with TAC, and 52.2% over multi-channel deep clustering end-to-end systems.

Conclusion: M2Former effectively addresses the speaker interference problem in far-field multi-speaker ASR and significantly outperforms state-of-the-art baselines.

Abstract: With the development of teleconferencing and in-vehicle voice assistants, far-field multi-speaker speech recognition has become a hot research topic. Recently, a multi-channel transformer (MCT) has been proposed, which demonstrates the ability of the transformer to model far-field acoustic environments. However, MCT cannot encode high-dimensional acoustic features for each speaker from mixed input audio because of the interference between speakers. Based on these, we propose the multi-channel multi-speaker transformer (M2Former) for far-field multi-speaker ASR in this paper. Experiments on the SMS-WSJ benchmark show that the M2Former outperforms the neural beamformer, MCT, dual-path RNN with transform-average-concatenate and multi-channel deep clustering based end-to-end systems by 9.2%, 14.3%, 24.9%, and 52.2% respectively, in terms of relative word error rate reduction.

Method: Systematic empirical evaluation of state-of-the-art speaker authentication systems using a large-scale speech synthesis dataset to test security vulnerabilities.

Result: Two major vulnerabilities identified: 1) modern voice cloning models trained on small samples easily bypass commercial speaker verification systems; 2) anti-spoofing detectors fail to generalize across different audio synthesis methods, showing significant gap between in-domain performance and real-world robustness.

Conclusion: Current security measures need reconsideration, requiring architectural innovations, adaptive defenses, and transition towards multi-factor authentication to address the serious threats posed by audio deepfakes.

Abstract: As audio deepfakes transition from research artifacts to widely available commercial tools, robust biometric authentication faces pressing security threats in high-stakes industries. This paper presents a systematic empirical evaluation of state-of-the-art speaker authentication systems based on a large-scale speech synthesis dataset, revealing two major security vulnerabilities: 1) modern voice cloning models trained on very small samples can easily bypass commercial speaker verification systems; and 2) anti-spoofing detectors struggle to generalize across different methods of audio synthesis, leading to a significant gap between in-domain performance and real-world robustness. These findings call for a reconsideration of security measures and stress the need for architectural innovations, adaptive defenses, and the transition towards multi-factor authentication.

Method: Three core contributions: 1) Large-scale synthesized binaural audio dataset for rich spatial cues; 2) Hybrid feature projector with parallel semantic and spatial encoders for decoupled representations, integrated via dense fusion; 3) Progressive training curriculum from supervised fine-tuning to reinforcement learning via Group Relative Policy Optimization (GRPO).

Result: The model demonstrates strong capability for spatial understanding on their comprehensive benchmark, enabling spatial perception and advancing from mono semantic recognition to spatial intelligence.

Conclusion: The work provides a clear pathway for leveraging large models’ reasoning abilities toward holistic acoustic scene analysis by enabling spatial perception, bridging the gap between mono semantic recognition and spatial intelligence.

Abstract: Existing large audio-language models perceive the world as “mono” – a single stream of audio that ignores the critical spatial dimension (“where”) required for universal acoustic scene analysis. To bridge this gap, we first introduce a hierarchical framework for Auditory Scene Analysis (ASA). Guided by this framework, we introduce a system that enables models like Qwen2-Audio to understand and reason about the complex acoustic world. Our framework achieves this through three core contributions: First, we build a large-scale, synthesized binaural audio dataset to provide the rich spatial cues. Second, we design a hybrid feature projector, which leverages parallel semantic and spatial encoders to extract decoupled representations. These distinct streams are integrated via a dense fusion mechanism, ensuring the model receives a holistic view of the acoustic scene. Finally, we employ a progressive training curriculum, advancing from supervised fine-tuning (SFT) to reinforcement learning via Group Relative Policy Optimization (GRPO), to explicitly evolve the model’s capabilities towards reasoning. On our comprehensive benchmark, the model demonstrates comparatively strong capability for spatial understanding. By enabling this spatial perception, our work provides a clear pathway for leveraging the powerful reasoning abilities of large models towards holistic acoustic scene analysis, advancing from “mono” semantic recognition to spatial intelligence.

Method: Proposes an automatic annotation pipeline to construct Frequency-Time structured chain-of-thought rationales (~340K demonstrations). Uses two-stage training: 1) Supervised fine-tuning cold-start, then 2) Frequency Time-Group Relative Policy Optimization (FT-GRPO) with rule-based frequency-time constraints to prevent reward hacking and hallucinated rationales.

Result: FT-GRPO achieves state-of-the-art performance on all-type audio deepfake detection while producing interpretable, frequency-time grounded rationales that explain detection decisions.

Conclusion: The proposed FT-GRPO framework successfully addresses the limitations of traditional supervised fine-tuning (black-box decisions) and vanilla reinforcement fine-tuning (reward hacking, hallucinated rationales) by combining structured chain-of-thought rationales with constrained optimization for interpretable and effective audio deepfake detection.

Abstract: Recent advances in audio large language models (ALLMs) have made high-quality synthetic audio widely accessible, increasing the risk of malicious audio deepfakes across speech, environmental sounds, singing voice, and music. Real-world audio deepfake detection (ADD) therefore requires all-type detectors that generalize across heterogeneous audio and provide interpretable decisions. Given the strong multi-task generalization ability of ALLMs, we first investigate their performance on all-type ADD under both supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). However, SFT using only binary real/fake labels tends to reduce the model to a black-box classifier, sacrificing interpretability. Meanwhile, vanilla RFT under sparse supervision is prone to reward hacking and can produce hallucinated, ungrounded rationales. To address this, we propose an automatic annotation and polishing pipeline that constructs Frequency-Time structured chain-of-thought (CoT) rationales, producing ~340K cold-start demonstrations. Building on CoT data, we propose Frequency Time-Group Relative Policy Optimization (FT-GRPO), a two-stage training paradigm that cold-starts ALLMs with SFT and then applies GRPO under rule-based frequency-time constraints. Experiments demonstrate that FT-GRPO achieves state-of-the-art performance on all-type ADD while producing interpretable, FT-grounded rationales. The data and code are available online.

Method: Created CMDAR benchmark with 3,000 carefully curated question-answer pairs linked to diverse audio clips, covering five categories of complex reasoning and three question types (multiple-choice with single audio, multiple-choice with multiple audios, open-ended).

Result: Benchmarked 26 state-of-the-art audio language models: Qwen2.5-Omni achieved 76.67% accuracy on CMDAR-main, GPT-4o Audio reached 68.47%, but GPT-4o Audio substantially outperformed Qwen2.5-Omni on more challenging multiple-choice with multiple audios and open-ended tasks.

Conclusion: Current audio language models exhibit limitations in complex reasoning tasks, highlighting the need for improved audio reasoning capabilities. The benchmark provides detailed analysis and suggestions for future development of large audio language models.

Abstract: The ability to reason from audio, including speech, environmental sounds, and music, is essential for AI agents to interact effectively in real-world scenarios. Existing benchmarks mainly focus on static or single-scene settings and English audio data and do not fully capture scenarios where multiple speakers, unfolding events, and heterogeneous audio sources interact. To address these challenges, we introduce CMDAR, a Chinese benchmark for evaluating models on complex, multi-scene, and dynamically evolving audio reasoning tasks. CMDAR comprises 3,000 carefully curated question-answer pairs linked to diverse audio clips, covering five categories of complex reasoning and spanning three question types. We benchmark 26 state-of-the-art audio language models on CMDAR and observe that they exhibit limitations in complex reasoning tasks. In CMDAR-main, Qwen2.5-Omni achieves 76.67% accuracy, whereas GPT-4o Audio reaches 68.47%. However, GPT-4o Audio substantially outperforms Qwen2.5-Omni on the more challenging multiple-choice with multiple audios and open-ended tasks. And we provide detail analysis corresponding suggestions for the future development of large audio language models.

Method: Proposes segment-aware emotion conditioning (causal masking + monotonic stream alignment filtering) and segment-aware duration steering (local duration embedding steering + global EOS logit modulation). Uses LLM-based automatic prompt construction with a 30,000-sample annotated dataset.

Result: Achieves state-of-the-art intra-utterance consistency in multi-emotion and duration control while maintaining baseline-level speech quality of the underlying TTS model.

Conclusion: The training-free framework enables fine-grained intra-utterance emotion and duration control for pretrained zero-shot TTS without complex training procedures.

Abstract: While controllable Text-to-Speech (TTS) has achieved notable progress, most existing methods remain limited to inter-utterance-level control, making fine-grained intra-utterance expression challenging due to their reliance on non-public datasets or complex multi-stage training. In this paper, we propose a training-free controllable framework for pretrained zero-shot TTS to enable intra-utterance emotion and duration expression. Specifically, we propose a segment-aware emotion conditioning strategy that combines causal masking with monotonic stream alignment filtering to isolate emotion conditioning and schedule mask transitions, enabling smooth intra-utterance emotion shifts while preserving global semantic coherence. Based on this, we further propose a segment-aware duration steering strategy to combine local duration embedding steering with global EOS logit modulation, allowing local duration adjustment while ensuring globally consistent termination. To eliminate the need for segment-level manual prompt engineering, we construct a 30,000-sample multi-emotion and duration-annotated text dataset to enable LLM-based automatic prompt construction. Extensive experiments demonstrate that our training-free method not only achieves state-of-the-art intra-utterance consistency in multi-emotion and duration control, but also maintains baseline-level speech quality of the underlying TTS model. Audio samples are available at https://aclanonymous111.github.io/TED-TTS-DemoPage/.

Method: Developed MOSS Transcribe Diarize, a unified multimodal large language model trained on extensive real wild data with 128k context window for up to 90-minute inputs, performing joint SATS in end-to-end paradigm.

Result: Outperforms state-of-the-art commercial systems on multiple public and in-house benchmarks, demonstrating strong scaling and robust generalization.

Conclusion: The proposed end-to-end multimodal LLM approach effectively addresses limitations of existing SATS systems and achieves superior performance for speaker-attributed, time-stamped transcription tasks.

Abstract: Speaker-Attributed, Time-Stamped Transcription (SATS) aims to transcribe what is said and to precisely determine the timing of each speaker, which is particularly valuable for meeting transcription. Existing SATS systems rarely adopt an end-to-end formulation and are further constrained by limited context windows, weak long-range speaker memory, and the inability to output timestamps. To address these limitations, we present MOSS Transcribe Diarize, a unified multimodal large language model that jointly performs Speaker-Attributed, Time-Stamped Transcription in an end-to-end paradigm. Trained on extensive real wild data and equipped with a 128k context window for up to 90-minute inputs, MOSS Transcribe Diarize scales well and generalizes robustly. Across comprehensive evaluations, it outperforms state-of-the-art commercial systems on multiple public and in-house benchmarks.

Method: Introduced AGL1K benchmark spanning 72 countries/territories, proposed Audio Localizability metric to quantify informativeness of recordings, curated 1,444 audio clips from crowd-sourced platforms, and evaluated 16 audio language models.

Result: ALMs demonstrate emerging audio geo-localization capability; closed-source models substantially outperform open-source models; linguistic clues often dominate as prediction scaffolds; analysis includes reasoning traces, regional bias, error causes, and metric interpretability.

Conclusion: AGL1K establishes a benchmark for audio geo-localization that may advance ALMs with better geospatial reasoning capability, addressing the previous gap in audio-location paired datasets.

Abstract: Geo-localization aims to infer the geographic origin of a given signal. In computer vision, geo-localization has served as a demanding benchmark for compositional reasoning and is relevant to public safety. In contrast, progress on audio geo-localization has been constrained by the lack of high-quality audio-location pairs. To address this gap, we introduce AGL1K, the first audio geo-localization benchmark for audio language models (ALMs), spanning 72 countries and territories. To extract reliably localizable samples from a crowd-sourced platform, we propose the Audio Localizability metric that quantifies the informativeness of each recording, yielding 1,444 curated audio clips. Evaluations on 16 ALMs show that ALMs have emerged with audio geo-localization capability. We find that closed-source models substantially outperform open-source models, and that linguistic clues often dominate as a scaffold for prediction. We further analyze ALMs’ reasoning traces, regional bias, error causes, and the interpretability of the localizability metric. Overall, AGL1K establishes a benchmark for audio geo-localization and may advance ALMs with better geospatial reasoning capability.

Method: Leverage foundation models (large-scale neural architectures pretrained on multimodal data) and apply various probing methods to their embeddings to examine relationships between audio FX and estimated emotion.

Result: Uncovered complex, nonlinear relationships between audio FX and emotion, revealing patterns tied to specific effects and evaluating the robustness of foundation audio models.

Conclusion: The findings advance understanding of how audio production practices affect perception, with implications for music cognition, performance, and affective computing.

Abstract: Audio effects (FX) such as reverberation, distortion, modulation, and dynamic range processing play a pivotal role in shaping emotional responses during music listening. While prior studies have examined links between low-level audio features and affective perception, the systematic impact of audio FX on emotion remains underexplored. This work investigates how foundation models - large-scale neural architectures pretrained on multimodal data - can be leveraged to analyze these effects. Such models encode rich associations between musical structure, timbre, and affective meaning, offering a powerful framework for probing the emotional consequences of sound design techniques. By applying various probing methods to embeddings from deep learning models, we examine the complex, nonlinear relationships between audio FX and estimated emotion, uncovering patterns tied to specific effects and evaluating the robustness of foundation audio models. Our findings aim to advance understanding of the perceptual impact of audio production practices, with implications for music cognition, performance, and affective computing.

Method: Created MSU-Bench with 1,800 human-curated generative QA pairs from classical composers’ works, organized into four difficulty levels (onset to form), and evaluated 15+ state-of-the-art models in zero-shot and fine-tuned settings across textual and visual modalities.

Result: Revealed pronounced modality gaps, unstable level-wise performance, and challenges in maintaining multilevel correctness. Fine-tuning substantially improved results across modalities while preserving general knowledge.

Conclusion: MSU-Bench serves as a robust foundation for future multimodal reasoning research, with all resources publicly released to facilitate further investigation into musical score understanding.

Abstract: Understanding complete musical scores entails integrated reasoning over pitch, rhythm, harmony, and large-scale structure, yet the ability of Large Language Models and Vision-Language Models to interpret full musical notation remains insufficiently examined. We introduce the Musical Score Understanding Benchmark (MSU-Bench), the first large-scale, human-curated benchmark for score-level musical understanding across textual (ABC notation) and visual (PDF) modalities. MSU-Bench contains 1,800 generative Question-Answering pairs from works by Bach, Beethoven, Chopin, Debussy, and others, organised into four levels of increasing difficulty, ranging from onset information to texture and form. Evaluations of more than fifteen state-of-the-art models, in both zero-shot and fine-tuned settings, reveal pronounced modality gaps, unstable level-wise performance, and challenges in maintaining multilevel correctness. Fine-tuning substantially improves results across modalities while preserving general knowledge, positioning MSU-Bench as a robust foundation for future research in multimodal reasoning. To facilitate further research, we publicly release MSU-Bench and all associated resources.

Method: Three-level architecture: micro level models attention heads as interacting spins with Hamiltonians; meso level uses Neural Differential Manifold (NDM) with Hamiltonian flows and HJB optimal control; macro level maintains generative semantic workspace and information-phase portrait. Uses symplectic layers to preserve geometric invariants.

Result: On simple toy problems (numerical integration and dynamical systems), the physical transformer outperforms naive baselines in stability and long-horizon accuracy by respecting underlying geometric and Hamiltonian structure.

Conclusion: The framework provides a path toward physical AI that unifies digital reasoning with physically grounded manifolds, potentially leading to more interpretable and unified models of reasoning, control, and real-world interaction.

Abstract: Digital AI systems spanning large language models, vision models, and generative architectures that operate primarily in symbolic, linguistic, or pixel domains. They have achieved striking progress, but almost all of this progress lives in virtual spaces. These systems transform embeddings and tokens, yet do not themselves touch the world and rarely admit a physical interpretation. In this work we propose a physical transformer that couples modern transformer style computation with geometric representation and physical dynamics. At the micro level, attention heads, and feed-forward blocks are modeled as interacting spins governed by effective Hamiltonians plus non-Hamiltonian bath terms. At the meso level, their aggregated state evolves on a learned Neural Differential Manifold (NDM) under Hamiltonian flows and Hamilton, Jacobi, Bellman (HJB) optimal control, discretized by symplectic layers that approximately preserve geometric and energetic invariants. At the macro level, the model maintains a generative semantic workspace and a two-dimensional information-phase portrait that tracks uncertainty and information gain over a reasoning trajectory. Within this hierarchy, reasoning tasks are formulated as controlled information flows on the manifold, with solutions corresponding to low cost trajectories that satisfy geometric, energetic, and workspace-consistency constraints. On simple toy problems involving numerical integration and dynamical systems, the physical transformer outperforms naive baselines in stability and long-horizon accuracy, highlighting the benefits of respecting underlying geometric and Hamiltonian structure. More broadly, the framework suggests a path toward physical AI that unify digital reasoning with physically grounded manifolds, opening a route to more interpretable and potentially unified models of reasoning, control, and interaction with the real world.

Method: 1) Created WebGym with 300K tasks across diverse real-world websites with rubric-based evaluations. 2) Developed high-throughput asynchronous rollout system for 4-5x speedup in trajectory sampling. 3) Used RL training on agent’s own interaction traces with task rewards as feedback. 4) Scaled task set breadth, depth, and size. 5) Fine-tuned Qwen-3-VL-8B-Instruct vision-language model on WebGym.

Result: Fine-tuned agent achieved 42.9% success rate on out-of-distribution test set (websites never seen during training), significantly outperforming GPT-4o (27.1%) and GPT-5-Thinking (29.8%). The baseline success rate was 26.2% before fine-tuning.

Conclusion: WebGym enables effective training of visual web agents that generalize to unseen websites through large-scale RL training and model fine-tuning, demonstrating substantial improvements over proprietary models on realistic web navigation tasks.

Abstract: We present WebGym, the largest-to-date open-source environment for training realistic visual web agents. Real websites are non-stationary and diverse, making artificial or small-scale task sets insufficient for robust policy learning. WebGym contains nearly 300,000 tasks with rubric-based evaluations across diverse, real-world websites and difficulty levels. We train agents with a simple reinforcement learning (RL) recipe, which trains on the agent’s own interaction traces (rollouts), using task rewards as feedback to guide learning. To enable scaling RL, we speed up sampling of trajectories in WebGym by developing a high-throughput asynchronous rollout system, designed specifically for web agents. Our system achieves a 4-5x rollout speedup compared to naive implementations. Second, we scale the task set breadth, depth, and size, which results in continued performance improvement. Fine-tuning a strong base vision-language model, Qwen-3-VL-8B-Instruct, on WebGym results in an improvement in success rate on an out-of-distribution test set from 26.2% to 42.9%, significantly outperforming agents based on proprietary models such as GPT-4o and GPT-5-Thinking that achieve 27.1% and 29.8%, respectively. This improvement is substantial because our test set consists only of tasks on websites never seen during training, unlike many other prior works on training visual web agents.

Method: Adapts Manifold-Constrained Hyper-Connections to GNNs by expanding node representations across parallel streams and constraining stream-mixing matrices to the Birkhoff polytope via Sinkhorn-Knopp normalization.

Result: mHC-GNN shows exponentially slower over-smoothing (rate (1-γ)^{L/n} vs. (1-γ)^L), distinguishes graphs beyond 1-WL, maintains over 74% accuracy at 128 layers while standard GNNs collapse beyond 16 layers, with improvements exceeding 50 percentage points at extreme depths.

Conclusion: mHC-GNN effectively addresses over-smoothing and expressiveness limitations in deep GNNs, with the manifold constraint being essential for performance (removing it causes up to 82% degradation).

Abstract: Graph Neural Networks (GNNs) suffer from over-smoothing in deep architectures and expressiveness bounded by the 1-Weisfeiler-Leman (1-WL) test. We adapt Manifold-Constrained Hyper-Connections (\mhc)~\citep{xie2025mhc}, recently proposed for Transformers, to graph neural networks. Our method, mHC-GNN, expands node representations across $n$ parallel streams and constrains stream-mixing matrices to the Birkhoff polytope via Sinkhorn-Knopp normalization. We prove that mHC-GNN exhibits exponentially slower over-smoothing (rate $(1-γ)^{L/n}$ vs.\ $(1-γ)^L$) and can distinguish graphs beyond 1-WL. Experiments on 10 datasets with 4 GNN architectures show consistent improvements. Depth experiments from 2 to 128 layers reveal that standard GNNs collapse to near-random performance beyond 16 layers, while mHC-GNN maintains over 74% accuracy even at 128 layers, with improvements exceeding 50 percentage points at extreme depths. Ablations confirm that the manifold constraint is essential: removing it causes up to 82% performance degradation. Code is available at \href{https://github.com/smlab-niser/mhc-gnn}{https://github.com/smlab-niser/mhc-gnn}

Method: MixTTE integrates link-level modeling with industrial route-level TTE systems using: 1) spatio-temporal external attention module for global traffic dynamic dependencies across million-scale networks, 2) stabilized graph mixture-of-experts network for heterogeneous traffic patterns with inference efficiency, and 3) asynchronous incremental learning for real-time adaptation to traffic distribution shifts.

Result: Experiments on real-world datasets show MixTTE significantly reduces prediction errors compared to seven baselines. The system has been deployed in DiDi, substantially improving accuracy and stability of the TTE service.

Conclusion: MixTTE provides a scalable and adaptive framework that effectively addresses limitations of existing industrial TTE systems by capturing city-scale traffic dynamics while maintaining efficiency, leading to improved prediction accuracy and service stability in real-world deployment.

Abstract: Accurate Travel Time Estimation (TTE) is critical for ride-hailing platforms, where errors directly impact user experience and operational efficiency. While existing production systems excel at holistic route-level dependency modeling, they struggle to capture city-scale traffic dynamics and long-tail scenarios, leading to unreliable predictions in large urban networks. In this paper, we propose \model, a scalable and adaptive framework that synergistically integrates link-level modeling with industrial route-level TTE systems. Specifically, we propose a spatio-temporal external attention module to capture global traffic dynamic dependencies across million-scale road networks efficiently. Moreover, we construct a stabilized graph mixture-of-experts network to handle heterogeneous traffic patterns while maintaining inference efficiency. Furthermore, an asynchronous incremental learning strategy is tailored to enable real-time and stable adaptation to dynamic traffic distribution shifts. Experiments on real-world datasets validate MixTTE significantly reduces prediction errors compared to seven baselines. MixTTE has been deployed in DiDi, substantially improving the accuracy and stability of the TTE service.

Method: The authors use Li-Yau estimates for heat kernel log-gradients and Minakshisundaram-Pleijel parametrix expansion of perturbed heat equations to analyze Riemannian diffusion models under mild curvature assumptions.

Result: The paper establishes that polynomially small step sizes (instead of exponentially small) suffice to guarantee small sampling error in total variation distance, requiring only L2-accurate score estimates and standard curvature assumptions.

Conclusion: This work provides stronger theoretical foundations for diffusion models on non-Euclidean spaces, enabling sharper analysis and more practical implementations of diffusion models on manifolds.

Abstract: Diffusion models have demonstrated remarkable empirical success in the recent years and are considered one of the state-of-the-art generative models in modern AI. These models consist of a forward process, which gradually diffuses the data distribution to a noise distribution spanning the whole space, and a backward process, which inverts this transformation to recover the data distribution from noise. Most of the existing literature assumes that the underlying space is Euclidean. However, in many practical applications, the data are constrained to lie on a submanifold of Euclidean space. Addressing this setting, De Bortoli et al. (2022) introduced Riemannian diffusion models and proved that using an exponentially small step size yields a small sampling error in the Wasserstein distance, provided the data distribution is smooth and strictly positive, and the score estimate is $L_\infty$-accurate. In this paper, we greatly strengthen this theory by establishing that, under $L_2$-accurate score estimate, a {\em polynomially small stepsize} suffices to guarantee small sampling error in the total variation distance, without requiring smoothness or positivity of the data distribution. Our analysis only requires mild and standard curvature assumptions on the underlying manifold. The main ingredients in our analysis are Li-Yau estimate for the log-gradient of heat kernel, and Minakshisundaram-Pleijel parametrix expansion of the perturbed heat equation. Our approach opens the door to a sharper analysis of diffusion models on non-Euclidean spaces.

Method: I-GEM revisits Gradient Episodic Memory (GEM) within the Low-Rank Adapter (LoRA) subspace, using a fixed-budget, GPU-resident dual projected-gradient approximation to GEM’s quadratic projection, constraining non-interference solely within adapter parameters.

Result: On 3-task AG News split with domain drift using GPT-2 (355M) and LoRA (r=8), I-GEM matches GEM’s average accuracy (within ~0.04 points) and outperforms A-GEM by ~1.4 points, while reducing projection time by factor of ~1000 compared to GEM.

Conclusion: Applying GEM constraints in the LoRA subspace provides a practical pathway for continual learning at LLM scale, offering GEM-like stability with orders-of-magnitude lower computational overhead.

Abstract: Full fine-tuning of Large Language Models (LLMs) is computationally costly, motivating Continual Learning (CL) approaches that utilize parameter-efficient adapters. We revisit Gradient Episodic Memory (GEM) within the Low-Rank Adapter (LoRA) subspace and introduce I-GEM: a fixed-budget, GPU-resident dual projected-gradient approximation to GEM’s quadratic projection. By constraining non-interference solely within the adapter parameters, I-GEM preserves GEM-like stability with orders-of-magnitude lower mean projection overhead. On a 3-task AG News split with induced domain drift, using GPT-2 (355M) and LoRA ($r=8$), I-GEM matches GEM’s average accuracy (within $\sim!0.04$ pts) and outperforms A-GEM by $\sim!1.4$ pts. Crucially, it reduces projection time vs.\ GEM by a factor of $\sim!10^3$. These results suggest that applying GEM constraints in the LoRA subspace is a practical pathway for continual learning at the LLM scale.

Method: Major extension of hdlib Python library with four key extensions: 1) enhanced supervised classification with feature selection, 2) new regression model for continuous variables, 3) clustering model for unsupervised learning, 4) graph-based learning model. Plus pioneering implementation of Quantum Hyperdimensional Computing with quantum-powered arithmetic operations and new Quantum Machine Learning model for supervised learning.

Result: Significantly enhanced machine learning capabilities within the VSA framework, making hdlib a more comprehensive toolkit for both classical and quantum hyperdimensional computing applications. The library remains open-source under MIT license with improved documentation and examples.

Conclusion: hdlib v2 represents a substantial advancement in VSA/Hyperdimensional Computing tooling, bridging the gap between symbolic and statistical machine learning while pioneering quantum implementations, making advanced VSA techniques more accessible to researchers and practitioners.

Abstract: Following the initial publication of hdlib, a Python library for designing Vector-Symbolic Architectures (VSA), we introduce a major extension that significantly enhances its machine learning capabilities. VSA, also known as Hyperdimensional Computing, is a computing paradigm that represents and processes information using high-dimensional vectors. While the first version of hdlib established a robust foundation for creating and manipulating these vectors, this update addresses the growing need for more advanced, data-driven modeling within the VSA framework. Here, we present four extensions: significant enhancements to the existing supervised classification model also enabling feature selection, and a new regression model for predicting continuous variables, a clustering model for unsupervised learning, and a graph-based learning model. Furthermore, we propose the first implementation ever of Quantum Hyperdimensional Computing with quantum-powered arithmetic operations and a new Quantum Machine Learning model for supervised learning. hdlib remains open-source and available on GitHub at https://github.com/cumbof/hdlib under the MIT license, and distributed through the Python Package Index (pip install hdlib) and Conda (conda install -c conda-forge hdlib). Documentation and examples of these new features are available on the official Wiki at https://github.com/cumbof/hdlib/wiki.

Method: Integrates LLM-based potential functions for reward shaping with RL, VAE-enhanced dynamic reward scaling, and active learning with label propagation. Uses LSTM-based RL agent with LLM-derived semantic rewards and VAE reconstruction errors for unsupervised anomaly signals.

Result: Achieves state-of-the-art detection accuracy on Yahoo-A1 and SMD benchmarks under limited labeling budgets and operates effectively in data-constrained settings.

Conclusion: Combining LLMs with RL and advanced unsupervised techniques shows promise for robust, scalable anomaly detection in real-world applications.

Abstract: Detecting anomalies in time series data is crucial for finance, healthcare, sensor networks, and industrial monitoring applications. However, time series anomaly detection often suffers from sparse labels, complex temporal patterns, and costly expert annotation. We propose a unified framework that integrates Large Language Model (LLM)-based potential functions for reward shaping with Reinforcement Learning (RL), Variational Autoencoder (VAE)-enhanced dynamic reward scaling, and active learning with label propagation. An LSTM-based RL agent leverages LLM-derived semantic rewards to guide exploration, while VAE reconstruction errors add unsupervised anomaly signals. Active learning selects the most uncertain samples, and label propagation efficiently expands labeled data. Evaluations on Yahoo-A1 and SMD benchmarks demonstrate that our method achieves state-of-the-art detection accuracy under limited labeling budgets and operates effectively in data-constrained settings. This study highlights the promise of combining LLMs with RL and advanced unsupervised techniques for robust, scalable anomaly detection in real-world applications.

Method: CamS serializes molecular graphs into structure-rich causal sequences by: 1) mining data-driven connection-aware motifs, 2) serializing motifs via scaffold-rooted BFS to establish core-to-periphery order, and 3) enabling hierarchical modeling by concatenating sequences from fine to coarse motif scales. The method is instantiated as CamS-LLaMA by pre-training a vanilla LLaMA backbone on CamS sequences.

Result: CamS-LLaMA achieves state-of-the-art performance on MoleculeNet and the activity-cliff benchmark MoleculeACE, outperforming both SMILES-based language models and strong graph baselines. Interpretability analysis confirms effective attention toward cliff-determining differences.

Conclusion: CamS successfully bridges the gap between SMILES-based and graph-native approaches by enabling decoder-only Transformers to learn molecular graphs via standard next-token prediction, while preserving both topological connectivity and crucial chemical details through multi-scale causal serialization.

Abstract: We present Connection-Aware Motif Sequencing (CamS), a graph-to-sequence representation that enables decoder-only Transformers to learn molecular graphs via standard next-token prediction (NTP). For molecular property prediction, SMILES-based NTP scales well but lacks explicit topology, whereas graph-native masked modeling captures connectivity but risks disrupting the pivotal chemical details (e.g., activity cliffs). CamS bridges this gap by serializing molecular graphs into structure-rich causal sequences. CamS first mines data-driven connection-aware motifs. It then serializes motifs via scaffold-rooted breadth-first search (BFS) to establish a stable core-to-periphery order. Crucially, CamS enables hierarchical modeling by concatenating sequences from fine to coarse motif scales, allowing the model to condition global scaffolds on dense, uncorrupted local structural evidence. We instantiate CamS-LLaMA by pre-training a vanilla LLaMA backbone on CamS sequences. It achieves state-of-the-art performance on MoleculeNet and the activity-cliff benchmark MoleculeACE, outperforming both SMILES-based language models and strong graph baselines. Interpretability analysis confirms that our multi-scale causal serialization effectively drives attention toward cliff-determining differences.

Method: Conformal Bayesian Dermatological Classifier (CBDC) framework that integrates Statistical Learning Theory, Topological Data Analysis (TDA), and Bayesian Conformal Inference. Provides distribution-dependent generalization bounds, topological stability theorem for CNN embeddings under perturbations, and finite conformal coverage guarantees.

Result: CBDC achieves classification accuracy while generating calibrated, clinically interpretable predictions on HAM10000, PH2, and ISIC 2020 datasets. Provides theoretical guarantees for generalization, stability, and uncertainty quantification.

Conclusion: CBDC represents a theoretical and practical leap for deep dermatological diagnostics, bridging the interface between machine learning theory and clinical applicability.

Abstract: The rapid growth of dermatological imaging and mobile diagnostic tools calls for systems that not only demonstrate empirical performance but also provide strong theoretical guarantees. Deep learning models have shown high predictive accuracy; however, they are often criticized for lacking well, calibrated uncertainty estimates without which these models are hardly deployable in a clinical setting. To this end, we present the Conformal Bayesian Dermatological Classifier (CBDC), a well, founded framework that combines Statistical Learning Theory, Topological Data Analysis (TDA), and Bayesian Conformal Inference. CBDC offers distribution, dependent generalization bounds that reflect dermatological variability, proves a topological stability theorem that guarantees the invariance of convolutional neural network embeddings under photometric and morphological perturbations and provides finite conformal coverage guarantees for trustworthy uncertainty quantification. Through exhaustive experiments on the HAM10000, PH2, and ISIC 2020 datasets, we show that CBDC not only attains classification accuracy but also generates calibrated predictions that are interpretable from a clinical perspective. This research constitutes a theoretical and practical leap for deep dermatological diagnostics, thereby opening the machine learning theory clinical applicability interface.

Method: The authors propose NCMI as an information-theoretic surrogate loss, observe that model’s NCMI is inversely proportional to its accuracy, and introduce an alternating algorithm to efficiently minimize NCMI during training.

Result: NCMI-trained models substantially outperform state-of-the-art losses: 2.77% top-1 accuracy improvement on ImageNet with ResNet-50, 8.6% macro-F1 improvement on CAMELYON-17, with consistent gains across various architectures and batch sizes.

Conclusion: NCMI is a practical and competitive alternative to cross-entropy that offers significant performance improvements at comparable computational cost, suggesting it could become a new standard for training DNN classifiers.

Abstract: In this paper, we propose a novel information theoretic surrogate loss; normalized conditional mutual information (NCMI); as a drop in alternative to the de facto cross-entropy (CE) for training deep neural network (DNN) based classifiers. We first observe that the model’s NCMI is inversely proportional to its accuracy. Building on this insight, we introduce an alternating algorithm to efficiently minimize the NCMI. Across image recognition and whole-slide imaging (WSI) subtyping benchmarks, NCMI-trained models surpass state of the art losses by substantial margins at a computational cost comparable to that of CE. Notably, on ImageNet, NCMI yields a 2.77% top-1 accuracy improvement with ResNet-50 comparing to the CE; on CAMELYON-17, replacing CE with NCMI improves the macro-F1 by 8.6% over the strongest baseline. Gains are consistent across various architectures and batch sizes, suggesting that NCMI is a practical and competitive alternative to CE.

Method: Generalizes TURTLE’s cost function with power law prior for imbalanced distributions and introduces sparse logits in labeling to simplify search space.

Result: Improves accuracy for imbalanced data, prevents over-prediction of minority clusters, and enhances overall clustering on synthetic and real datasets.

Conclusion: PET-TURTLE successfully addresses TURTLE’s limitations with imbalanced data while maintaining/improving performance on balanced datasets through power law prior and sparse logits.

Abstract: Foundation vision, audio, and language models enable zero-shot performance on downstream tasks via their latent representations. Recently, unsupervised learning of data group structure with deep learning methods has gained popularity. TURTLE, a state of the art deep clustering algorithm, uncovers data labeling without supervision by alternating label and hyperplane updates, maximizing the hyperplane margin, in a similar fashion to support vector machines (SVMs). However, TURTLE assumes clusters are balanced; when data is imbalanced, it yields non-ideal hyperplanes that cause higher clustering error. We propose PET-TURTLE, which generalizes the cost function to handle imbalanced data distributions by a power law prior. Additionally, by introducing sparse logits in the labeling process, PET-TURTLE optimizes a simpler search space that in turn improves accuracy for balanced datasets. Experiments on synthetic and real data show that PET-TURTLE improves accuracy for imbalanced sources, prevents over-prediction of minority clusters, and enhances overall clustering.

Method: LendNova uses advanced NLP techniques and language models to operate directly on raw, jargon-heavy credit bureau text, learning task-relevant representations automatically without manual feature engineering or preprocessing.

Result: Evaluation on real-world data demonstrates strong potential for accurate and efficient risk assessment, establishing a baseline for intelligent credit risk agents and showing the feasibility of language models in this domain.

Conclusion: LendNova transforms risk modeling by capturing patterns and risk signals embedded in text, reducing costs and improving scalability, while laying groundwork for future foundation systems enabling more accurate, adaptable, and automated financial decision-making.

Abstract: Credit risk assessment is essential in the financial sector, but has traditionally depended on costly feature-based models that often fail to utilize all available information in raw credit records. This paper introduces LendNova, the first practical automated end-to-end pipeline for credit risk assessment, designed to utilize all available information in raw credit records by leveraging advanced NLP techniques and language models. LendNova transforms risk modeling by operating directly on raw, jargon-heavy credit bureau text using a language model that learns task-relevant representations without manual feature engineering. By automatically capturing patterns and risk signals embedded in the text, it replaces manual preprocessing steps, reducing costs and improving scalability. Evaluation on real-world data further demonstrates its strong potential in accurate and efficient risk assessment. LendNova establishes a baseline for intelligent credit risk agents, demonstrating the feasibility of language models in this domain. It lays the groundwork for future research toward foundation systems that enable more accurate, adaptable, and automated financial decision-making.

Method: Proposes a threat detection framework using machine learning, specifically artificial neural networks (ANN). The approach involves extensive preprocessing and exploratory data analysis to address data imbalance, feature inconsistency, and noise in network traffic data.

Result: The ANN model demonstrates good detection performance with high robustness, achieving strong results on conventional metrics including accuracy, recall, F1-score, and ROC-AUC.

Conclusion: Neural network-based solutions show potential for effectively identifying latent threat dynamics in large-scale social media networks and can complement existing intrusion detection systems for proactive cybersecurity operations.

Abstract: The accelerated development of social media websites has posed intricate security issues in cyberspace, where these sites have increasingly become victims of criminal activities including attempts to intrude into them, abnormal traffic patterns, and organized attacks. The conventional rule-based security systems are not always scalable and dynamic to meet such a threat. This paper introduces a threat detection framework based on machine learning that can be used to classify malicious behavior in the social media network environment based on the nature of network traffic. Exploiting a rich network traffic dataset, the massive preprocessing and exploratory data analysis is conducted to overcome the problem of data imbalance, feature inconsistency, and noise. A model of artificial neural network (ANN) is then created to acquire intricate, non-linear tendencies of malicious actions. The proposed model is tested on conventional performance metrics, such as accuracy, accuracy, recall, F1-score, and ROC-AUC, and shows good detection and high levels of strength. The findings suggest that neural network-based solutions have the potential to be used effectively to identify the latent threat dynamics within the context of a large-scale social media network and that they can be employed to complement the existing intrusion detection system and better to conduct proactive cybersecurity operations.

Method: Four synergistic optimizations: (1) Fused Triton kernels eliminating 75% memory traffic via RMSNorm (7x), SwiGLU (5x), and QK-RoPE (2.3x) fusion; (2) Cut Cross-Entropy reducing logit memory from 5GB to 135MB through online softmax computation; (3) LoRA+ with theoretically-derived 16x differential learning rates between adapter matrices; (4) Best-Fit Decreasing sequence packing recovering 60-75% of compute wasted on padding.

Result: On Qwen2.5-0.5B with A100-40GB: 41,184 tokens/second for full fine-tuning (3.51x faster than Unsloth’s 11,736 tokens/second). For LoRA at rank 32: 11,699 tokens/second (4.10x faster than Unsloth MAX’s 2,857 tokens/second). Discovered Unsloth’s reported 46,000 tokens/second benchmark had zero gradient norms (model wasn’t training).

Conclusion: Chronicals provides a comprehensive open-source solution for efficient LLM fine-tuning with mathematical foundations, achieving significant speedups through memory and compute optimizations, with all implementations, benchmarks, and proofs available on GitHub and PyPI.

Abstract: Large language model fine-tuning is bottlenecked by memory: a 7B parameter model requires 84GB–14GB for weights, 14GB for gradients, and 56GB for FP32 optimizer states–exceeding even A100-40GB capacity. We present Chronicals, an open-source training framework achieving 3.51x speedup over Unsloth through four synergistic optimizations: (1) fused Triton kernels eliminating 75% of memory traffic via RMSNorm (7x), SwiGLU (5x), and QK-RoPE (2.3x) fusion; (2) Cut Cross-Entropy reducing logit memory from 5GB to 135MB through online softmax computation; (3) LoRA+ with theoretically-derived 16x differential learning rates between adapter matrices; and (4) Best-Fit Decreasing sequence packing recovering 60-75% of compute wasted on padding. On Qwen2.5-0.5B with A100-40GB, Chronicals achieves 41,184 tokens/second for full fine-tuning versus Unsloth’s 11,736 tokens/second (3.51x). For LoRA at rank 32, we reach 11,699 tokens/second versus Unsloth MAX’s 2,857 tokens/second (4.10x). Critically, we discovered that Unsloth’s reported 46,000 tokens/second benchmark exhibited zero gradient norms–the model was not training. We provide complete mathematical foundations: online softmax correctness proofs, FlashAttention IO complexity bounds O(N^2 d^2 M^{-1}), LoRA+ learning rate derivations from gradient magnitude analysis, and bin-packing approximation guarantees. All implementations, benchmarks, and proofs are available at https://github.com/Ajwebdevs/Chronicals with pip installation via https://pypi.org/project/chronicals/.

Method: The authors propose Neural Manifold Noise Correlation (NMNC), which performs credit assignment using perturbations restricted to the neural manifold rather than isotropic noise. They show theoretically and empirically that the Jacobian row space aligns with the neural manifold in trained networks, and that manifold dimensionality scales slowly with network size.

Result: NMNC substantially improves performance and sample efficiency over vanilla noise correlation in convolutional networks trained on CIFAR-10, ImageNet-scale models, and recurrent networks. NMNC also yields representations more similar to the primate visual system than vanilla noise correlation.

Conclusion: These findings offer a mechanistic hypothesis for how biological circuits could support credit assignment, and suggest that biologically inspired constraints (like low-dimensional neural manifolds) may enable, rather than limit, effective learning at scale.

Abstract: Credit assignment–how changes in individual neurons and synapses affect a network’s output–is central to learning in brains and machines. Noise correlation, which estimates gradients by correlating perturbations of activity with changes in output, provides a biologically plausible solution to credit assignment but scales poorly as accurately estimating the Jacobian requires that the number of perturbations scale with network size. Moreover, isotropic noise conflicts with neurobiological observations that neural activity lies on a low-dimensional manifold. To address these drawbacks, we propose neural manifold noise correlation (NMNC), which performs credit assignment using perturbations restricted to the neural manifold. We show theoretically and empirically that the Jacobian row space aligns with the neural manifold in trained networks, and that manifold dimensionality scales slowly with network size. NMNC substantially improves performance and sample efficiency over vanilla noise correlation in convolutional networks trained on CIFAR-10, ImageNet-scale models, and recurrent networks. NMNC also yields representations more similar to the primate visual system than vanilla noise correlation. These findings offer a mechanistic hypothesis for how biological circuits could support credit assignment, and suggest that biologically inspired constraints may enable, rather than limit, effective learning at scale.

Method: The framework uses a model-driven priority score derived from empirical success statistics to select problems. It prioritizes problems with intermediate success rates (neither consistently solved nor consistently failed) as these provide stronger learning signals. The system includes heap-based prioritized sampling and periodic retesting mechanisms to prevent starvation and forgetting of solved/unsolved problems.

Result: The method creates a continuously adapting, automatic prioritization process that requires no predefined difficulty tiers, auxiliary predictors, or external labels. It offers a scalable alternative to manually designed curricula while aligning data selection directly with GRPO-based post-training dynamics.

Conclusion: The proposed framework provides a principled and scalable approach to problem prioritization for RL post-training of LLMs, automatically focusing training on problems that offer the strongest learning signals based on empirical success statistics, without requiring manual curriculum design.

Abstract: We introduce a problem-level prioritization framework for RL post-training of large language models. Building on insights from prioritized replay in deep RL, as well as prior observations that rollouts with intermediate success rates tend to produce stronger learning signals under methods such as GRPO, our approach selects problems according to a simple, model-driven priority score derived from empirical success statistics. In contrast to conventional curriculum strategies that emphasize easier tasks early in training, the resulting schedule naturally focuses training on problems that are neither consistently solved nor consistently failed, while deprioritizing those that contribute little gradient information. The method yields a continuously adapting and automatic prioritization process that requires no predefined difficulty tiers, auxiliary predictors, or external labels. We further introduce lightweight mechanisms for practical deployment, including heap-based prioritized sampling and periodic retesting of solved and unsolved problems to mitigate starvation and forgetting. Overall, the approach offers a principled and scalable alternative to manually designed curricula while aligning data selection directly with the dynamics of GRPO-based post-training.

Method: SpikingGPF uses spiking neuron architecture to learn sparse prompt vectors for selective feature prompting, and employs sparse representation theory to represent prompts as sparse combinations of prompt atoms.

Result: Extensive experiments on benchmarks show SpikingGPF is effective and robust, providing more compact, lightweight prompting with improved noise resistance.

Conclusion: SpikingGPF successfully addresses redundancy and noise sensitivity in graph prompting through sparse prompting with spiking neurons, offering efficient and robust adaptation of pre-trained GNNs.

Abstract: Graph Prompt Feature (GPF) learning has been widely used in adapting pre-trained GNN model on the downstream task. GPFs first introduce some prompt atoms and then learns the optimal prompt vector for each graph node using the linear combination of prompt atoms. However, existing GPFs generally conduct prompting over node’s all feature dimensions which is obviously redundant and also be sensitive to node feature noise. To overcome this issue, for the first time, this paper proposes learning sparse graph prompts by leveraging the spiking neuron mechanism, termed Spiking Graph Prompt Feature (SpikingGPF). Our approach is motivated by the observation that spiking neuron can perform inexpensive information processing and produce sparse outputs which naturally fits the task of our graph sparse prompting. Specifically, SpikingGPF has two main aspects. First, it learns a sparse prompt vector for each node by exploiting a spiking neuron architecture, enabling prompting on selective node features. This yields a more compact and lightweight prompting design while also improving robustness against node noise. Second, SpikingGPF introduces a novel prompt representation learning model based on sparse representation theory, i.e., it represents each node prompt as a sparse combination of prompt atoms. This encourages a more compact representation and also facilitates efficient computation. Extensive experiments on several benchmarks demonstrate the effectiveness and robustness of SpikingGPF.

Method: MAFS integrates statistical priors with deep learning: 1) Starts with filter-based priors for stable initialization, 2) Uses multi-head attention to examine features from multiple perspectives in parallel, capturing complex nonlinear relationships, 3) Includes a reordering module to consolidate outputs across attention heads, resolving conflicts and minimizing information loss.

Result: MAFS consistently achieves superior coverage and stability compared to existing filter-based and deep learning-based alternatives across simulated and real-world datasets (cancer gene expression and Alzheimer’s disease data).

Conclusion: MAFS offers a scalable, interpretable, and robust solution for feature selection in high-dimensional biomedical data by combining statistical guidance with deep modeling capacity, yielding interpretable importance scores while maximizing retention of informative signals.

Abstract: Feature selection is essential for high-dimensional biomedical data, enabling stronger predictive performance, reduced computational cost, and improved interpretability in precision medicine applications. Existing approaches face notable challenges. Filter methods are highly scalable but cannot capture complex relationships or eliminate redundancy. Deep learning-based approaches can model nonlinear patterns but often lack stability, interpretability, and efficiency at scale. Single-head attention improves interpretability but is limited in capturing multi-level dependencies and remains sensitive to initialization, reducing reproducibility. Most existing methods rarely combine statistical interpretability with the representational power of deep learning, particularly in ultra-high-dimensional settings. Here, we introduce MAFS (Multi-head Attention-based Feature Selection), a hybrid framework that integrates statistical priors with deep learning capabilities. MAFS begins with filter-based priors for stable initialization and guide learning. It then uses multi-head attention to examine features from multiple perspectives in parallel, capturing complex nonlinear relationships and interactions. Finally, a reordering module consolidates outputs across attention heads, resolving conflicts and minimizing information loss to generate robust and consistent feature rankings. This design combines statistical guidance with deep modeling capacity, yielding interpretable importance scores while maximizing retention of informative signals. Across simulated and real-world datasets, including cancer gene expression and Alzheimer’s disease data, MAFS consistently achieves superior coverage and stability compared with existing filter-based and deep learning-based alternatives, offering a scalable, interpretable, and robust solution for feature selection in high-dimensional biomedical data.

Method: Proposes Uni-FinLLM, a unified multimodal large language model using a shared Transformer backbone and modular task heads to jointly process financial text, numerical time series, fundamentals, and visual data through cross-modal attention and multi-task optimization.

Result: Significantly outperforms baselines: raises stock directional accuracy to 67.4% (from 61.7%), credit-risk accuracy to 84.1% (from 79.6%), and macro early-warning accuracy to 82.3%.

Conclusion: A unified multimodal LLM can jointly model asset behavior and systemic vulnerabilities, offering a scalable decision-support engine for finance that captures cross-scale dependencies.

Abstract: Financial institutions and regulators require systems that integrate heterogeneous data to assess risks from stock fluctuations to systemic vulnerabilities. Existing approaches often treat these tasks in isolation, failing to capture cross-scale dependencies. We propose Uni-FinLLM, a unified multimodal large language model that uses a shared Transformer backbone and modular task heads to jointly process financial text, numerical time series, fundamentals, and visual data. Through cross-modal attention and multi-task optimization, it learns a coherent representation for micro-, meso-, and macro-level predictions. Evaluated on stock forecasting, credit-risk assessment, and systemic-risk detection, Uni-FinLLM significantly outperforms baselines. It raises stock directional accuracy to 67.4% (from 61.7%), credit-risk accuracy to 84.1% (from 79.6%), and macro early-warning accuracy to 82.3%. Results validate that a unified multimodal LLM can jointly model asset behavior and systemic vulnerabilities, offering a scalable decision-support engine for finance.

Method: Analyzes decentralized gradient descent under label poisoning attacks, comparing both robust aggregators and weighted mean aggregator. Theoretical analysis examines how learning errors depend on network topology and contamination rates.

Result: Weighted mean aggregator’s performance is topology-independent, while robust aggregators’ errors depend on network topology. Weighted mean can outperform robust aggregators when: (1) global contamination rate < local contamination rate, (2) regular agent network is disconnected, or (3) network is sparse with high local contamination rate.

Conclusion: Network topology plays crucial role in robustness to label poisoning. Weighted mean aggregator, often considered vulnerable, can actually be more effective than robust aggregators under certain network conditions, challenging conventional wisdom about defense mechanisms.

Abstract: Robustness to malicious attacks is crucial for practical decentralized signal processing and machine learning systems. A typical example of such attacks is label poisoning, meaning that some agents possess corrupted local labels and share models trained on these poisoned data. To defend against malicious attacks, existing works often focus on designing robust aggregators; meanwhile, the weighted mean aggregator is typically considered a simple, vulnerable baseline. This paper analyzes the robustness of decentralized gradient descent under label poisoning attacks, considering both robust and weighted mean aggregators. Theoretical results reveal that the learning errors of robust aggregators depend on the network topology, whereas the performance of weighted mean aggregator is topology-independent. Remarkably, the weighted mean aggregator, although often considered vulnerable, can outperform robust aggregators under sufficient heterogeneity, particularly when: (i) the global contamination rate (i.e., the fraction of poisoned agents for the entire network) is smaller than the local contamination rate (i.e., the maximal fraction of poisoned neighbors for the regular agents); (ii) the network of regular agents is disconnected; or (iii) the network of regular agents is sparse and the local contamination rate is high. Empirical results support our theoretical findings, highlighting the important role of network topology in the robustness to label poisoning attacks.

Method: The study systematically analyzes scaling across two dimensions: data scale (0.1K-40K training samples) and compute scale (multiple neural network architectures with varying FLOPs). It examines both DNNs and physics-informed NNs (PINNs) and evaluates three core performance metrics: prediction error (MAE), constraint violations, and speed.

Result: The research reveals consistent power-law relationships between each resource dimension (data and compute) and performance metrics. For ACOPF, accuracy scales with dataset size and training compute. The study identifies divergence between prediction accuracy and constraint feasibility and characterizes the compute-optimal frontier.

Conclusion: This work provides quantitative guidance for ML-OPF design and deployments by establishing scaling laws that enable predictable and principled ML pipeline design for OPF applications.

Abstract: Optimal power flow (OPF) is one of the fundamental tasks for power system operations. While machine learning (ML) approaches such as deep neural networks (DNNs) have been widely studied to enhance OPF solution speed and performance, their practical deployment faces two critical scaling questions: What is the minimum training data volume required for reliable results? How should ML models’ complexity balance accuracy with real-time computational limits? Existing studies evaluate discrete scenarios without quantifying these scaling relationships, leading to trial-and-error-based ML development in real-world applications. This work presents the first systematic scaling study for ML-based OPF across two dimensions: data scale (0.1K-40K training samples) and compute scale (multiple NN architectures with varying FLOPs). Our results reveal consistent power-law relationships on both DNNs and physics-informed NNs (PINNs) between each resource dimension and three core performance metrics: prediction error (MAE), constraint violations and speed. We find that for ACOPF, the accuracy metric scales with dataset size and training compute. These scaling laws enable predictable and principled ML pipeline design for OPF. We further identify the divergence between prediction accuracy and constraint feasibility and characterize the compute-optimal frontier. This work provides quantitative guidance for ML-OPF design and deployments.

Method: Propose using actual complex linear transformations for rotary positional embeddings instead of the current implementation approach, which reduces parameter count in attention blocks.

Result: Achieves ~50% parameter reduction in attention blocks with negligible impact on model performance both in-sample and out-of-sample.

Conclusion: True complex linear transformations provide more efficient parameter usage and cleaner representation space interpretation for rotary positional embeddings.

Abstract: Rotary positional embedding has become the state-of-the-art approach to encode position information in transformer-based models. While it is often succinctly expressed in complex linear algebra, we note that the actual implementation of $Q/K/V$-projections is not equivalent to a complex linear transformation. We argue that complex linear transformation is a more natural parametrization and saves near 50% parameters within the attention block. We show empirically that removing such redundancy has negligible impact on the model performance both in sample and out of sample. Our modification achieves more efficient parameter usage, as well as a cleaner interpretation of the representation space.

Method: Introduces a family of Separable Weighted Leaf-Collision Proximities that admit exact sparse matrix factorization, restricting computation to leaf-level collisions and avoiding explicit pairwise comparisons.

Result: Empirical benchmarks show substantial runtime and memory improvements over traditional approaches, enabling tree ensemble proximities to scale to hundreds of thousands of samples on standard CPU hardware.

Conclusion: The proposed framework provides a general, efficient approach for computing tree ensemble proximities using sparse linear algebra, overcoming previous scalability limitations.

Abstract: Tree ensemble methods such as Random Forests naturally induce supervised similarity measures through their decision tree structure, but existing implementations of proximities derived from tree ensembles typically suffer from quadratic time or memory complexity, limiting their scalability. In this work, we introduce a general framework for efficient proximity computation by defining a family of Separable Weighted Leaf-Collision Proximities. We show that any proximity measure in this family admits an exact sparse matrix factorization, restricting computation to leaf-level collisions and avoiding explicit pairwise comparisons. This formulation enables low-memory, scalable proximity computation using sparse linear algebra in Python. Empirical benchmarks demonstrate substantial runtime and memory improvements over traditional approaches, allowing tree ensemble proximities to scale efficiently to datasets with hundreds of thousands of samples on standard CPU hardware.

Method: QGA integrates Q-value regularization with double Q-learning into a Decision Transformer backbone. It features a Q-value guided dual-exploration mechanism where the DT model is conditioned on multiple return-to-go targets and locally perturbed actions, with the Q-value module providing principled evaluation for candidate actions.

Result: Experiments show QGA consistently achieves superior or highly competitive results on public benchmarks and simulation environments. In large-scale real-world A/B testing, it achieves a 3.27% increase in Ad GMV and 2.49% improvement in Ad ROI.

Conclusion: QGA effectively addresses limitations of existing auto-bidding methods by combining imitation learning with value optimization, enabling better performance through principled exploration beyond data distribution while leveraging dataset experience.

Abstract: With the rapid development of e-commerce, auto-bidding has become a key asset in optimizing advertising performance under diverse advertiser environments. The current approaches focus on reinforcement learning (RL) and generative models. These efforts imitate offline historical behaviors by utilizing a complex structure with expensive hyperparameter tuning. The suboptimal trajectories further exacerbate the difficulty of policy learning. To address these challenges, we proposes QGA, a novel Q-value regularized Generative Auto-bidding method. In QGA, we propose to plug a Q-value regularization with double Q-learning strategy into the Decision Transformer backbone. This design enables joint optimization of policy imitation and action-value maximization, allowing the learned bidding policy to both leverage experience from the dataset and alleviate the adverse impact of the suboptimal trajectories. Furthermore, to safely explore the policy space beyond the data distribution, we propose a Q-value guided dual-exploration mechanism, in which the DT model is conditioned on multiple return-to-go targets and locally perturbed actions. This entire exploration process is dynamically guided by the aforementioned Q-value module, which provides principled evaluation for each candidate action. Experiments on public benchmarks and simulation environments demonstrate that QGA consistently achieves superior or highly competitive results compared to existing alternatives. Notably, in large-scale real-world A/B testing, QGA achieves a 3.27% increase in Ad GMV and a 2.49% improvement in Ad ROI.

Method: Proposes RadioDiff-Flux, a novel two-stage latent diffusion framework that decouples static environmental modeling from dynamic refinement. The first stage generates coarse latent representations using only static scene features (cacheable across similar scenarios). The second stage adapts these representations to dynamic conditions and transmitter locations using a pre-trained model, avoiding repeated early-stage computation by leveraging the consistency of latent midpoints across semantically similar scenes.

Result: RadioDiff-Flux achieves up to 50× acceleration with less than 0.15% accuracy loss compared to conventional diffusion models, demonstrating practical utility for fast, scalable radio map generation in 6G networks.

Conclusion: The proposed framework successfully addresses the latency-accuracy trade-off in radio map construction for 6G networks by exploiting structural properties of diffusion processes, enabling real-time performance while maintaining high fidelity.

Abstract: Accurate radio map (RM) construction is essential to enabling environment-aware and adaptive wireless communication. However, in future 6G scenarios characterized by high-speed network entities and fast-changing environments, it is very challenging to meet real-time requirements. Although generative diffusion models (DMs) can achieve state-of-the-art accuracy with second-level delay, their iterative nature leads to prohibitive inference latency in delay-sensitive scenarios. In this paper, by uncovering a key structural property of diffusion processes: the latent midpoints remain highly consistent across semantically similar scenes, we propose RadioDiff-Flux, a novel two-stage latent diffusion framework that decouples static environmental modeling from dynamic refinement, enabling the reuse of precomputed midpoints to bypass redundant denoising. In particular, the first stage generates a coarse latent representation using only static scene features, which can be cached and shared across similar scenarios. The second stage adapts this representation to dynamic conditions and transmitter locations using a pre-trained model, thereby avoiding repeated early-stage computation. The proposed RadioDiff-Flux significantly reduces inference time while preserving fidelity. Experiment results show that RadioDiff-Flux can achieve up to 50 acceleration with less than 0.15% accuracy loss, demonstrating its practical utility for fast, scalable RM generation in future 6G networks.

Method: SHS models per-token edits as events driven by cumulative hazard (CTMC) or cumulative jump mass (DTMC), placing events by stratifying this cumulative quantity with a single random phase per position. Tokens jump when accumulated hazard crosses unit-spaced thresholds, preserving expected jumps while minimizing variance.

Result: SHS achieves minimum possible variance among unbiased integer estimators (bounded by 1/4) without altering per-jump destination sampling, retaining multimodality. Also includes phase-allocation variant for lexical constraints that prioritizes early edits at high-risk positions.

Conclusion: SHS provides a drop-in, hyperparameter-free inference principle that reduces variance in discrete generative models, improving reproducibility by preventing characteristic failure modes of under/over-editing while maintaining multimodality.

Abstract: CTMC/DTMC-based discrete generative models, including uniform-noise discrete diffusion (e.g., D3PM/CTDD) and discrete flow matching, enable non-autoregressive sequence generation by repeatedly replacing tokens through a time-inhomogeneous Markov process. Inference is typically implemented with step-based simulation: each token decides to jump via independent Bernoulli (or categorical) draws at every discretization step. Under uniform-noise initialization, where self-correction requires multiple edits per position, these independent decisions induce substantial variance in both the number and timing of edits, leading to characteristic failure modes such as under-editing (residual noise) or over-editing (cascading unnecessary substitutions), decreasing reproducibility. We propose Stratified Hazard Sampling (SHS), a drop-in and hyperparameter-free inference principle for any sampler that admits a stay-vs.-replace decomposition. SHS models per-token edits as events driven by cumulative hazard (CTMC) or cumulative jump mass (DTMC) and places events by stratifying this cumulative quantity: with a single random phase per position, a token jumps whenever its accumulated hazard crosses unit-spaced thresholds. This preserves the expected number of jumps while achieving the minimum possible variance among unbiased integer estimators (bounded by 1/4), without altering per-jump destination sampling and thus retaining multimodality. We also introduce a phase-allocation variant for blacklist-style lexical constraints that prioritizes early edits at high-risk positions to mitigate late-masking artifacts.

Method: A novel multivariate neural network approach that combines linear and nonlinear feed-forward neural structures with online learning and forecast combination. Incorporates fundamental relationships from wind/solar generation, electricity demand, energy fuel and carbon markets, plus autoregressive dynamics and calendar effects.

Result: Significantly reduces computational cost while delivering superior forecasting accuracy with 12-13% RMSE and 15-18% MAE reductions compared to state-of-the-art benchmark models. Results validated through six-year forecasting study on major European electricity markets.

Conclusion: The proposed hybrid neural network approach effectively balances accuracy and computational efficiency for electricity price forecasting by integrating linear and nonlinear structures with comprehensive market factors and online learning techniques.

Abstract: Precise day-ahead forecasts for electricity prices are crucial to ensure efficient portfolio management, support strategic decision-making for power plant operations, enable efficient battery storage optimization, and facilitate demand response planning. However, developing an accurate prediction model is highly challenging in an uncertain and volatile market environment. For instance, although linear models generally exhibit competitive performance in predicting electricity prices with minimal computational requirements, they fail to capture relevant nonlinear relationships. Nonlinear models, on the other hand, can improve forecasting accuracy with a surge in computational costs. We propose a novel multivariate neural network approach that combines linear and nonlinear feed-forward neural structures. Unlike previous hybrid models, our approach integrates online learning and forecast combination for efficient training and accuracy improvement. It also incorporates all relevant characteristics, particularly the fundamental relationships arising from wind and solar generation, electricity demand patterns, related energy fuel and carbon markets, in addition to autoregressive dynamics and calendar effects. Compared to the current state-of-the-art benchmark models, the proposed forecasting method significantly reduces computational cost while delivering superior forecasting accuracy (12-13% RMSE and 15-18% MAE reductions). Our results are derived from a six-year forecasting study conducted on major European electricity markets.

Method: Quantum Neural Tangent Kernel-Upper Confidence Bound (QNTK-UCB) freezes QNN at random initialization and uses its static QNTK as kernel for ridge regression, bypassing unstable training dynamics.

Result: Theoretical analysis shows improved parameter scaling of Ω((TK)³) vs classical NeuralUCB’s Ω((TK)⁸) for similar regret guarantees. Empirical evaluations show superior sample efficiency in low-data regimes.

Conclusion: QNTK provides implicit regularization and sharper spectral decay, enabling potential quantum advantage in online learning by leveraging quantum inductive bias without unstable training.

Abstract: Stochastic contextual bandits are fundamental for sequential decision-making but pose significant challenges for existing neural network-based algorithms, particularly when scaling to quantum neural networks (QNNs) due to issues such as massive over-parameterization, computational instability, and the barren plateau phenomenon. This paper introduces the Quantum Neural Tangent Kernel-Upper Confidence Bound (QNTK-UCB) algorithm, a novel algorithm that leverages the Quantum Neural Tangent Kernel (QNTK) to address these limitations. By freezing the QNN at a random initialization and utilizing its static QNTK as a kernel for ridge regression, QNTK-UCB bypasses the unstable training dynamics inherent in explicit parameterized quantum circuit training while fully exploiting the unique quantum inductive bias. For a time horizon $T$ and $K$ actions, our theoretical analysis reveals a significantly improved parameter scaling of $Ω((TK)^3)$ for QNTK-UCB, a substantial reduction compared to $Ω((TK)^8)$ required by classical NeuralUCB algorithms for similar regret guarantees. Empirical evaluations on non-linear synthetic benchmarks and quantum-native variational quantum eigensolver tasks demonstrate QNTK-UCB’s superior sample efficiency in low-data regimes. This work highlights how the inherent properties of QNTK provide implicit regularization and a sharper spectral decay, paving the way for achieving ``quantum advantage’’ in online learning.

Method: Used 60-second labeled sequences of 1 Hz surface drilling data to predict SSI. Employed grid search for hyperparameter optimization. Compared Adversarial Domain Generalization (ADG), Invariant Risk Minimization (IRM), and baseline models, plus evaluated transfer learning effectiveness.

Result: ADG and IRM achieved 10% and 8% performance improvements over baseline respectively. Severe events detection improved from 20% (baseline) to 60%. ADG slightly outperformed IRM, and transfer learning further enhanced performance.

Conclusion: Domain generalization approaches are effective for drilling applications, with ADG emerging as the most promising technique. Transfer learning provides additional performance benefits, demonstrating practical value for real-world drilling operations.

Abstract: This paper provides a comprehensive comparison of domain generalization techniques applied to time series data within a drilling context, focusing on the prediction of a continuous Stick-Slip Index (SSI), a critical metric for assessing torsional downhole vibrations at the drill bit. The study aims to develop a robust regression model that can generalize across domains by training on 60 second labeled sequences of 1 Hz surface drilling data to predict the SSI. The model is tested in wells that are different from those used during training. To fine-tune the model architecture, a grid search approach is employed to optimize key hyperparameters. A comparative analysis of the Adversarial Domain Generalization (ADG), Invariant Risk Minimization (IRM) and baseline models is presented, along with an evaluation of the effectiveness of transfer learning (TL) in improving model performance. The ADG and IRM models achieve performance improvements of 10% and 8%, respectively, over the baseline model. Most importantly, severe events are detected 60% of the time, against 20% for the baseline model. Overall, the results indicate that both ADG and IRM models surpass the baseline, with the ADG model exhibiting a slight advantage over the IRM model. Additionally, applying TL to a pre-trained model further improves performance. Our findings demonstrate the potential of domain generalization approaches in drilling applications, with ADG emerging as the most effective approach.

Method: Proposes Residual-Projected Multi-Collaboration Closed-Loop and Single Instance Quantization (RPIQ) framework with multi-collaborative closed-loop compensation scheme based on Single Instance Calibration and Gauss-Seidel Iterative Quantization.

Result: Compresses models to 4-bit representation with 60-75% memory reduction compared to full-precision models. Maintains performance close to full-precision models across language and visual tasks, with excellent recognition in text understanding and visual question answering.

Conclusion: RPIQ enables efficient deployment of large models on assistive devices for visually impaired users while advancing computational efficiency and reliability of large models for accurate and rapid information access.

Abstract: Visually impaired users face significant challenges in daily information access and real-time environmental perception, and there is an urgent need for intelligent assistive systems with accurate recognition capabilities. Although large-scale models provide effective solutions for perception and reasoning, their practical deployment on assistive devices is severely constrained by excessive memory consumption and high inference costs. Moreover, existing quantization strategies often ignore inter-block error accumulation, leading to degraded model stability. To address these challenges, this study proposes a novel quantization framework – Residual-Projected Multi-Collaboration Closed-Loop and Single Instance Quantization(RPIQ), whose quantization process adopts a multi-collaborative closed-loop compensation scheme based on Single Instance Calibration and Gauss-Seidel Iterative Quantization. Experiments on various types of large-scale models, including language models such as OPT, Qwen, and LLaMA, as well as vision-language models such as CogVLM2, demonstrate that RPIQ can compress models to 4-bit representation while significantly reducing peak memory consumption (approximately 60%-75% reduction compared to original full-precision models). The method maintains performance highly close to full-precision models across multiple language and visual tasks, and exhibits excellent recognition and reasoning capabilities in key applications such as text understanding and visual question answering in complex scenarios. While verifying the effectiveness of RPIQ for deployment in real assistive systems, this study also advances the computational efficiency and reliability of large models, enabling them to provide visually impaired users with the required information accurately and rapidly.

Method: Proposes RESGA and SAEGA methods that adapt gradient ascent to LLMs, optimizing randomly initialized prompts to achieve better alignment with identified persona directions, with fluent gradient ascent to control prompt fluency.

Result: Demonstrates effectiveness across Llama 3.1, Qwen 2.5, and Gemma 3 for steering sycophancy, hallucination, and myopic reward personas. On sycophancy, automatically discovered prompts achieve significant improvement (49.90% compared with 79.24%).

Conclusion: By grounding prompt discovery in mechanistically meaningful features, the method offers a new paradigm for controllable and interpretable behavior modification in LLMs.

Abstract: Controlling emergent behavioral personas (e.g., sycophancy, hallucination) in Large Language Models (LLMs) is critical for AI safety, yet remains a persistent challenge. Existing solutions face a dilemma: manual prompt engineering is intuitive but unscalable and imprecise, while automatic optimization methods are effective but operate as “black boxes” with no interpretable connection to model internals. We propose a novel framework that adapts gradient ascent to LLMs, enabling targeted prompt discovery. In specific, we propose two methods, RESGA and SAEGA, that both optimize randomly initialized prompts to achieve better aligned representation with an identified persona direction. We introduce fluent gradient ascent to control the fluency of discovered persona steering prompts. We demonstrate RESGA and SAEGA’s effectiveness across Llama 3.1, Qwen 2.5, and Gemma 3 for steering three different personas,sycophancy, hallucination, and myopic reward. Crucially, on sycophancy, our automatically discovered prompts achieve significant improvement (49.90% compared with 79.24%). By grounding prompt discovery in mechanistically meaningful features, our method offers a new paradigm for controllable and interpretable behavior modification.

Method: ChemBART uses a SMILES-based LLM pre-trained on chemical reactions with mask-filling pre-training. The model is applied to multiple downstream tasks including precursor/reagent generation, temperature-yield regression, molecular property classification, and reinforcement learning with Monte Carlo tree search for multi-step synthesis route design.

Result: ChemBART achieves the paradigm of “one model, one pre-training, multiple tasks” and demonstrates superior performance in synthesis planning. Wet-lab validation confirmed that ChemBART-designed multi-step synthesis routes achieved shorter pathways with approximately 30% yield improvement over literature benchmarks.

Conclusion: The work validates the power of reaction-focused pre-training and showcases ChemBART’s broad utility in advancing the complete synthesis planning cycle, demonstrating that a unified LLM approach can effectively address multiple chemical challenges simultaneously.

Abstract: Recent advances in large language models (LLMs) have demonstrated transformative potential across diverse fields. While LLMs have been applied to molecular simplified molecular input line entry system (SMILES) in computer-aided synthesis planning (CASP), existing methodologies typically address single tasks, such as precursor prediction. We introduce ChemBART, a SMILES-based LLM pre-trained on chemical reactions, which enables a unified model for multiple downstream chemical tasks–achieving the paradigm of “one model, one pre-training, multiple tasks.” By leveraging outputs from a mask-filling pre-training task on reaction expressions, ChemBART effectively solves a variety of chemical problems, including precursor/reagent generation, temperature-yield regression, molecular property classification, and optimizing the policy and value functions within a reinforcement learning framework, integrated with Monte Carlo tree search for multi-step synthesis route design. Unlike single-molecule pre-trained LLMs constrained to specific applications, ChemBART addresses broader chemical challenges and integrates them for comprehensive synthesis planning. Crucially, ChemBART-designed multi-step synthesis routes and reaction conditions directly inspired wet-lab validation, which confirmed shorter pathways with ~30% yield improvement over literature benchmarks. Our work validates the power of reaction-focused pre-training and showcases the broad utility of ChemBART in advancing the complete synthesis planning cycle.

Method: A code-oriented LLM is placed in a closed-loop synthesis framework with 22 fine-tuning cycles. It generates PyTorch convolutional networks, which are validated, evaluated via single-epoch accuracy, and filtered for structural redundancy using MinHash-Jaccard. High-performing novel architectures are converted to prompt-code pairs for iterative LoRA fine-tuning initialized from LEMUR dataset.

Result: Valid generation rate stabilized at 50.6% (peaking at 74.5%), mean first-epoch accuracy rose from 28.06% to 50.99%, and fraction of candidates exceeding 40% accuracy grew from 2.04% to 96.81%. The model synthesized 455 high-performing architectures absent from original corpus.

Conclusion: LLMs can internalize empirical, non-textual rewards to transcend their training data, providing a scalable blueprint for transforming stochastic generators into autonomous, performance-driven neural designers.

Abstract: Large language models (LLMs) excel in program synthesis, yet their ability to autonomously navigate neural architecture design–balancing syntactic reliability, performance, and structural novelty–remains underexplored. We address this by placing a code-oriented LLM within a closed-loop synthesis framework, analyzing its evolution over 22 supervised fine-tuning cycles. The model synthesizes PyTorch convolutional networks which are validated, evaluated via low-fidelity performance signals (single-epoch accuracy), and filtered using a MinHash-Jaccard criterion to prevent structural redundancy. High-performing, novel architectures are converted into prompt-code pairs for iterative fine-tuning via parameter-efficient LoRA adaptation, initialized from the LEMUR dataset. Across cycles, the LLM internalizes empirical architectural priors, becoming a robust generator. The valid generation rate stabilizes at 50.6 percent (peaking at 74.5 percent), while mean first-epoch accuracy rises from 28.06 percent to 50.99 percent, and the fraction of candidates exceeding 40 percent accuracy grows from 2.04 percent to 96.81 percent. Analyses confirm the model moves beyond replicating existing motifs, synthesizing 455 high-performing architectures absent from the original corpus. By grounding code synthesis in execution feedback, this work provides a scalable blueprint for transforming stochastic generators into autonomous, performance-driven neural designers, establishing that LLMs can internalize empirical, non-textual rewards to transcend their training data.

Method: Max-p aggregation scheme that delivers finite-sample multi-distribution coverage using any conformity scores; optimization programs for efficiency; algorithm to learn optimal conformity scores.

Result: Method provides valid worst-case coverage across multiple distributions while reducing set size compared to naive max-p aggregation, achieving sizes comparable to single-source prediction sets.

Conclusion: The framework connects to group-wise distributionally robust optimization, sub-population shift, fairness, and multi-source learning, offering practical solutions for distribution-robust conformal prediction.

Abstract: In many fairness and distribution robustness problems, one has access to labeled data from multiple source distributions yet the test data may come from an arbitrary member or a mixture of them. We study the problem of constructing a conformal prediction set that is uniformly valid across multiple, heterogeneous distributions, in the sense that no matter which distribution the test point is from, the coverage of the prediction set is guaranteed to exceed a pre-specified level. We first propose a max-p aggregation scheme that delivers finite-sample, multi-distribution coverage given any conformity scores associated with each distribution. Upon studying several efficiency optimization programs subject to uniform coverage, we prove the optimality and tightness of our aggregation scheme, and propose a general algorithm to learn conformity scores that lead to efficient prediction sets after the aggregation under standard conditions. We discuss how our framework relates to group-wise distributionally robust optimization, sub-population shift, fairness, and multi-source learning. In synthetic and real-data experiments, our method delivers valid worst-case coverage across multiple distributions while greatly reducing the set size compared with naively applying max-p aggregation to single-source conformity scores, and can be comparable in size to single-source prediction sets with popular, standard conformity scores.

Method: SPICE learns a prior over Q-values using deep ensembles during training. At test time, it performs Bayesian updates on this prior using in-context information. To handle poor priors from suboptimal training data, it uses an Upper-Confidence Bound rule for online inference that favors exploration and adaptation.

Result: Theoretical analysis proves SPICE achieves regret-optimal behavior in stochastic bandits and finite-horizon MDPs, even when pretrained on suboptimal trajectories. Empirical validation across bandit and control benchmarks shows SPICE achieves near-optimal decisions on unseen tasks, substantially reduces regret compared to prior methods, rapidly adapts to unseen tasks, and remains robust under distribution shift.

Conclusion: SPICE provides a practical Bayesian ICRL approach that enables effective adaptation to unseen environments without parameter updates, works with suboptimal training data, and offers theoretical guarantees on regret-optimal behavior across different problem settings.

Abstract: In-context reinforcement learning (ICRL) promises fast adaptation to unseen environments without parameter updates, but current methods either cannot improve beyond the training distribution or require near-optimal data, limiting practical adoption. We introduce SPICE, a Bayesian ICRL method that learns a prior over Q-values via deep ensemble and updates this prior at test-time using in-context information through Bayesian updates. To recover from poor priors resulting from training on sub-optimal data, our online inference follows an Upper-Confidence Bound rule that favours exploration and adaptation. We prove that SPICE achieves regret-optimal behaviour in both stochastic bandits and finite-horizon MDPs, even when pretrained only on suboptimal trajectories. We validate these findings empirically across bandit and control benchmarks. SPICE achieves near-optimal decisions on unseen tasks, substantially reduces regret compared to prior ICRL and meta-RL approaches while rapidly adapting to unseen tasks and remaining robust under distribution shift.

Method: CMF learns latent representations where local Riemannian geometry (metric tensor and curvature) remains invariant under counterfactual interventions on sensitive attributes, enforcing constraints on decoder Jacobian and Hessian.

Result: CMF effectively disentangles sensitive geometric warping while preserving task utility on synthetic Structural Causal Models, offering rigorous quantification of fairness-utility trade-off via geometric metrics.

Conclusion: CMF provides a novel framework bridging causal inference and geometric deep learning to address fairness by preserving the underlying geometric structure of data across demographic groups.

Abstract: Fairness in machine learning is increasingly critical, yet standard approaches often treat data as static points in a high-dimensional space, ignoring the underlying generative structure. We posit that sensitive attributes (e.g., race, gender) do not merely shift data distributions but causally warp the geometry of the data manifold itself. To address this, we introduce Causal Manifold Fairness (CMF), a novel framework that bridges causal inference and geometric deep learning. CMF learns a latent representation where the local Riemannian geometry, defined by the metric tensor and curvature, remains invariant under counterfactual interventions on sensitive attributes. By enforcing constraints on the Jacobian and Hessian of the decoder, CMF ensures that the rules of the latent space (distances and shapes) are preserved across demographic groups. We validate CMF on synthetic Structural Causal Models (SCMs), demonstrating that it effectively disentangles sensitive geometric warping while preserving task utility, offering a rigorous quantification of the fairness-utility trade-off via geometric metrics.

Method: Replicated Anthropic’s main results using open-source sparse autoencoders (SAEs) for Llama 3.1, testing feature extraction and steering capabilities while examining sensitivity to layer selection, steering magnitude, and context.

Result: Successfully reproduced basic feature extraction and steering, but found substantial fragility: feature steering is sensitive to layer selection, steering magnitude, and context; observed non-standard activation behavior; and demonstrated difficulty distinguishing thematically similar features.

Conclusion: Current SAE-based interpretability methods fall short of systematic reliability needed for safety-critical applications, suggesting a necessary shift from prioritizing interpretability of internal representations toward reliable prediction and control of model output.

Abstract: Recent work by Anthropic on Mechanistic interpretability claims to understand and control Large Language Models by extracting human-interpretable features from their neural activation patterns using sparse autoencoders (SAEs). If successful, this approach offers one of the most promising routes for human oversight in AI safety. We conduct an initial stress-test of these claims by replicating their main results with open-source SAEs for Llama 3.1. While we successfully reproduce basic feature extraction and steering capabilities, our investigation suggests that major caution is warranted regarding the generalizability of these claims. We find that feature steering exhibits substantial fragility, with sensitivity to layer selection, steering magnitude, and context. We observe non-standard activation behavior and demonstrate the difficulty to distinguish thematically similar features from one another. While SAE-based interpretability produces compelling demonstrations in selected cases, current methods often fall short of the systematic reliability required for safety-critical applications. This suggests a necessary shift in focus from prioritizing interpretability of internal representations toward reliable prediction and control of model output. Our work contributes to a more nuanced understanding of what mechanistic interpretability has achieved and highlights fundamental challenges for AI safety that remain unresolved.

Method: Joint encoding of KV-cache blocks that fuses similar blocks across requests and input chunks into shared representations while preserving standard cache structure. Theoretically analyzed using Poisson process model.

Result: Achieves up to 4.38× KV-cache compression with negligible accuracy loss across diverse LLMs and benchmarks. In real serving, improves token throughput by ~40% on single-machine vLLM benchmark.

Conclusion: Joint encoding effectively alleviates KV-cache memory bottleneck, supports high-concurrency serving without specialized hardware, and outperforms existing structured and adaptive compression baselines.

Abstract: Modern large language models (LLMs) drive interactive AI systems but are bottlenecked by the memory-heavy growth of key-value (KV) caches, which limits real-time throughput under concurrent loads. Existing KV-cache compression methods rely on rigid heuristics, disrupt tensor layouts, or require specialized compute, hindering scalability and deployment. We propose joint encoding of KV-cache blocks, which fuses similar blocks across requests and input chunks into shared representations while preserving standard cache structure. This alleviates the KV-cache memory bottleneck, supporting high-concurrency serving without specialized hardware. Theoretically, we analyze the rate-distortion tradeoff of fused cache blocks under a Poisson process model. Empirically, our method achieves up to 4.38 $\times$ KV-cache compression with negligible accuracy loss across diverse LLMs and benchmarks, outperforming recent structured and adaptive compression baselines. In real LLM serving, joint encoding improves the token throughput by $\sim$40% on a single-machine vLLM benchmark, demonstrating substantial gains in inference throughput. Code is available at https://github.com/sef1/kv_fast_fusion kv_joint_encoding.

Method: Uses multi-source prediction model combined with novel concept drift adaptation method for anomaly detection in IIoT streaming data, focusing on handling data complexity and dynamism.

Result: Achieves up to 89.71% accuracy (AUC) in trace-driven evaluations, outperforming state-of-the-art methods while meeting efficiency and scalability requirements.

Conclusion: Proposed adaptive anomaly detection method effectively handles IIoT streaming data complexity, providing accurate, efficient, and scalable solution for next-generation network reliability.

Abstract: To ensure reliability and service availability, next-generation networks are expected to rely on automated anomaly detection systems powered by advanced machine learning methods with the capability of handling multi-dimensional data. Such multi-dimensional, heterogeneous data occurs mostly in today’s industrial Internet of Things (IIoT), where real-time detection of anomalies is critical to prevent impending failures and resolve them in a timely manner. However, existing anomaly detection methods often fall short of effectively coping with the complexity and dynamism of multi-dimensional data streams in IIoT. In this paper, we propose an adaptive method for detecting anomalies in IIoT streaming data utilizing a multi-source prediction model and concept drift adaptation. The proposed anomaly detection algorithm merges a prediction model into a novel drift adaptation method resulting in accurate and efficient anomaly detection that exhibits improved scalability. Our trace-driven evaluations indicate that the proposed method outperforms the state-of-the-art anomaly detection methods by achieving up to an 89.71% accuracy (in terms of Area under the Curve (AUC)) while meeting the given efficiency and scalability requirements.

Method: BAFA treats auditing as uncertainty estimation over fairness metrics, maintaining a version space of surrogate models consistent with queried scores and using active query selection to narrow uncertainty intervals via constrained empirical risk minimization.

Result: BAFA achieves target error thresholds with up to 40x fewer queries than stratified sampling (e.g., 144 vs 5,956 queries at ε=0.02 for CivilComments), shows better performance over time, and has lower variance across runs.

Conclusion: Active sampling can significantly reduce resources needed for independent fairness auditing of LLMs, enabling more practical continuous model evaluations.

Abstract: Large Language Models (LLMs) exhibit systematic biases across demographic groups. Auditing is proposed as an accountability tool for black-box LLM applications, but suffers from resource-intensive query access. We conceptualise auditing as uncertainty estimation over a target fairness metric and introduce BAFA, the Bounded Active Fairness Auditor for query-efficient auditing of black-box LLMs. BAFA maintains a version space of surrogate models consistent with queried scores and computes uncertainty intervals for fairness metrics (e.g., $Δ$ AUC) via constrained empirical risk minimisation. Active query selection narrows these intervals to reduce estimation error. We evaluate BAFA on two standard fairness dataset case studies: \textsc{CivilComments} and \textsc{Bias-in-Bios}, comparing against stratified sampling, power sampling, and ablations. BAFA achieves target error thresholds with up to 40$\times$ fewer queries than stratified sampling (e.g., 144 vs 5,956 queries at $\varepsilon=0.02$ for \textsc{CivilComments}) for tight thresholds, demonstrates substantially better performance over time, and shows lower variance across runs. These results suggest that active sampling can reduce resources needed for independent fairness auditing with LLMs, supporting continuous model evaluations.

Method: ATLAS employs an external lightweight latent verifier that analyzes intermediate hidden states to predict reasoning quality, then adaptively selects whether and how strongly to apply steering on a per-example, per-step basis during inference.

Result: Experiments on multiple mathematical reasoning benchmarks show ATLAS consistently outperforms vanilla decoding and fixed steering baselines, achieving higher accuracy while substantially reducing test-time token usage.

Conclusion: Verifier-guided latent adaptation provides an effective and scalable mechanism for controlling reasoning efficiency without sacrificing solution quality, representing the first integration of learned latent verification into test-time steering for LLMs.

Abstract: Recent work on activation and latent steering has demonstrated that modifying internal representations can effectively guide large language models (LLMs) toward improved reasoning and efficiency without additional training. However, most existing approaches rely on fixed steering policies and static intervention strengths, which limit their robustness across problem instances and often result in over- or under-steering. We propose Adaptive Test-time Latent Steering, called (ATLAS), a task-specific framework that dynamically controls steering decisions at inference time using an external, lightweight latent verifier. Given intermediate hidden states, the verifier predicts the quality of ongoing reasoning and adaptively selects whether and how strongly to apply steering, enabling per-example and per-step adjustment with minimal overhead. To our knowledge, ATLAS is the first method to integrate learned latent verification into test-time steering for enhancing LLMs reasoning. Experiments on multiple mathematical reasoning benchmarks show that ATLAS consistently outperforms both vanilla decoding and fixed steering baselines, achieving higher accuracy while substantially reducing test-time token usage. These results demonstrate that verifier-guided latent adaptation provides an effective and scalable mechanism for controlling reasoning efficiency without sacrificing solution quality. All source code will be publicly available.

Method: Hierarchical framework with three stages: 1) CNN-BiLSTM-Attention model classifies finger activations from multichannel sEMG, 2) ergonomic typing priors infer letters from finger patterns, 3) fine-tuned T5 transformer reconstructs full sentences.

Result: Achieved 85.4% finger-classification accuracy, 5.4% character error rate (CER), and 6.5% word error rate (WER) on 30 users from emg2qwerty dataset, outperforming baselines.

Conclusion: MyoText establishes a principled pathway from neuromuscular signals to text, providing a blueprint for virtual/augmented-reality typing interfaces without physical keyboards, advancing wearable neural input for ubiquitous computing.

Abstract: Surface electromyography (sEMG) provides a direct neural interface for decoding muscle activity and offers a promising foundation for keyboard-free text input in wearable and mixed-reality systems. Previous sEMG-to-text studies mainly focused on recognizing letters directly from sEMG signals, forming an important first step toward translating muscle activity into text. Building on this foundation, we present MyoText, a hierarchical framework that decodes sEMG signals to text through physiologically grounded intermediate stages. MyoText first classifies finger activations from multichannel sEMG using a CNN-BiLSTM-Attention model, applies ergonomic typing priors to infer letters, and reconstructs full sentences with a fine-tuned T5 transformer. This modular design mirrors the natural hierarchy of typing, linking muscle intent to language output and reducing the search space for decoding. Evaluated on 30 users from the emg2qwerty dataset, MyoText outperforms baselines by achieving 85.4% finger-classification accuracy, 5.4% character error rate (CER), and 6.5% word error rate (WER). Beyond accuracy gains, this methodology establishes a principled pathway from neuromuscular signals to text, providing a blueprint for virtual and augmented-reality typing interfaces that operate entirely without physical keyboards. By integrating ergonomic structure with transformer-based linguistic reasoning, MyoText advances the feasibility of seamless, wearable neural input for future ubiquitous computing environments.

Method: TASC uses a state-space model with constant trend while preserving low-rank signal structure. It employs Kalman filter and Rauch-Tung-Striebel smoother, fitting a generative time-series model with expectation-maximization, then performs counterfactual inference.

Result: Evaluation on simulated and real-world datasets (policy evaluation and sports prediction) shows TASC offers advantages in settings with strong temporal trends and high levels of observation noise.

Conclusion: TASC successfully incorporates temporal structure into synthetic control framework, improving causal inference in time-series panel data with trends and noise, addressing limitations of existing methods.

Abstract: The synthetic control (SC) framework is widely used for observational causal inference with time-series panel data. SC has been successful in diverse applications, but existing methods typically treat the ordering of pre-intervention time indices interchangeable. This invariance means they may not fully take advantage of temporal structure when strong trends are present. We propose Time-Aware Synthetic Control (TASC), which employs a state-space model with a constant trend while preserving a low-rank structure of the signal. TASC uses the Kalman filter and Rauch-Tung-Striebel smoother: it first fits a generative time-series model with expectation-maximization and then performs counterfactual inference. We evaluate TASC on both simulated and real-world datasets, including policy evaluation and sports prediction. Our results suggest that TASC offers advantages in settings with strong temporal trends and high levels of observation noise.

Method: Introduces “polymath learning” - a framework for designing one training sample that elicits multidisciplinary impact. The approach involves: (1) identifying key math reasoning skills, (2) determining characteristics of optimal polymath samples, and (3) engineering synthetic samples that integrate multidisciplinary elements.

Result: Three key findings: (1) A single strategically selected math reasoning sample produces significant performance improvements across physics, chemistry, and biology; (2) Math skills salient to reasoning reveal characteristics of optimal polymath samples; (3) Engineered synthetic samples integrating multidisciplinary elements outperform training with naturally occurring individual samples.

Conclusion: Sample quality and design, rather than quantity, may be key to unlocking enhanced reasoning capabilities in LLMs. The results suggest a shift toward “sample engineering” - precision engineering of training samples rather than simply increasing data volume.

Abstract: The reasoning ability of large language models (LLMs) can be unleashed with reinforcement learning (RL) (OpenAI, 2024; DeepSeek-AI et al., 2025a; Zeng et al., 2025). The success of existing RL attempts in LLMs usually relies on high-quality samples of thousands or beyond. In this paper, we challenge fundamental assumptions about data requirements in RL for LLMs by demonstrating the remarkable effectiveness of one-shot learning. Specifically, we introduce polymath learning, a framework for designing one training sample that elicits multidisciplinary impact. We present three key findings: (1) A single, strategically selected math reasoning sample can produce significant performance improvements across multiple domains, including physics, chemistry, and biology with RL; (2) The math skills salient to reasoning suggest the characteristics of the optimal polymath sample; and (3) An engineered synthetic sample that integrates multidiscipline elements outperforms training with individual samples that naturally occur. Our approach achieves superior performance to training with larger datasets across various reasoning benchmarks, demonstrating that sample quality and design, rather than quantity, may be the key to unlock enhanced reasoning capabilities in language models. Our results suggest a shift, dubbed as sample engineering, toward precision engineering of training samples rather than simply increasing data volume.

Method: PersonaLedger uses a large language model conditioned on rich user personas to generate diverse transaction streams, coupled with a configurable programmatic engine that enforces financial rules. The LLM and engine interact in a closed loop: after each event, the engine updates user state, enforces rules, and returns a context-aware “nextprompt” to guide the LLM toward feasible actions.

Result: Created a public dataset of 30 million transactions from 23,000 users and a benchmark suite with two tasks: illiquidity classification and identity theft segmentation. The system produces realistic, privacy-preserving financial data that maintains both behavioral diversity and logical groundedness.

Conclusion: PersonaLedger provides a realistic, privacy-preserving resource for financial AI research, enabling rigorous evaluation of forecasting and anomaly detection models while accelerating innovation through publicly available code, rules, and generation logs.

Abstract: Strict privacy regulations limit access to real transaction data, slowing open research in financial AI. Synthetic data can bridge this gap, but existing generators do not jointly achieve behavioral diversity and logical groundedness. Rule-driven simulators rely on hand-crafted workflows and shallow stochasticity, which miss the richness of human behavior. Learning-based generators such as GANs capture correlations yet often violate hard financial constraints and still require training on private data. We introduce PersonaLedger, a generation engine that uses a large language model conditioned on rich user personas to produce diverse transaction streams, coupled with an expert configurable programmatic engine that maintains correctness. The LLM and engine interact in a closed loop: after each event, the engine updates the user state, enforces financial rules, and returns a context aware “nextprompt” that guides the LLM toward feasible next actions. With this engine, we create a public dataset of 30 million transactions from 23,000 users and a benchmark suite with two tasks, illiquidity classification and identity theft segmentation. PersonaLedger offers a realistic, privacy preserving resource that supports rigorous evaluation of forecasting and anomaly detection models. PersonaLedger offers the community a rich, realistic, and privacy preserving resource – complete with code, rules, and generation logs – to accelerate innovation in financial AI and enable rigorous, reproducible evaluation.

Method: The paper proposes a flexible framework adapting counterfactual explanations to non-deterministic generative AI systems, using downstream classifiers to reveal output characteristics. It introduces an algorithm for generating prompt-counterfactual explanations (PCEs) that identify what about the prompt causes specific output characteristics.

Result: Three case studies demonstrate PCEs for different output characteristics: political leaning, toxicity, and sentiment. The results show PCEs can streamline prompt engineering to suppress undesirable outputs and enhance red-teaming efforts to uncover prompts that elicit undesirable outputs.

Conclusion: This work establishes a foundation for prompt-focused interpretability in generative AI, which will become essential as models handle higher-stakes tasks and face regulatory requirements for transparency and accountability.

Abstract: As generative AI systems become integrated into real-world applications, organizations increasingly need to be able to understand and interpret their behavior. In particular, decision-makers need to understand what causes generative AI systems to exhibit specific output characteristics. Within this general topic, this paper examines a key question: what is it about the input – the prompt – that causes an LLM-based generative AI system to produce output that exhibits specific characteristics, such as toxicity, negative sentiment, or political bias. To examine this question, we adapt a common technique from the Explainable AI literature: counterfactual explanations. We explain why traditional counterfactual explanations cannot be applied directly to generative AI systems, due to several differences in how generative AI systems function. We then propose a flexible framework that adapts counterfactual explanations to non-deterministic, generative AI systems in scenarios where downstream classifiers can reveal key characteristics of their outputs. Based on this framework, we introduce an algorithm for generating prompt-counterfactual explanations (PCEs). Finally, we demonstrate the production of counterfactual explanations for generative AI systems with three case studies, examining different output characteristics (viz., political leaning, toxicity, and sentiment). The case studies further show that PCEs can streamline prompt engineering to suppress undesirable output characteristics and can enhance red-teaming efforts to uncover additional prompts that elicit undesirable outputs. Ultimately, this work lays a foundation for prompt-focused interpretability in generative AI: a capability that will become indispensable as these models are entrusted with higher-stakes tasks and subject to emerging regulatory requirements for transparency and accountability.

Method: Developed RATS (Rapid Augmentations for Time Series) in Rust with Python bindings (RATSpy), implementing multiple augmentation methods including basic transformations, frequency-domain operations, and time warping techniques with a unified pipeline interface and built-in parallelization.

Result: Benchmarking on 143 datasets shows RATSpy achieves average 74.5% speedup over tsaug (up to 94.8% on large datasets) with up to 47.9% less peak memory usage.

Conclusion: RATS provides a high-performance solution for time series augmentation that addresses scalability limitations of existing Python libraries, making it suitable for large-scale production systems.

Abstract: Time series augmentation is critical for training robust deep learning models, particularly in domains where labelled data is scarce and expensive to obtain. However, existing augmentation libraries for time series, mainly written in Python, suffer from performance bottlenecks, where running time grows exponentially as dataset sizes increase – an aspect limiting their applicability in large-scale, production-grade systems. We introduce RATS (Rapid Augmentations for Time Series), a high-performance library for time series augmentation written in Rust with Python bindings (RATSpy). RATS implements multiple augmentation methods spanning basic transformations, frequency-domain operations and time warping techniques, all accessible through a unified pipeline interface with built-in parallelisation. Comprehensive benchmarking of RATSpy versus a commonly used library (tasug) on 143 datasets demonstrates that RATSpy achieves an average speedup of 74.5% over tsaug (up to 94.8% on large datasets), with up to 47.9% less peak memory usage.

Method: Theoretical and empirical analysis of preconditioned gradient descent (PGD), particularly Gauss-Newton, examining its impact on spectral bias and grokking. Experimental validation of how PGD enables uniform parameter space exploration in the NTK regime.

Result: PGD mitigates spectral bias issues and reduces grokking delays. Experimental results confirm that grokking represents transitional behavior between lazy (NTK) and rich learning regimes, and PGD facilitates this transition.

Conclusion: PGD provides a mechanism to overcome spectral bias limitations and accelerate the transition from lazy to rich learning regimes, deepening understanding of optimization dynamics, spectral bias, and neural network learning phases.

Abstract: Spectral bias, the tendency of neural networks to learn low frequencies first, can be both a blessing and a curse. While it enhances the generalization capabilities by suppressing high-frequency noise, it can be a limitation in scientific tasks that require capturing fine-scale structures. The delayed generalization phenomenon known as grokking is another barrier to rapid training of neural networks. Grokking has been hypothesized to arise as learning transitions from the NTK to the feature-rich regime. This paper explores the impact of preconditioned gradient descent (PGD), such as Gauss-Newton, on spectral bias and grokking phenomena. We demonstrate through theoretical and empirical results how PGD can mitigate issues associated with spectral bias. Additionally, building on the rich learning regime grokking hypothesis, we study how PGD can be used to reduce delays associated with grokking. Our conjecture is that PGD, without the impediment of spectral bias, enables uniform exploration of the parameter space in the NTK regime. Our experimental results confirm this prediction, providing strong evidence that grokking represents a transitional behavior between the lazy regime characterized by the NTK and the rich regime. These findings deepen our understanding of the interplay between optimization dynamics, spectral bias, and the phases of neural network learning.

Method: Integrates HyperSHAP-based hyperparameter importance analysis into ParEGO MOO algorithm, dynamically adapts configuration space by fixing unimportant hyperparameters to focus search on influential ones.

Result: Empirical validation on PyMOO and YAHPO-Gym tasks shows improved convergence speed and Pareto front quality compared to baseline methods.

Conclusion: Dynamic prioritization of influential hyperparameters based on objective trade-offs accelerates convergence and yields better multi-objective solutions than traditional approaches.

Abstract: Choosing a suitable ML model is a complex task that can depend on several objectives, e.g., accuracy, model size, fairness, inference time, or energy consumption. In practice, this requires trading off multiple, often competing, objectives through multi-objective optimization (MOO). However, existing MOO methods typically treat all hyperparameters as equally important, overlooking that hyperparameter importance (HPI) can vary significantly depending on the trade-off between objectives. We propose a novel dynamic optimization approach that prioritizes the most influential hyperparameters based on varying objective trade-offs during the search process, which accelerates empirical convergence and leads to better solutions. Building on prior work on HPI for MOO post-analysis, we now integrate HPI, calculated with HyperSHAP, into the optimization. For this, we leverage the objective weightings naturally produced by the MOO algorithm ParEGO and adapt the configuration space by fixing the unimportant hyperparameters, allowing the search to focus on the important ones. Eventually, we validate our method with diverse tasks from PyMOO and YAHPO-Gym. Empirical results demonstrate improvements in convergence speed and Pareto front quality compared to baselines.

Method: Created a large-scale dataset of 129,000+ labeled multivariate time-series sequences from 153 rides by 51 participants under No, Low, and High Time Pressure conditions. Proposed MotoTimePressure, a deep learning model combining convolutional preprocessing, dual-stage temporal attention, and Squeeze-and-Excitation feature recalibration.

Result: High Time Pressure induces 48% higher speeds, 36.4% greater speed variability, 58% more risky turns, 36% more sudden braking, and 50% higher rear brake forces. MotoTimePressure achieves 91.53% accuracy and 98.93% ROC AUC, outperforming eight baselines. Using predicted time pressure improves collision risk accuracy from 91.25% to 93.51%.

Conclusion: Time pressure prediction enables proactive ITS interventions (adaptive alerts, haptic feedback, V2I signaling, speed guidance) to support safer two-wheeler mobility under the Safe System Approach. The dataset and model provide valuable tools for understanding and mitigating time-pressure-induced risks.

Abstract: Time pressure critically influences risky maneuvers and crash proneness among powered two-wheeler riders, yet its prediction remains underexplored in intelligent transportation systems. We present a large-scale dataset of 129,000+ labeled multivariate time-series sequences from 153 rides by 51 participants under No, Low, and High Time Pressure conditions. Each sequence captures 63 features spanning vehicle kinematics, control inputs, behavioral violations, and environmental context. Our empirical analysis shows High Time Pressure induces 48% higher speeds, 36.4% greater speed variability, 58% more risky turns at intersections, 36% more sudden braking, and 50% higher rear brake forces versus No Time Pressure. To benchmark this dataset, we propose MotoTimePressure, a deep learning model combining convolutional preprocessing, dual-stage temporal attention, and Squeeze-and-Excitation feature recalibration, achieving 91.53% accuracy and 98.93% ROC AUC, outperforming eight baselines. Since time pressure cannot be directly measured in real time, we demonstrate its utility in collision prediction and threshold determination. Using MTPS-predicted time pressure as features, improves Informer-based collision risk accuracy from 91.25% to 93.51%, approaching oracle performance (93.72%). Thresholded time pressure states capture rider cognitive stress and enable proactive ITS interventions, including adaptive alerts, haptic feedback, V2I signaling, and speed guidance, supporting safer two-wheeler mobility under the Safe System Approach.

Method: Defines Decentralized Discrete Flow Matching objective by expressing probability generating velocity as linear combination of expert flows. Compares decentralized vs centralized training paradigms using LLaVA and InternVL 2.5-1B models with fixed CLIP vision encoder and full-parameter fine-tuning during instruction tuning.

Result: Experiments demonstrate equivalence between decentralized and centralized training settings for multimodal language models across diverse benchmarks, suggesting decentralized approaches can achieve comparable performance.

Conclusion: Decentralized training for autoregressive generation is theoretically sound and practically viable, with demonstrated equivalence to centralized approaches for multimodal language models, potentially enabling more scalable and distributed training paradigms.

Abstract: We present a theoretical analysis of decentralization of autoregressive generation. We define the Decentralized Discrete Flow Matching objective, by expressing probability generating velocity as a linear combination of expert flows. We also conduct experiments demonstrating the equivalence between decentralized and centralized training settings for multimodal language models across diverse set of benchmarks. Specifically, we compare two distinct paradigms: LLaVA and InternVL 2.5-1B, which uses a fixed CLIP vision encoder and performs full-parameter fine-tuning (ViT+MLP+LLM) during the instruction tuning stage.

Method: Develops an operator-level analytical framework based on probability-domain softening operators with four core components: bias-variance decompositions, homotopy path formalization of multi-stage pruning, convergence guarantees, and equivalence class characterizations. Introduces axiomatic definition of softening operators based on ranking preservation, continuity, entropy monotonicity, identity, and boundary behavior.

Result: Establishes theoretical guarantees including: operator-agnostic bias-variance decompositions showing when sparse students outperform dense teachers, formal explanation of why iterative pruning succeeds where one-shot fails, O(1/n) convergence rates for n-stage distillation, and identification of distinct probability-domain operators that yield identical student models under capacity constraints.

Conclusion: The framework provides unified theoretical foundations for sparse knowledge distillation, with guarantees that hold uniformly across operator classes, supporting practical applications in black-box distillation, partial-access settings, and privacy-preserving model compression.

Abstract: We develop a unified theoretical framework for sparse knowledge distillation based on probability-domain softening operators. While the equivalence $p^{1/T} \propto \mathrm{softmax}(z/T)$ is well known, our contribution is an operator-level analytical framework built on this foundation rather than the equivalence itself. The framework comprises four core components: (i) operator-agnostic bias–variance decompositions that characterize when sparse students outperform dense teachers, (ii) a homotopy path formalization of multi-stage pruning in function space explaining why iterative compression succeeds where one-shot pruning fails, (iii) convergence guarantees establishing $O(1/n)$ rates for $n$-stage distillation with explicit parameter dependence, and (iv) equivalence class characterizations identifying distinct probability-domain operators that yield identical student models under capacity constraints. We introduce an axiomatic definition of probability-domain softening operators based on ranking preservation, continuity, entropy monotonicity, identity, and boundary behavior, and show that multiple non-equivalent operator families satisfy these axioms. All learning-theoretic guarantees are shown to hold uniformly across this operator class, independent of implementation details. These results provide theoretical grounding for black-box teacher distillation, partial-access settings such as top-$k$ truncation and text-only outputs, and privacy-preserving model compression.

Method: Introduces VC-IFEval benchmark with systematically constructed dataset that incorporates vision-dependent constraints into instruction design. Also fine-tunes MLLMs on the dataset to improve visual instruction-following performance.

Result: Fine-tuning MLLMs on the VC-IFEval dataset achieves substantial gains in visual instruction-following accuracy and adherence. Extensive evaluation across representative MLLMs provides new insights into their strengths and limitations.

Conclusion: VC-IFEval enables more rigorous and fine-grained assessment of MLLMs’ ability to align outputs with both visual input and textual instructions, addressing a critical gap in multimodal instruction-following evaluation.

Abstract: Evaluating the instruction-following (IF) capabilities of Multimodal Large Language Models (MLLMs) is essential for rigorously assessing how faithfully model outputs adhere to user-specified intentions. Nevertheless, existing benchmarks for evaluating MLLMs’ instruction-following capability primarily focus on verbal instructions in the textual modality. These limitations hinder a thorough analysis of instruction-following capabilities, as they overlook the implicit constraints embedded in the semantically rich visual modality. To address this gap, we introduce VC-IFEval, a new benchmark accompanied by a systematically constructed dataset that evaluates MLLMs’ instruction-following ability under multimodal settings. Our benchmark systematically incorporates vision-dependent constraints into instruction design, enabling a more rigorous and fine-grained assessment of how well MLLMs align their outputs with both visual input and textual instructions. Furthermore, by fine-tuning MLLMs on our dataset, we achieve substantial gains in visual instruction-following accuracy and adherence. Through extensive evaluation across representative MLLMs, we provide new insights into the strengths and limitations of current models.

Method: CF-GU bootstraps causal discovery under domain knowledge constraints to generate multiple plausible DAGs, quantifies graph uncertainty using normalized Shannon entropy, and provides confidence bounds on CF metrics.

Result: Experiments show CF-GU can support or refute fairness audits under different domain knowledge assumptions on synthetic data, and successfully identifies known biases in COMPAS and Adult datasets with high confidence even with minimal domain knowledge.

Conclusion: Incorporating causal graph uncertainty into fairness auditing provides more robust and reliable bias evaluations, making CF-GU a valuable tool for trustworthy ML systems.

Abstract: Evaluating machine learning (ML) model bias is key to building trustworthy and robust ML systems. Counterfactual Fairness (CF) audits allow the measurement of bias of ML models with a causal framework, yet their conclusions rely on a single causal graph that is rarely known with certainty in real-world scenarios. We propose CF with Graph Uncertainty (CF-GU), a bias evaluation procedure that incorporates the uncertainty of specifying a causal graph into CF. CF-GU (i) bootstraps a Causal Discovery algorithm under domain knowledge constraints to produce a bag of plausible Directed Acyclic Graphs (DAGs), (ii) quantifies graph uncertainty with the normalized Shannon entropy, and (iii) provides confidence bounds on CF metrics. Experiments on synthetic data show how contrasting domain knowledge assumptions support or refute audits of CF, while experiments on real-world data (COMPAS and Adult datasets) pinpoint well-known biases with high confidence, even when supplied with minimal domain knowledge constraints.

Method: Treats denoising as sequential decision process, introduces timestep-aware critic with noisy-step rewards. Trains CLIP-based reward predictor on noisy latents, uses per-step signal to compute advantage estimates for policy-gradient updates of reverse diffusion kernel.

Result: Achieves better or comparable forgetting to strong baselines across multiple concepts while maintaining image quality and benign prompt fidelity. Ablations show per-step critics and noisy-conditioned rewards are key to stability and effectiveness.

Conclusion: The RL framework is simple to implement, supports off-policy reuse, plugs into standard text-to-image backbones, and provides an effective approach for diffusion unlearning with improved stability and credit assignment.

Abstract: Machine unlearning in text-to-image diffusion models aims to remove targeted concepts while preserving overall utility. Prior diffusion unlearning methods typically rely on supervised weight edits or global penalties; reinforcement-learning (RL) approaches, while flexible, often optimize sparse end-of-trajectory rewards, yielding high-variance updates and weak credit assignment. We present a general RL framework for diffusion unlearning that treats denoising as a sequential decision process and introduces a timestep-aware critic with noisy-step rewards. Concretely, we train a CLIP-based reward predictor on noisy latents and use its per-step signal to compute advantage estimates for policy-gradient updates of the reverse diffusion kernel. Our algorithm is simple to implement, supports off-policy reuse, and plugs into standard text-to-image backbones. Across multiple concepts, the method achieves better or comparable forgetting to strong baselines while maintaining image quality and benign prompt fidelity; ablations show that (i) per-step critics and (ii) noisy-conditioned rewards are key to stability and effectiveness. We release code and evaluation scripts to facilitate reproducibility and future research on RL-based diffusion unlearning.

Method: The authors introduce “epiplexity” - a formalization of information that captures what computationally bounded observers can learn from data. It measures structural content while excluding time-bounded entropy (random unpredictable content like pseudorandom generators). They develop practical procedures to estimate epiplexity and demonstrate its properties through theoretical analysis and empirical validation.

Result: Epiplexity shows that information can be created with computation, depends on data ordering, and likelihood modeling can produce more complex programs than the original data generating process. Practical estimation procedures capture differences across data sources, track with downstream performance, and highlight dataset interventions that improve out-of-distribution generalization.

Conclusion: Epiplexity provides a theoretical foundation for data selection, guiding how to select, generate, or transform data for learning systems. Unlike model selection principles, it focuses on data selection and offers a framework for understanding what computationally bounded systems can actually learn from data.

Abstract: Can we learn more from data than existed in the generating process itself? Can new and useful information be constructed from merely applying deterministic transformations to existing data? Can the learnable content in data be evaluated without considering a downstream task? On these questions, Shannon information and Kolmogorov complexity come up nearly empty-handed, in part because they assume observers with unlimited computational capacity and fail to target the useful information content. In this work, we identify and exemplify three seeming paradoxes in information theory: (1) information cannot be increased by deterministic transformations; (2) information is independent of the order of data; (3) likelihood modeling is merely distribution matching. To shed light on the tension between these results and modern practice, and to quantify the value of data, we introduce epiplexity, a formalization of information capturing what computationally bounded observers can learn from data. Epiplexity captures the structural content in data while excluding time-bounded entropy, the random unpredictable content exemplified by pseudorandom number generators and chaotic dynamical systems. With these concepts, we demonstrate how information can be created with computation, how it depends on the ordering of the data, and how likelihood modeling can produce more complex programs than present in the data generating process itself. We also present practical procedures to estimate epiplexity which we show capture differences across data sources, track with downstream performance, and highlight dataset interventions that improve out-of-distribution generalization. In contrast to principles of model selection, epiplexity provides a theoretical foundation for data selection, guiding how to select, generate, or transform data for learning systems.

Method: Proposes a new optimization problem formulation that includes pre-trained models and random sketch operators for model and gradient sparsification. Develops specialized SGD variants including general sampling, distributed version, and variance-reduced SGD adapted to this formulation.

Result: Establishes insightful properties of the objective function, shows connections to standard formulations, achieves tighter convergence rates, relaxes assumptions, and bridges gap between theory and practical applications like Dropout and sparse training.

Conclusion: The sparsification-aware optimization approach presents promising opportunities to enhance theoretical understanding of model training while maintaining practical relevance to important training techniques.

Abstract: We introduce a novel optimization problem formulation that departs from the conventional way of minimizing machine learning model loss as a black-box function. Unlike traditional formulations, the proposed approach explicitly incorporates an initially pre-trained model and random sketch operators, allowing for sparsification of both the model and gradient during training. We establish the insightful properties of the proposed objective function and highlight its connections to the standard formulation. Furthermore, we present several variants of the Stochastic Gradient Descent (SGD) method adapted to the new problem formulation, including SGD with general sampling, a distributed version, and SGD with variance reduction techniques. We achieve tighter convergence rates and relax assumptions, bridging the gap between theoretical principles and practical applications, covering several important techniques such as Dropout and Sparse training. This work presents promising opportunities to enhance the theoretical understanding of model training through a sparsification-aware optimization approach.

Method: Proposes Time-Transformer AAE with adversarial autoencoder and Time-Transformer decoder. Time-Transformer uses parallel design combining Temporal Convolutional Networks (for local features) and Transformer (for global dependencies) with bidirectional cross attention for feature fusion.

Result: Outperforms state-of-the-art models on 5 out of 6 datasets, especially on data with both global and local properties. Demonstrates effectiveness on artificial dataset and real-world applications like data augmentation for small/imbalanced datasets.

Conclusion: Time-Transformer AAE effectively addresses the challenge of learning both local and global temporal properties in time series generation, showing superior performance and practical utility for data augmentation tasks.

Abstract: Generating time series data is a promising approach to address data deficiency problems. However, it is also challenging due to the complex temporal properties of time series data, including local correlations as well as global dependencies. Most existing generative models have failed to effectively learn both the local and global properties of time series data. To address this open problem, we propose a novel time series generative model named ‘Time-Transformer AAE’, which consists of an adversarial autoencoder (AAE) and a newly designed architecture named ‘Time-Transformer’ within the decoder. The Time-Transformer first simultaneously learns local and global features in a layer-wise parallel design, combining the abilities of Temporal Convolutional Networks and Transformer in extracting local features and global dependencies respectively. Second, a bidirectional cross attention is proposed to provide complementary guidance across the two branches and achieve proper fusion between local and global features. Experimental results demonstrate that our model can outperform existing state-of-the-art models in 5 out of 6 datasets, specifically on those with data containing both global and local properties. Furthermore, we highlight our model’s advantage on handling this kind of data via an artificial dataset. Finally, we show our model’s ability to address a real-world problem: data augmentation to support learning with small datasets and imbalanced datasets.

Method: The study uses separate models for dwell time and running time predictions, employing XGBoost and graph neural networks (GNN) to leverage spatial correlations. A hierarchical model combining random forest classifier and GNN is proposed to handle shuttle bypass scenarios.

Result: Results show promising low errors even when predicting several stops ahead, though no single model is universally superior. Dwell time prediction is identified as the key determinant of overall accuracy in low-traffic areas or under speed limits.

Conclusion: The meta-analysis across six pilot sites provides insights into current autonomous public transport prediction models and enables more data-informed decision-making as the field advances, with dwell time prediction being particularly critical in certain operational conditions.

Abstract: Urban mobility is on the cusp of transformation with the emergence of shared, connected, and cooperative automated vehicles. Yet, for them to be accepted by customers, trust in their punctuality is vital. Many pilot initiatives operate without a fixed schedule, enhancing the importance of reliable arrival time (AT) predictions. This study presents an AT prediction system for automated shuttles, utilizing separate models for dwell and running time predictions, validated on real-world data from six cities. Alongside established methods such as XGBoost, we explore the benefits of leveraging spatial correlations using graph neural networks (GNN). To accurately handle the case of a shuttle bypassing a stop, we propose a hierarchical model combining a random forest classifier and a GNN. The results for the final AT prediction are promising, showing low errors even when predicting several stops ahead. Yet, no single model emerges as universally superior, and we provide insights into the characteristics of pilot sites that influence the model selection process and prediction performance. Finally, we identify dwell time prediction as the key determinant in overall AT prediction accuracy when automated shuttles are deployed in low-traffic areas or under regulatory speed limits. Our meta-analysis across six pilot sites in different cities provides insights into the current state of autonomous public transport prediction models and paves the way for more data-informed decision-making as the field advances.

Method: Proposes Reliability-based Curriculum Learning (RCL) that distills robust supervision from multiple frozen MLLMs into a compact target model. Uses three-stage curriculum learning that progressively incorporates pseudo-labels based on inter-model agreement and model confidence for stable, noise-aware training.

Result: Achieves state-of-the-art performance on standard SFDA datasets (Office-Home, DomainNet-126, VisDA-C), outperforming zero-shot MLLMs and their ensembles, without accessing source data or tuning foundation models.

Conclusion: RCL effectively leverages multiple MLLMs’ knowledge through curriculum learning to achieve robust SFDA performance while maintaining efficiency by using frozen models and avoiding source data access.

Abstract: Existing SFDA methods struggle to fully use pre-trained knowledge and often rely on a single model’s predictions or handcrafted prompts, limiting robustness under domain shift. Multimodal Large Language Models (MLLMs) offer a promising alternative: they encode rich visual-semantic knowledge and generalize well without task-specific tuning. However, their use in SFDA is hindered by instruction-following failures, inconsistent outputs, and high inference costs. We propose Reliability-based Curriculum Learning (RCL), a novel framework that distills robust supervision from multiple frozen MLLMs into a compact target model. RCL organizes adaptation as a three-stage curriculum that progressively incorporates pseudo-labels based on inter-model agreement and model confidence, enabling stable and noise-aware training. Our approach achieves state-of-the-art performance on standard SFDA datasets, Office-Home, DomainNet-126, and VisDA-C, outperforming zero-shot MLLMs, their ensembles, all without accessing source data or tuning foundation models. Our code is available at: https://github.com/Dong-Jie-Chen/RCL.

Method: Frames causal dose-response problem as covariate shift, uses weighted conformal prediction with propensity estimation, conformal predictive systems, and likelihood ratios. Incorporates kernel functions as weights to approximate local coverage for every treatment value.

Result: Demonstrated using synthetic benchmark dataset showing significance of covariate shift assumptions for achieving robust prediction intervals in dose-response models.

Conclusion: Proposed methodology provides practical solution for generating prediction intervals in dose-response models, addressing the gap in applying conformal prediction to continuous treatments with proper uncertainty quantification.

Abstract: Understanding the dose-response relation between a continuous treatment and the outcome for an individual can greatly drive decision-making, particularly in areas like personalized drug dosing and personalized healthcare interventions. Point estimates are often insufficient in these high-risk environments, highlighting the need for uncertainty quantification to support informed decisions. Conformal prediction, a distribution-free and model-agnostic method for uncertainty quantification, has seen limited application in continuous treatments or dose-response models. To address this gap, we propose a novel methodology that frames the causal dose-response problem as a covariate shift, leveraging weighted conformal prediction. By incorporating propensity estimation, conformal predictive systems, and likelihood ratios, we present a practical solution for generating prediction intervals for dose-response models. Additionally, our method approximates local coverage for every treatment value by applying kernel functions as weights in weighted conformal prediction. Finally, we use a new synthetic benchmark dataset to demonstrate the significance of covariate shift assumptions in achieving robust prediction intervals for dose-response models.

Method: Theoretical analysis of debiasing methods for model evaluation using LLM-as-a-judge paradigm. Main result proves mathematical limit: when judge accuracy ≤ evaluated model accuracy, no debiasing method can reduce required ground truth labels by more than half. Also includes empirical evaluation to compare practical savings against theoretical limit.

Result: Theoretical limit shows maximum possible reduction in ground truth labels is 50% when judge is no more accurate than evaluated model. Empirical results show practical savings are even more modest than this theoretical limit. Findings reveal severe limitations of LLM-as-a-judge paradigm for evaluating frontier models that may be better than the judge.

Conclusion: LLM-as-a-judge evaluation has fundamental limitations, especially for assessing newly released models that could outperform the judge. Debiasing methods offer only modest improvements at best. The work highlights the need for new approaches to scalable model evaluation and identifies promising research directions.

Abstract: High quality annotations are increasingly a bottleneck in the explosively growing machine learning ecosystem. Scalable evaluation methods that avoid costly annotation have therefore become an important research ambition. Many hope to use strong existing models in lieu of costly labels to provide cheap model evaluations. Unfortunately, this method of using models as judges introduces biases, such as self-preferencing, that can distort model comparisons. An emerging family of debiasing tools promises to fix these issues by using a few high quality labels to debias a large number of model judgments. In this paper, we study how far such debiasing methods, in principle, can go. Our main result shows that when the judge is no more accurate than the evaluated model, no debiasing method can decrease the required amount of ground truth labels by more than half. Our result speaks to the severe limitations of the LLM-as-a-judge paradigm at the evaluation frontier where the goal is to assess newly released models that are possibly better than the judge. Through an empirical evaluation, we demonstrate that the sample size savings achievable in practice are even more modest than what our theoretical limit suggests. Along the way, our work provides new observations about debiasing methods for model evaluation, and points out promising avenues for future work.

Method: Uses fine-grained quantization techniques to compress selected activations in Tensor Parallel deployments, specifically targeting inter-accelerator communication. The compression reduces data transfer volume between accelerators by 3.5-4.5x.

Result: Achieves up to 2x reduction in time-to-first-token (TTFT) with negligible degradation in model performance. The activation compression effectively reduces communication bottlenecks in distributed LLM inference.

Conclusion: Fine-grained quantization for compressing inter-accelerator communication in Tensor Parallel deployments is an effective approach to significantly reduce inference latency without compromising model quality, addressing a key bottleneck in distributed LLM serving.

Abstract: Large Language Models (LLMs) have pushed the frontier of artificial intelligence but are comprised of hundreds of billions of parameters and operations. For faster inference latency, LLMs are deployed on multiple hardware accelerators through various Model Parallelism strategies. Our paper looks into the details on one such strategy - Tensor Parallel - and proposes to reduce latency by compressing inter-accelerator communication. We leverage fine grained quantization techniques to compress selected activations by 3.5 - 4.5x. Our proposed method leads up to 2x reduction of time-to-first-token (TTFT) with negligible model performance degradation.

Method: Restructures the computational graph by grouping and evaluating consecutive layer pairs in parallel. This approach requires no retraining and can be combined with lightweight fine-tuning of the parallelized layers to recover some lost accuracy.

Result: Achieves 1.19x throughput gain on Llama 2 7B while reducing average benchmark accuracy by only 1.5%. Demonstrates practical value for large-scale LLM deployment and shows lost accuracy can be partially recovered with lightweight fine-tuning.

Conclusion: The method successfully bridges the gap between theoretical insights about LLM layer redundancy and practical inference speedups, offering a deployable solution for accelerating LLMs without significant accuracy degradation.

Abstract: The remarkable capabilities of Large Language Models (LLMs) are overshadowed by their immense computational cost. While recent work has shown that many LLM layers can be reordered or even removed with minimal impact on accuracy, these insights have not been translated into significant inference speedups. To bridge this gap, we introduce a novel method that restructures the computational graph by grouping and evaluating consecutive layer pairs in parallel. This approach, requiring no retraining, yields a 1.19x throughput gain on Llama 2 7B while reducing the average benchmark accuracy by only 1.5%. We demonstrate the practical value of this method for large-scale LLM deployment and show that some of the lost accuracy can be recovered with lightweight fine-tuning of the parallelized layers.

Method: Developed a constraint programming model that leverages knowledge of forest structure and DP mechanism characteristics to formally reconstruct the most likely dataset that could have produced a given DP random forest.

Result: DP reduces reconstruction attack success but doesn’t eliminate it; only forests with predictive performance no better than a constant classifier are fully robust to reconstruction attacks.

Conclusion: DP random forests can still leak training data despite privacy guarantees, and practical recommendations are provided for constructing more resilient DP random forests while maintaining non-trivial predictive performance.

Abstract: Recent research has shown that structured machine learning models such as tree ensembles are vulnerable to privacy attacks targeting their training data. To mitigate these risks, differential privacy (DP) has become a widely adopted countermeasure, as it offers rigorous privacy protection. In this paper, we introduce a reconstruction attack targeting state-of-the-art $ε$-DP random forests. By leveraging a constraint programming model that incorporates knowledge of the forest’s structure and DP mechanism characteristics, our approach formally reconstructs the most likely dataset that could have produced a given forest. Through extensive computational experiments, we examine the interplay between model utility, privacy guarantees and reconstruction accuracy across various configurations. Our results reveal that random forests trained with meaningful DP guarantees can still leak portions of their training data. Specifically, while DP reduces the success of reconstruction attacks, the only forests fully robust to our attack exhibit predictive performance no better than a constant classifier. Building on these insights, we also provide practical recommendations for the construction of DP random forests that are more resilient to reconstruction attacks while maintaining a non-trivial predictive performance.

Method: A lightweight predictive UQ method tailored for DeepONets that generalizes to other operator networks. Includes an inference strategy using precomputed trunk outputs and sparse placement matrix to reduce evaluation time by 5x+. Extends framework to Fourier Neural Operators (FNO) and describes generalized method for other operator networks.

Result: Numerical experiments on linear and nonlinear PDEs show the framework’s uncertainty estimates are unbiased and provide accurate out-of-distribution uncertainty predictions with sufficiently large training datasets. Enables fast inference and uncertainty estimates for efficient outer-loop analyses.

Conclusion: The method provides a practical route to uncertainty-aware operator learning in time-sensitive settings, demonstrating applications in Bayesian optimization and active learning to improve accuracy and data-efficiency for outer-loop optimization procedures.

Abstract: With the increased prevalence of neural operators being used to provide rapid solutions to partial differential equations (PDEs), understanding the accuracy of model predictions and the associated error levels is necessary for deploying reliable surrogate models in scientific applications. Existing uncertainty quantification (UQ) frameworks employ ensembles or Bayesian methods, which can incur substantial computational costs during both training and inference. We propose a lightweight predictive UQ method tailored for Deep operator networks (DeepONets) that also generalizes to other operator networks. Numerical experiments on linear and nonlinear PDEs demonstrate that the framework’s uncertainty estimates are unbiased and provide accurate out-of-distribution uncertainty predictions with a sufficiently large training dataset. Our framework provides fast inference and uncertainty estimates that can efficiently drive outer-loop analyses that would be prohibitively expensive with conventional solvers. We demonstrate how predictive uncertainties can be used in the context of Bayesian optimization and active learning problems to yield improvements in accuracy and data-efficiency for outer-loop optimization procedures. In the active learning setup, we extend the framework to Fourier Neural Operators (FNO) and describe a generalized method for other operator networks. To enable real-time deployment, we introduce an inference strategy based on precomputed trunk outputs and a sparse placement matrix, reducing evaluation time by more than a factor of five. Our method provides a practical route to uncertainty-aware operator learning in time-sensitive settings.

Method: The authors created a large-scale human-collected dataset specifically designed to measure lying behavior in LLMs. They used this benchmark to evaluate various LLMs, separating accuracy (correctness of beliefs) from honesty (truth-telling behavior). They also tested representation engineering interventions to improve honesty.

Result: Larger models show higher accuracy but not higher honesty. Most frontier LLMs score well on truthfulness benchmarks but exhibit significant lying under pressure, resulting in low honesty scores. Simple interventions like representation engineering can improve honesty.

Conclusion: There’s a critical need for robust evaluations that directly measure lying behavior and effective interventions to ensure LLM trustworthiness, as accuracy improvements don’t guarantee honesty and current truthfulness benchmarks may be misleading.

Abstract: As large language models (LLMs) become more capable and agentic, the requirement for trust in their outputs grows significantly, yet at the same time concerns have been mounting that models may learn to lie in pursuit of their goals. To address these concerns, a body of work has emerged around the notion of “honesty” in LLMs, along with interventions aimed at mitigating deceptive behaviors. However, some benchmarks claiming to measure honesty in fact simply measure accuracy–the correctness of a model’s beliefs–in disguise. Moreover, no benchmarks currently exist for directly measuring whether language models lie. In this work, we introduce a large-scale human-collected dataset for directly measuring lying, allowing us to disentangle accuracy from honesty. Across a diverse set of LLMs, we find that while larger models obtain higher accuracy on our benchmark, they do not become more honest. Surprisingly, most frontier LLMs obtain high scores on truthfulness benchmarks yet exhibit a substantial propensity to lie under pressure, resulting in low honesty scores on our benchmark. We find that simple methods, such as representation engineering interventions, can improve honesty. These results underscore the growing need for robust evaluations and effective interventions to ensure LLMs remain trustworthy.

Method: 1) Proposes Contextual Stress Framework (CSF) defining toxicity as stress-inducing norm violation within context; 2) Introduces PONOS metric quantifying toxicity through proportion of negative observed sentiments in social reception; 3) Validates approach on novel dataset.

Result: The approach demonstrates improved contextual sensitivity and adaptability when used alongside existing toxicity detection models, showing better alignment with how toxicity actually manifests in social contexts.

Conclusion: Toxicity should be understood as context-dependent social phenomenon rather than text property. The CSF framework and PONOS metric offer more nuanced, socially-grounded approach to toxicity detection that accounts for community norms and collective reception.

Abstract: Most toxicity detection models treat toxicity as an intrinsic property of text, overlooking the role of context in shaping its impact. In this position paper, drawing on insights from psychology, neuroscience, and computational social science, we reconceptualise toxicity as a socially emergent signal of stress. We formalise this perspective in the Contextual Stress Framework (CSF), which defines toxicity as a stress-inducing norm violation within a given context and introduces an additional dimension for toxicity detection. As one possible realisation of CSF, we introduce PONOS (Proportion Of Negative Observed Sentiments), a metric that quantifies toxicity through collective social reception rather than lexical features. We validate this approach on a novel dataset, demonstrating improved contextual sensitivity and adaptability when used alongside existing models.

Method: Presents first thorough review of offline MBO: formalizes problem for single/multi-objective settings, reviews benchmarks/evaluation metrics, categorizes approaches into surrogate modeling (accurate OOD function approximation) and generative modeling (exploring high-dimensional design spaces).

Result: Comprehensive review organizes the field, identifies key challenges, and proposes future directions including safe control of superintelligent systems.

Conclusion: Offline MBO is rapidly evolving with significant potential for accelerating scientific discovery; the review bridges the gap in literature and provides roadmap for advancing the field while addressing critical challenges like reward hacking and safe AI control.

Abstract: Offline optimization is a fundamental challenge in science and engineering, where the goal is to optimize black-box functions using only offline datasets. This setting is particularly relevant when querying the objective function is prohibitively expensive or infeasible, with applications spanning protein engineering, material discovery, neural architecture search, and beyond. The main difficulty lies in accurately estimating the objective landscape beyond the available data, where extrapolations are fraught with significant epistemic uncertainty. This uncertainty can lead to objective hacking(reward hacking), exploiting model inaccuracies in unseen regions, or other spurious optimizations that yield misleadingly high performance estimates outside the training distribution. Recent advances in model-based optimization(MBO) have harnessed the generalization capabilities of deep neural networks to develop offline-specific surrogate and generative models. Trained with carefully designed strategies, these models are more robust against out-of-distribution issues, facilitating the discovery of improved designs. Despite its growing impact in accelerating scientific discovery, the field lacks a comprehensive review. To bridge this gap, we present the first thorough review of offline MBO. We begin by formalizing the problem for both single-objective and multi-objective settings and by reviewing recent benchmarks and evaluation metrics. We then categorize existing approaches into two key areas: surrogate modeling, which emphasizes accurate function approximation in out-of-distribution regions, and generative modeling, which explores high-dimensional design spaces to identify high-performing designs. Finally, we examine the key challenges and propose promising directions for advancement in this rapidly evolving field including safe control of superintelligent systems.

Method: Proposed reinforcement learning approach with action masking, incorporating proven optimal greedy retrieval strategy for the flexible problem variant where store/retrieve operations can be performed in arbitrary order.

Result: Evaluation on 170 problem instances (2-8 buffer lanes, 5-15 colors) shows significant reduction in color changes compared to existing methods, with robustness across different buffer sizes and imbalanced color distributions.

Conclusion: The reinforcement learning approach with optimal greedy retrieval effectively minimizes color changes in flexible paint shop sequencing problems, outperforming existing methods.

Abstract: In the paint shop problem, an unordered incoming sequence of cars assigned to different colors has to be reshuffled with the objective of minimizing the number of color changes. To reshuffle the incoming sequence, manufacturers can employ a first-in-first-out multi-lane buffer system allowing store and retrieve operations. So far, prior studies primarily focused on simple decision heuristics like greedy or simplified problem variants that do not allow full flexibility when performing store and retrieve operations. In this study, we propose a reinforcement learning approach to minimize color changes for the flexible problem variant, where store and retrieve operations can be performed in an arbitrary order. After proving that greedy retrieval is optimal, we incorporate this finding into the model using action masking. Our evaluation, based on 170 problem instances with 2-8 buffer lanes and 5-15 colors, shows that our approach reduces color changes compared to existing methods by considerable margins depending on the problem size. Furthermore, we demonstrate the robustness of our approach towards different buffer sizes and imbalanced color distributions.

Method: Proposes Ensemble Heuristic-Distilled MLPs (EHDM) which eliminates graph dependencies while effectively integrating complementary signals via a gating mechanism. Uses heuristic methods as teachers for MLP distillation.

Result: Experiments on ten datasets show average 7.93% improvement over previous GNN-to-MLP approaches with 1.95-3.32 times less training time.

Conclusion: EHDM is an efficient and effective link prediction method that achieves strong performance without graph dependencies, demonstrating that weaker heuristic teachers can produce strong MLP students with much lower computational costs.

Abstract: Link prediction is a crucial graph-learning task with applications including citation prediction and product recommendation. Distilling Graph Neural Networks (GNNs) teachers into Multi-Layer Perceptrons (MLPs) students has emerged as an effective approach to achieve strong performance and reducing computational cost by removing graph dependency. However, existing distillation methods only use standard GNNs and overlook alternative teachers such as specialized model for link prediction (GNN4LP) and heuristic methods (e.g., common neighbors). This paper first explores the impact of different teachers in GNN-to-MLP distillation. Surprisingly, we find that stronger teachers do not always produce stronger students: MLPs distilled from GNN4LP can underperform those distilled from simpler GNNs, while weaker heuristic methods can teach MLPs to near-GNN performance with drastically reduced training costs. Building on these insights, we propose Ensemble Heuristic-Distilled MLPs (EHDM), which eliminates graph dependencies while effectively integrating complementary signals via a gating mechanism. Experiments on ten datasets show an average 7.93% improvement over previous GNN-to-MLP approaches with 1.95-3.32 times less training time, indicating EHDM is an efficient and effective link prediction method.

Method: DisCO replaces group relative objectives with discriminative scoring functions, uses non-clipping RL surrogates instead of clipping-based ones, and employs constrained optimization with KL divergence constraints.

Result: DisCO significantly outperforms GRPO and its variants (like DAPO), achieving average gains of 7% over GRPO and 6% over DAPO across six benchmark tasks for a 1.5B model.

Conclusion: DisCO offers superior performance by eliminating difficulty bias, improving training stability, and enabling advanced discriminative techniques to handle data imbalance in large reasoning model reinforcement.

Abstract: The recent success and openness of DeepSeek-R1 have brought widespread attention to Group Relative Policy Optimization (GRPO) as a reinforcement learning method for large reasoning models (LRMs). In this work, we analyze the GRPO objective under a binary reward setting and reveal an inherent limitation of question-level difficulty bias. We also identify a connection between GRPO and traditional discriminative methods in supervised learning. Motivated by these insights, we introduce a new Discriminative Constrained Optimization (DisCO) framework for reinforcing LRMs, grounded in the principle of discriminative learning. The main differences between DisCO and GRPO and its recent variants are: (1) it replaces the group relative objective with a discriminative objective defined by a scoring function; (2) it abandons clipping-based surrogates in favor of non-clipping RL surrogate objectives used as scoring functions; (3) it employs a simple yet effective constrained optimization approach to enforce the KL divergence constraint. As a result, DisCO offers notable advantages over GRPO and its variants: (i) it completely eliminates difficulty bias by adopting discriminative objectives; (ii) it addresses the entropy instability in GRPO and its variants through the use of non-clipping scoring functions and a constrained optimization approach, yielding long and stable training dynamics; (iii) it allows the incorporation of advanced discriminative learning techniques to address data imbalance, where a significant number of questions have more negative than positive generated answers during training. Our experiments on enhancing the mathematical reasoning capabilities of SFT-finetuned models show that DisCO significantly outperforms GRPO and its improved variants such as DAPO, achieving average gains of 7% over GRPO and 6% over DAPO across six benchmark tasks for a 1.5B model.

Method: DC-Clustering enables clustering over complex data partitioning scenarios by having each institution share only intermediate representations instead of raw data. The method supports flexible selection between k-means and spectral clustering algorithms and achieves final results with just a single round of communication with a central server.

Result: Extensive experiments on synthetic and open benchmark datasets show that DC-Clustering achieves clustering performance comparable to centralized clustering where all data are pooled together. The method effectively bridges the performance gap between federated and centralized approaches.

Conclusion: DC-Clustering addresses an important gap in federated learning research by enabling effective knowledge discovery from distributed heterogeneous data with complex partitioning. Its practical properties - privacy preservation, communication efficiency, and algorithmic flexibility - make it promising for privacy-sensitive domains like healthcare and finance.

Abstract: In recent years, the growing need to leverage sensitive data across institutions has led to increased attention on federated learning (FL), a decentralized machine learning paradigm that enables model training without sharing raw data. However, existing FL-based clustering methods, known as federated clustering, typically assume simple data partitioning scenarios such as horizontal or vertical splits, and cannot handle more complex distributed structures. This study proposes data collaboration clustering (DC-Clustering), a novel federated clustering method that supports clustering over complex data partitioning scenarios where horizontal and vertical splits coexist. In DC-Clustering, each institution shares only intermediate representations instead of raw data, ensuring privacy preservation while enabling collaborative clustering. The method allows flexible selection between k-means and spectral clustering, and achieves final results with a single round of communication with the central server. We conducted extensive experiments using synthetic and open benchmark datasets. The results show that our method achieves clustering performance comparable to centralized clustering where all data are pooled. DC-Clustering addresses an important gap in current FL research by enabling effective knowledge discovery from distributed heterogeneous data. Its practical properties – privacy preservation, communication efficiency, and flexibility – make it a promising tool for privacy-sensitive domains such as healthcare and finance.

Method: Extensive evaluation comparing different inference-time strategies on VLMs, including: 1) comparing generation-time capabilities (majority voting) vs verification-centric strategies (best of N with self-verification), 2) examining behaviors associated with RL-tuned models like the ‘Aha moment’, and 3) analyzing visual information integration in self-verification processes.

Result: Three key findings: 1) Simple majority voting consistently and substantially outperforms verification-centric strategies for VLMs, 2) RL-tuned model behaviors like the ‘Aha moment’ don’t yield reliable reasoning improvements, 3) Visual information is not effectively integrated into VLMs’ self-verification process, limiting inference-time scaling benefits.

Conclusion: Current RL-trained VLMs derive limited benefit from self-verification in the visual modality, constraining the effectiveness of inference-time scaling techniques for visual mathematical reasoning, highlighting a key limitation in current VLM architectures and training approaches.

Abstract: Inference time techniques such as decoding time scaling and self refinement have been shown to substantially improve mathematical reasoning in large language models (LLMs), largely attributed to emergent self correction and self verification behaviors often elicited through reinforcement learning (RL). In this work, we ask whether the same recipe transfers to vision language models (VLMs), especially RL finetuned variants that claim strong visual mathematical reasoning. Through extensive evaluation, we reach three main findings that differ markedly from text only models. First, generation time capability matters more than verification and refinement: simple majority voting consistently and substantially outperforms verification centric strategies such as best of N with self verification. Second, behaviors often associated with RL tuned models at inference time, such as the ‘Aha moment,’ do not yield reliable reasoning performance improvements. Third, visual information is not effectively integrated into the model’s self verification process. Overall, our analysis highlights a key limitation: current RL trained VLMs derive limited benefit from self verification in the visual modality, which constrains the effectiveness of inference time scaling for visual mathematical reasoning.

Method: Created through a multi-step curation pipeline using over 35,000 real user-expert interactions from agriculture domain. Combines natural user queries, expert-authored responses, and image-based context to create diverse scenarios covering crop health, pest diagnosis, and management.

Result: MIRAGE includes more than 7,000 unique biological entities (plant species, pests, diseases) making it one of the most taxonomically diverse vision-language benchmarks. Features underspecified, context-rich scenarios requiring models to infer knowledge gaps and handle rare entities.

Conclusion: MIRAGE provides a high-fidelity benchmark for evaluating multimodal models on grounded reasoning, clarification strategies, and long-form generation in real-world, knowledge-intensive domains like agriculture, addressing limitations of existing benchmarks.

Abstract: We introduce MIRAGE, a new benchmark for multimodal expert-level reasoning and decision-making in consultative interaction settings. Designed for the agriculture domain, MIRAGE captures the full complexity of expert consultations by combining natural user queries, expert-authored responses, and image-based context, offering a high-fidelity benchmark for evaluating models on grounded reasoning, clarification strategies, and long-form generation in a real-world, knowledge-intensive domain. Grounded in over 35,000 real user-expert interactions and curated through a carefully designed multi-step pipeline, MIRAGE spans diverse crop health, pest diagnosis, and crop management scenarios. The benchmark includes more than 7,000 unique biological entities, covering plant species, pests, and diseases, making it one of the most taxonomically diverse benchmarks available for vision-language models, grounded in the real world. Unlike existing benchmarks that rely on well-specified user inputs and closed-set taxonomies, MIRAGE features underspecified, context-rich scenarios with open-world settings, requiring models to infer latent knowledge gaps, handle rare entities, and either proactively guide the interaction or respond. Project Page: https://mirage-benchmark.github.io

Method: The authors develop a unified theoretical framework for replay-based continual learning, deriving both hypothesis-based and prediction-based information-theoretic generalization bounds. The hypothesis-based bounds capture trade-offs between memory buffer size and information dependency, while prediction-based bounds provide tighter, computationally tractable bounds using low-dimensional variables. The framework is general and applicable to various learning algorithms, with SGLD used as a representative example.

Result: The paper establishes novel generalization bounds that explicitly show how memory buffers and current tasks affect generalization performance. The bounds capture the trade-off between the number of stored exemplars and information dependency, with prediction-based bounds providing tighter, tractable upper bounds. Comprehensive experiments demonstrate these bounds effectively capture generalization dynamics in replay-based continual learning settings.

Conclusion: The paper provides a solid theoretical foundation for understanding replay-based continual learning, offering practical bounds that can guide algorithm design and analysis. The information-theoretic framework bridges the gap between empirical performance and theoretical understanding in continual learning research.

Abstract: Continual learning (CL) has emerged as a dominant paradigm for acquiring knowledge from sequential tasks while avoiding catastrophic forgetting. Although many CL methods have been proposed to show impressive empirical performance, the theoretical understanding of their generalization behavior remains limited, particularly for replay-based approaches. This paper establishes a unified theoretical framework for replay-based CL, deriving a series of information-theoretic generalization bounds that explicitly elucidate the impact of the memory buffer alongside the current task on generalization performance. Specifically, our hypothesis-based bounds capture the trade-off between the number of selected exemplars and the information dependency between the hypothesis and the memory buffer. Our prediction-based bounds yield tighter and computationally tractable upper bounds on the generalization error by leveraging low-dimensional variables. Theoretical analysis is general and broadly applicable to a wide range of learning algorithms, exemplified by stochastic gradient Langevin dynamics (SGLD) as a representative method. Comprehensive experimental evaluations demonstrate the effectiveness of our derived bounds in capturing the generalization dynamics in replay-based CL settings.

Method: U-shaped Physics-Informed Network (U-PINet) with hierarchical operator design inspired by near-far field decomposition. Uses physics-guided graph construction to model EM coupling among mesh elements and embeds EM governing equations as residual constraints for end-to-end physically consistent prediction.

Result: Achieves solver-level RCS accuracy with orders-of-magnitude runtime reduction. Exhibits strong generalization to unseen target geometries even with limited training data.

Conclusion: U-PINet successfully bridges the gap between accurate but slow CEM solvers and fast but physically inconsistent neural networks, providing a novel physics-informed framework for efficient RCS computation with strong generalization capabilities.

Abstract: Accurate radar cross section (RCS) computation is a fundamental task in radar engineering and electromagnetic (EM) scattering analysis, underpinning target signature characterization, detection, and recognition. Conventional computational electromagnetics (CEM) solvers provide high-fidelity RCS predictions but suffer from prohibitive computational costs when applied to 3-dimensional (3D) targets under multi-aspect configurations. In contrast, purely data-driven neural networks offer high efficiency yet often lack physical consistency and generalization capability. To address these challenges, this paper proposes a U-shaped Physics-Informed Network (U-PINet). To the best of our knowledge, it is the first framework to establish a fully end-to-end, physics-informed hierarchical architecture for fast and accurate RCS computation, grounded in the governing principles of CEM. Inspired by the near-far field decomposition in classical fast solvers, U-PINet explicitly models local EM coupling and long-range radiation effects through a hierarchical operator design. A physics-guided graph construction is further introduced to represent self- and mutual-coupling among mesh elements of complex 3D targets, enabling physically interpretable intermediate representations. By embedding EM governing equations as residual constraints, the proposed framework achieves end-to-end, physically consistent RCS prediction with significantly improved computational efficiency. Extensive numerical experiments demonstrate that U-PINet attains solver-level RCS accuracy with orders-of-magnitude runtime reduction, while exhibiting strong generalization to unseen target geometries under limited training data.

Method: IPA uses feature-aware projection framework that explicitly reconstructs original input in reduced hidden space, approximating top principal components for efficient pretraining with minimal inference overhead.

Result: IPA consistently outperforms LoRA and DoRA across language/vision benchmarks: +1.5 avg accuracy on commonsense reasoning, +2.3 on VTAB-1k, matches full LoRA performance with ~50% trainable parameters when projection frozen.

Conclusion: Feature-aware projection effectively addresses LoRA’s random initialization bottleneck, improving parameter efficiency and performance across domains with minimal overhead.

Abstract: Parameter-efficient fine-tuning (PEFT) methods, such as LoRA, reduce adaptation cost by injecting low-rank updates into pretrained weights. However, LoRA’s down-projection is randomly initialized and data-agnostic, discarding potentially useful information. Prior analyses show that this projection changes little during training, while the up-projection carries most of the adaptation, making the random input compression a performance bottleneck. We propose IPA, a feature-aware projection framework that explicitly aims to reconstruct the original input within a reduced hidden space. In the linear case, we instantiate IPA with algorithms approximating top principal components, enabling efficient projector pretraining with negligible inference overhead. Across language and vision benchmarks, IPA consistently improves over LoRA and DoRA, achieving on average 1.5 points higher accuracy on commonsense reasoning and 2.3 points on VTAB-1k, while matching full LoRA performance with roughly half the trainable parameters when the projection is frozen. Code available at https://github.com/valeoai/peft-ipa .

Method: Constructed a formal POMDP model for CAGE-2, then developed BF-PPO method based on PPO (Proximal Policy Optimization) with particle filter to handle computational complexity from large state space.

Result: BF-PPO outperformed CARDIFF (the highest ranked method on CAGE-2 leaderboard) in both learned defender strategy quality and required training time.

Conclusion: The POMDP-based BF-PPO method with particle filtering provides an effective and efficient solution for learning optimal defender strategies in CAGE-2 cybersecurity scenarios, surpassing current state-of-the-art approaches.

Abstract: CAGE-2 is an accepted benchmark for learning and evaluating defender strategies against cyberattacks. It reflects a scenario where a defender agent protects an IT infrastructure against various attacks. Many defender methods for CAGE-2 have been proposed in the literature. In this paper, we construct a formal model for CAGE-2 using the framework of Partially Observable Markov Decision Process (POMDP). Based on this model, we define an optimal defender strategy for CAGE-2 and introduce a method to efficiently learn this strategy. Our method, called BF-PPO, is based on PPO, and it uses particle filter to mitigate the computational complexity due to the large state space of the CAGE-2 model. We evaluate our method in the CAGE-2 CybORG environment and compare its performance with that of CARDIFF, the highest ranked method on the CAGE-2 leaderboard. We find that our method outperforms CARDIFF regarding the learned defender strategy and the required training time.

Method: 1) Created CaTS-Bench with 1746 human-rewritten captions across 11 domains for gold-standard evaluation; 2) Developed scalable pipeline for generating high-fidelity synthetic captions to address data scarcity; 3) Evaluated leading Vision-Language Models; 4) Released diagnostic suite with 910 multiple-choice questions and tailored numeric metrics.

Result: Proprietary models struggle with numeric nuances in temporal descriptions, while finetuning open-source models on synthetic data yields substantial performance gains. The synthetic caption generation pipeline produces high-quality data validated for fidelity.

Conclusion: CaTS-Bench establishes a reliable foundation for grounded, multimodal language generation in numeric domains, providing comprehensive evaluation of time-series-specific reasoning capabilities through human-annotated data, synthetic generation, and diagnostic tools.

Abstract: Time series captioning, the task of describing time series in natural language, requires numeric and temporal reasoning, trend interpretation, and contextual understanding. Existing benchmarks, however, often rely on fully synthetic or generic captions, and typically neglect metadata and visual representations. We introduce \textbf{CaTS-Bench}, a comprehensive benchmark for \textbf{C}ontext-\textbf{a}ware \textbf{T}ime \textbf{S}eries reasoning across $11$ diverse domains, centered on a gold-standard evaluation set of $1746$ human-rewritten captions that measure how effectively models translate numeric trends into immediately interpretable narratives. To address the scarcity of human-annotated data, we also propose a scalable pipeline for generating high-fidelity synthetic captions, the quality of which we validate. We evaluate leading Vision-Language Models on our benchmark, revealing that even proprietary models struggle to capture numeric nuances in temporal descriptions, while finetuning open-source models on synthetic data yields substantial performance gains. Finally, release a diagnostic suite of $910$ multiple-choice questions and tailored numeric metrics to gauge time-series-specific reasoning capabilities, establishing CaTS-Bench as a reliable foundation for grounded, multimodal language generation in numeric domains.

Method: The authors introduce a general framework that reduces the constrained learning problem to standard OCO with specially constructed surrogate cost functions. They present two algorithms: 1) an optimal-regret algorithm using projection onto constraint sets, and 2) a projection-free algorithm that achieves better violation bounds in rapidly varying environments.

Result: The algorithms achieve universal dynamic regret and cumulative constraint violation bounds that improve upon state-of-the-art results. The results hold for arbitrary cost and constraint functions without requiring any fixed common feasible point.

Conclusion: The paper successfully generalizes OCO to handle adversarial constraints with improved theoretical guarantees, offering both projection-based and projection-free solutions that work in the most general setting without common feasibility assumptions.

Abstract: We consider a generalization of the celebrated Online Convex Optimization (OCO) framework with adversarial online constraints. In this problem, an online learner interacts with an adversary sequentially over multiple rounds. At the beginning of each round, the learner chooses an action from a convex decision set. After that, the adversary reveals a convex cost function and a convex constraint function. The goal of the learner is to minimize the cumulative cost while satisfying the constraints as tightly as possible. We present two efficient algorithms with simple modular structures that give universal dynamic regret and cumulative constraint violation bounds, improving upon state-of-the-art results. While the first algorithm, which achieves the optimal regret bound, involves projection onto the constraint sets, the second algorithm is projection-free and achieves better violation bounds in rapidly varying environments. Our results hold in the most general case when both the cost and constraint functions are chosen arbitrarily, and the constraint functions need not contain any fixed common feasible point. We establish these results by introducing a general framework that reduces the constrained learning problem to an instance of the standard OCO problem with specially constructed surrogate cost functions.

Method: The authors analyze loss landscape geometry using the River-Valley model, distinguishing between U-shaped (flat) and V-shaped (steep) valleys. They empirically observe distinct dynamics at sample and Hessian levels, and conjecture that Looped-Attn’s recursive architecture induces a V-shaped valley landscape. Based on this insight, they propose SHIFT - a staged hierarchical framework for progressive training.

Result: The analysis reveals that Looped-Attn’s recursive architecture creates River-V-Valley landscapes that enable better loss convergence through valley hopping and encourage learning of complex patterns, compared to Single-Attn’s River-U-Valley landscapes.

Conclusion: The geometric properties of loss landscapes explain Looped-Attn’s advantage, and the proposed SHIFT training strategy can accelerate Looped-Attn training while maintaining comparable performance by leveraging these landscape insights.

Abstract: While looped transformers (termed as Looped-Attn) often outperform standard transformers (termed as Single-Attn) on complex reasoning tasks, the mechanism for this advantage remains underexplored. In this paper, we explain this phenomenon through the lens of loss landscape geometry, inspired by empirical observations of their distinct dynamics at both sample and Hessian levels. To formalize this, we extend the River-Valley landscape model by distinguishing between U-shaped valleys (flat) and V-shaped valleys (steep). Based on empirical observations, we conjecture that the recursive architecture of Looped-Attn induces a landscape-level inductive bias towards River-V-Valley. This inductive bias suggest a better loss convergence along the river due to valley hopping, and further encourage learning about complex patterns compared to the River-U-Valley induced by Single-Attn. Building on this insight, we propose SHIFT (Staged HIerarchical Framework for Progressive Training), a principled training strategy that accelerates the training process of Looped-Attn while achieving comparable performances.

Method: Transform unstructured trial data into meaningful representations using pretrained foundation models, then interpret via sparse autoencoder. Introduce Neural Effect Search - a recursive procedure using progressive stratification to address multiple-testing issues and effects entanglement.

Result: Algorithm robustness assessed on semi-synthetic experiments. Successfully demonstrated first unsupervised causal effect identification on a real-world scientific trial in experimental ecology context.

Conclusion: Neural Effect Search enables discovery of unknown causal effects directly from trial data, overcoming limitations of traditional hypothesis-driven RCTs and allowing for scalable causal effect estimation.

Abstract: Randomized Controlled Trials are one of the pillars of science; nevertheless, they rely on hand-crafted hypotheses and expensive analysis. Such constraints prevent causal effect estimation at scale, potentially anchoring on popular yet incomplete hypotheses. We propose to discover the unknown effects of a treatment directly from data. For this, we turn unstructured data from a trial into meaningful representations via pretrained foundation models and interpret them via a sparse autoencoder. However, discovering significant causal effects at the neural level is not trivial due to multiple-testing issues and effects entanglement. To address these challenges, we introduce Neural Effect Search, a novel recursive procedure solving both issues by progressive stratification. After assessing the robustness of our algorithm on semi-synthetic experiments, we showcase, in the context of experimental ecology, the first successful unsupervised causal effect identification on a real-world scientific trial.

Method: C-MCTD introduces three components: 1) Online Composer for globally-aware planning across entire plan compositions, 2) Distributed Composer for parallel exploration from multiple starting points to reduce complexity, and 3) Preplan Composer that uses cached plan graphs to accelerate inference.

Result: The proposed framework elevates planning from individual trajectory optimization to reasoning over complete plan compositions, overcoming MCTD’s limitations in global context and training length constraints.

Conclusion: C-MCTD represents a significant advancement over MCTD by enabling compositional reasoning and global-aware planning, making it more suitable for complex, long-horizon planning tasks.

Abstract: Monte Carlo Tree Diffusion (MCTD) integrates diffusion models with structured tree search to enable effective trajectory exploration through stepwise reasoning. However, MCTD remains fundamentally limited by training trajectory lengths. While periodic replanning allows plan concatenation for longer plan generation, the planning process remains locally confined, as MCTD searches within individual trajectories without access to global context. We propose Compositional Monte Carlo Tree Diffusion (C-MCTD), a framework that elevates planning from individual trajectory optimization to reasoning over complete plan compositions. C-MCTD introduces three complementary components: (1) Online Composer, which performs globally-aware planning by searching across entire plan compositions; (2) Distributed Composer, which reduces search complexity through parallel exploration from multiple starting points; and (3) Preplan Composer, which accelerates inference by leveraging cached plan graphs.

Method: The paper proposes constraining LoRA factors to be block-diagonal matrices. This structural constraint enables an alternative sharding strategy where LoRA adapters can be partitioned across devices without requiring any additional communication during LoRA computations, aligning well with tensor parallel execution of the base model.

Result: Block-diagonal LoRA achieves similar downstream performance to standard LoRA for comparable parameter counts (parameter efficient). It delivers significant end-to-end speed-ups over S-LoRA: up to 1.79x speed-up on 8 A100 GPUs for Llama-3.1-70B with 0.87x adapter parameters, and up to 1.63x speed-up for Llama-3.1-8B with 0.86x adapter parameters.

Conclusion: The block-diagonal constraint on LoRA factors provides a practical solution for efficient multi-adapter serving by eliminating communication overhead in sharding, making it superior to existing approaches like S-LoRA while maintaining the parameter efficiency benefits of standard LoRA.

Abstract: When serving a single base LLM with several different LoRA adapters simultaneously, the adapters cannot simply be merged with the base model’s weights as the adapter swapping would create overhead and requests using different adapters could not be batched. Rather, the LoRA computations have to be separated from the base LLM computations, and in a multi-device setup the LoRA adapters can be sharded in a way that is well aligned with the base model’s tensor parallel execution, as proposed in S-LoRA. However, the S-LoRA sharding strategy encounters some communication overhead, which may be small in theory, but can be large in practice. In this paper, we propose to constrain certain LoRA factors to be block-diagonal, which allows for an alternative way of sharding LoRA adapters that does not require any additional communication for the LoRA computations. We demonstrate in extensive experiments that our block-diagonal LoRA approach is similarly parameter efficient as standard LoRA (i.e., for a similar number of parameters it achieves similar downstream performance) and that it leads to significant end-to-end speed-up over S-LoRA. For example, when serving on eight A100 GPUs, we observe up to 1.79x (1.23x) end-to-end speed-up with 0.87x (1.74x) the number of adapter parameters for Llama-3.1-70B, and up to 1.63x (1.3x) end-to-end speed-up with 0.86x (1.73x) the number of adapter parameters for Llama-3.1-8B.

Method: LTE (Learning to reason from Trial and Error) - an approach that hints language models with their previously self-made mistakes during training, enabling learning from trial and error without requiring any external expert guidance.

Result: LTE outperforms normal GRPO by 5.02 in Pass@1 and 9.96 in Pass@k on average across six mathematical reasoning benchmarks for Qwen3-8B-Base. It even performs better than methods requiring external gold solutions after aligning experimental setup.

Conclusion: LTE successfully mitigates exploration stagnation and enhances both exploitation and exploration during training by enabling LMs to learn from their own mistakes, providing a scalable solution without external expert guidance.

Abstract: Reinforcement learning with verifiable rewards (RLVR) has significantly boosted the reasoning capability of language models (LMs) recently. However, existing RLVR approaches merely train LMs based on their own generated on-policy responses and are constrained by the initial capability of LMs, thus prone to exploration stagnation, in which LMs fail to solve more training problems and cannot further learn from the training data. Some work tries to address this by leveraging off-policy solutions to training problems, but relies on external expert guidance that is limited in availability and scalability. In this work, we propose LTE (Learning to reason from Trial and Error), an approach that hints LMs with their previously self-made mistakes, not requiring any external expert guidance. Experiments validate the effectiveness of LTE, which outperforms the normal group relative policy optimization (GRPO) by 5.02 in Pass@1 and 9.96 in Pass@k on average across six mathematical reasoning benchmarks for Qwen3-8B-Base and even performs better than methods that require external gold solutions as guidance after aligning the experimental setup. Further analysis confirms that LTE successfully mitigates exploration stagnation and enhances both exploitation and exploration during training. Our code is available at https://anonymous.4open.science/r/Learning-from-Trial-and-Error.

Method: Proposes constrained offline policy optimization framework with group-wise reward disparity constraints in off-policy gradient-based optimization. Uses doubly robust estimator to improve group-wise reward disparity estimation during training, with convergence guarantee for policy optimization.

Result: Empirical results on synthetic and real-world datasets show the method effectively reduces reward disparities while maintaining competitive overall performance.

Conclusion: The proposed framework successfully addresses group-sensitive fairness in offline contextual bandits by constraining reward disparities, providing a practical solution for fair policy learning from historical data.

Abstract: Offline contextual bandits allow one to learn policies from historical/offline data without requiring online interaction. However, offline policy optimization that maximizes overall expected rewards can unintentionally amplify the reward disparities across groups. As a result, some groups might benefit more than others from the learned policy, raising concerns about fairness, especially when the resources are limited. In this paper, we study a group-sensitive fairness constraint in offline contextual bandits, reducing group-wise reward disparities that may arise during policy learning. We tackle the following common-parity requirements: the reward disparity is constrained within some user-defined threshold or the reward disparity should be minimized during policy optimization. We propose a constrained offline policy optimization framework by introducing group-wise reward disparity constraints into an off-policy gradient-based optimization procedure. To improve the estimation of the group-wise reward disparity during training, we employ a doubly robust estimator and further provide a convergence guarantee for policy optimization. Empirical results in synthetic and real-world datasets demonstrate that our method effectively reduces reward disparities while maintaining competitive overall performance.

Method: Proposes SynthAgent with dual refinement: 1) Synthesizes diverse tasks through categorized exploration of web elements, 2) Refines tasks only when conflicts with observations are detected during trajectory collection, 3) Conducts trajectory refinement with global context after collection to mitigate noise/misalignments, 4) Fine-tunes open-source web agents on refined synthetic data.

Result: SynthAgent outperforms existing synthetic data methods, validating the importance of high-quality synthetic supervision for web agent adaptation.

Conclusion: The framework demonstrates that dual refinement of both tasks and trajectories significantly improves synthetic data quality, enabling better adaptation of web agents to new websites through high-quality synthetic supervision.

Abstract: Web agents struggle to adapt to new websites due to the scarcity of environment specific tasks and demonstrations. Recent works have explored synthetic data generation to address this challenge, however, they suffer from data quality issues where synthesized tasks contain hallucinations that cannot be executed, and collected trajectories are noisy with redundant or misaligned actions. In this paper, we propose SynthAgent, a fully synthetic supervision framework that aims at improving synthetic data quality via dual refinement of both tasks and trajectories. Our approach begins by synthesizing diverse tasks through categorized exploration of web elements, ensuring efficient coverage of the target environment. During trajectory collection, tasks are refined only when conflicts with observations are detected, which mitigates hallucinations while preserving task consistency. After collection, we conduct trajectory refinement with global context to mitigate potential noise or misalignments. Finally, we fine-tune open-source web agents on the refined synthetic data to adapt them to the target environment. Experimental results demonstrate that SynthAgent outperforms existing synthetic data methods, validating the importance of high-quality synthetic supervision. The code is publicly available at https://github.com/aiming-lab/SynthAgent.

Method: Proposes PhyKAN (physics-informed Kolmogorov-Arnold Network) that generalizes RWG basis into learnable, adaptive basis family. Combines local KAN branch with global branch embedded with Green’s function priors derived from EFIE to maintain physical consistency.

Result: PhyKAN achieves sub-0.01 reconstruction errors across canonical geometries and provides accurate, unsupervised radar cross section predictions.

Conclusion: PhyKAN offers an interpretable, physics-consistent bridge between classical electromagnetic solvers and modern neural network models, enabling adaptive basis learning while maintaining physical constraints.

Abstract: The Method of Moments (MoM) is constrained by the usage of static, geometry-defined basis functions, such as the Rao-Wilton-Glisson (RWG) basis. This letter reframes electromagnetic modeling around a learnable basis representation rather than solving for the coefficients over a fixed basis. We first show that the RWG basis is essentially a static and piecewise-linear realization of the Kolmogorov-Arnold representation theorem. Inspired by this insight, we propose PhyKAN, a physics-informed Kolmogorov-Arnold Network (KAN) that generalizes RWG into a learnable and adaptive basis family. Derived from the EFIE, PhyKAN integrates a local KAN branch with a global branch embedded with Green’s function priors to preserve physical consistency. It is demonstrated that, across canonical geometries, PhyKAN achieves sub-0.01 reconstruction errors as well as accurate, unsupervised radar cross section predictions, offering an interpretable, physics-consistent bridge between classical solvers and modern neural network models for electromagnetic modeling.

Method: Derives generalization theory for equality-constrained statistical learning, then proposes practical algorithm that solves sequence of unconstrained empirical learning problems to approximate constrained solutions.

Result: Shows solutions to equality-constrained problems can be approximated using samples and rich parametrizations; demonstrates effectiveness in fair learning, interpolating classifiers, and boundary value problems.

Conclusion: Provides theoretical foundation and practical algorithm for handling equality constraints in ML, enabling new formulations for fairness, interpolation, and boundary problems without hyperparameter tuning burdens.

Abstract: As machine learning applications grow increasingly ubiquitous and complex, they face an increasing set of requirements beyond accuracy. The prevalent approach to handle this challenge is to aggregate a weighted combination of requirement violation penalties into the training objective. To be effective, this approach requires careful tuning of these hyperparameters (weights), involving trial-and-error and cross-validation, which becomes ineffective even for a moderate number of requirements. These issues are exacerbated when the requirements involve parities or equalities, as is the case in fairness and boundary value problems. An alternative technique uses constrained optimization to formulate these learning problems. Yet, existing approximation and generalization guarantees do not apply to problems involving equality constraints. In this work, we derive a generalization theory for equality-constrained statistical learning problems, showing that their solutions can be approximated using samples and rich parametrizations. Using these results, we propose a practical algorithm based on solving a sequence of unconstrained, empirical learning problems. We showcase its effectiveness and the new formulations enabled by equality constraints in fair learning, interpolating classifiers, and boundary value problems.

Method: Three-stage pipeline: 1) Patch Low-Frequency forecasting Module extracts stable low-frequency trends from spectral patches, 2) residual learner models high-frequency variations, 3) fine-tuned LLM refines predictions using auxiliary context and domain knowledge via structured natural language prompts.

Result: Extensive experiments on financial and energy datasets show LoFT-LLM significantly outperforms strong baselines in both full-data and few-shot regimes, delivering superior accuracy, robustness, and interpretability.

Conclusion: LoFT-LLM effectively addresses challenges in time-series forecasting by integrating frequency-aware learning with LLM-based semantic calibration, providing a powerful solution for noisy, data-scarce domains like finance and energy.

Abstract: Time-series forecasting in real-world applications such as finance and energy often faces challenges due to limited training data and complex, noisy temporal dynamics. Existing deep forecasting models typically supervise predictions using full-length temporal windows, which include substantial high-frequency noise and obscure long-term trends. Moreover, auxiliary variables containing rich domain-specific information are often underutilized, especially in few-shot settings. To address these challenges, we propose LoFT-LLM, a frequency-aware forecasting pipeline that integrates low-frequency learning with semantic calibration via a large language model (LLM). Firstly, a Patch Low-Frequency forecasting Module (PLFM) extracts stable low-frequency trends from localized spectral patches. Secondly, a residual learner then models high-frequency variations. Finally, a fine-tuned LLM refines the predictions by incorporating auxiliary context and domain knowledge through structured natural language prompts. Extensive experiments on financial and energy datasets demonstrate that LoFT-LLM significantly outperforms strong baselines under both full-data and few-shot regimes, delivering superior accuracy, robustness, and interpretability.

Method: Proposes CDSP-MoE framework with a super-complete parameter backbone where logical experts are carved out via learnable topology masks. Uses a Lagged Gradient Game to penalize interfering connections in shared manifold, enabling spontaneous pruning of conflicting pathways.

Result: Achieves robust content-driven routing without human-defined task labels, maintains semantic specialization even under strict blind inference protocols where explicit instructions are absent.

Conclusion: CDSP-MoE enables interpretable modular structures through conflict-driven subspace pruning, addressing fundamental limitations of contemporary MoE designs while maintaining parameter efficiency.

Abstract: Mixture-of-Experts (MoE) architectures achieve parameter efficiency through conditional computation, yet contemporary designs suffer from two fundamental limitations: structural parameter isolation that causes catastrophic forgetting, and instruction-overfitting that degrades performance in instruction-free scenarios. We propose CDSP-MoE (Conflict-Driven Subspace Pruning MoE), a framework that addresses these issues through a paradigm shift from isolated expert containers to dynamic expert instantiation within a shared physical subspace. Grounded in the Universal Weight Subspace Hypothesis, CDSP-MoE maintains a super-complete parameter backbone where logical experts are carved out via learnable topology masks. Unlike prior work that uses gradient conflict for token reassignment or optimization surgery, we leverage it as a structural supervisory signal: a Lagged Gradient Game penalizes interfering connections in the shared manifold, enabling the topology to spontaneously prune conflicting pathways and evolve interpretable modular structures. Experimental results demonstrate that CDSP-MoE achieves robust content-driven routing without human-defined task labels, maintaining semantic specialization even under strict blind inference protocols where explicit instructions are absent. Code is available at: https://github.com/konodiodaaaaa1/Conflict-Driven-Subspace-Pruning-Mixture-of-Experts

Method: Analyzed gradient behavior of different KL estimator configurations, then empirically tested by RL fine-tuning Qwen2.5-7B, Llama-3.1-8B-Instruct, and Qwen3-4B-Instruct-2507 with different configurations, evaluating on both in-distribution and out-of-distribution tasks.

Result: In on-policy settings: (1) biased gradient estimators cause training instability; (2) unbiased gradient estimators lead to better performance on both in-domain and out-of-domain tasks. In off-policy settings, KL regularization helps stabilize training in asynchronous setups.

Conclusion: Proper KL estimator design matters significantly for RL fine-tuning of LLMs. Using unbiased gradient estimators improves training stability and performance across domains, while biased estimators cause instability. KL regularization also benefits off-policy training.

Abstract: The reasoning performance of large language models (LLMs) can be substantially improved by training them with reinforcement learning (RL). The RL objective for LLM training involves a regularization term, which is the reverse Kullback-Leibler (KL) divergence between the trained policy and the reference policy. Since computing the KL divergence exactly is intractable, various estimators are used in practice to estimate it from on-policy samples. Despite its wide adoption, including in several open-source libraries, there is no systematic study analyzing the numerous ways of incorporating KL estimators in the objective and their effect on the downstream performance of RL-trained models. Recent works show that prevailing practices for incorporating KL regularization do not provide correct gradients for stated objectives, creating a discrepancy between the objective and its implementation. In this paper, we further analyze these practices and study the gradients of several estimators configurations, revealing how design choices shape gradient bias. We substantiate these findings with empirical observations by RL fine-tuning \texttt{Qwen2.5-7B}, \texttt{Llama-3.1-8B-Instruct} and \texttt{Qwen3-4B-Instruct-2507} with different configurations and evaluating their performance on both in- and out-of-distribution tasks. Through our analysis, we observe that, in on-policy settings: (1) estimator configurations with biased gradients can result in training instabilities; and (2) using estimator configurations resulting in unbiased gradients leads to better performance on in-domain as well as out-of-domain tasks. We also investigate the performance resulting from different KL configurations in off-policy settings and observe that KL regularization can help stabilize off-policy RL training resulting from asynchronous setups.

Method: DiRL integrates FlexAttention-accelerated blockwise training with LMDeploy-optimized inference for efficient online model updates, enabling two-stage post-training (Supervised Fine-Tuning + Reinforcement Learning). DiPO introduces unbiased Group Relative Policy Optimization (GRPO) tailored for dLLMs.

Result: DiRL-8B-Instruct trained on high-quality math data achieves state-of-the-art math performance among dLLMs and surpasses comparable Qwen2.5 series models on several benchmarks.

Conclusion: DiRL provides an effective post-training framework for dLLMs that addresses computational inefficiency and objective mismatch issues, enabling strong performance on complex reasoning tasks like mathematics.

Abstract: Diffusion Language Models (dLLMs) have emerged as promising alternatives to Auto-Regressive (AR) models. While recent efforts have validated their pre-training potential and accelerated inference speeds, the post-training landscape for dLLMs remains underdeveloped. Existing methods suffer from computational inefficiency and objective mismatches between training and inference, severely limiting performance on complex reasoning tasks such as mathematics. To address this, we introduce DiRL, an efficient post-training framework that tightly integrates FlexAttention-accelerated blockwise training with LMDeploy-optimized inference. This architecture enables a streamlined online model update loop, facilitating efficient two-stage post-training (Supervised Fine-Tuning followed by Reinforcement Learning). Building on this framework, we propose DiPO, the first unbiased Group Relative Policy Optimization (GRPO) implementation tailored for dLLMs. We validate our approach by training DiRL-8B-Instruct on high-quality math data. Our model achieves state-of-the-art math performance among dLLMs and surpasses comparable models in the Qwen2.5 series on several benchmarks.

Method: Simultaneous prediction of three alloy properties using MTL on 54,028 metal alloys dataset, analysis of learned task graphs, and evaluation of Deep Imbalanced Regression techniques (PCGrad, LDS+GradNorm) to address data imbalance issues.

Result: MTL significantly degrades regression accuracy (hardness R² drops from 0.832 to 0.694) but improves classification performance (amorphous F1 increases from 0.703 to 0.744). Analysis shows negligible inter-task correlations. PCGrad recovers hardness performance (R² → 0.855) by resolving gradient conflicts, while LDS+GradNorm achieves best overall multi-task balance.

Conclusion: Alloy properties often behave independently, requiring specific strategies: independent models for maximum regression precision, PCGrad for minority tasks, and LDS+GradNorm when balanced joint prediction is needed. The assumption of shared physical principles in MTL for materials informatics needs careful validation.

Abstract: Multi-task learning (MTL) is widely adopted in materials informatics under the assumption that related properties share leverageable physical principles. This study critically examines this premise by simultaneously predicting electrical resistivity, Vickers hardness, and amorphous-forming ability using a dataset of 54,028 metal alloys.1 Contrary to expectations, we observe a striking dichotomy: MTL significantly degrades regression accuracy (e.g., hardness 2$R^2$ drops from 3$0.832$ to 4$0.694$) while improving classification performance (amorphous F1 increases from 5$0.703$ to 6$0.744$).7 Analysis of learned task graphs reveals negligible inter-task correlations, attributing regression failure to negative transfer driven by severe data imbalance (52,388 vs. 800 samples). To mitigate this, we evaluate Deep Imbalanced Regression techniques. PCGrad recovers hardness performance ($R^2 \rightarrow 0.855$) by resolving gradient conflicts, while LDS+GradNorm achieves the best overall multi-task balance. Our findings suggest that alloy properties often behave independently, necessitating specific strategies: independent models for maximum regression precision, PCGrad for minority tasks, and LDS+GradNorm when balanced joint prediction is required.

Method: Bi-objective optimization framework using DAG constraints for feature importance hierarchies constructed via Central Limit Theorem, with Temporal Integrated Gradients for feature importance measurement and Optimal Path Oracle to handle OOD issues in TIG computation.

Result: Empirical results on time-series data show IGBO effectively enforces DAG constraints with minimal accuracy loss, outperforming standard regularization baselines. Theoretical analysis proves convergence properties and robustness to mini-batch noise.

Conclusion: IGBO provides a theoretically grounded framework for training interpretable models with domain knowledge constraints, successfully balancing interpretability requirements with predictive performance through bi-objective optimization.

Abstract: This paper introduces Interpretability-Guided Bi-objective Optimization (IGBO), a framework that trains interpretable models by incorporating structured domain knowledge via a bi-objective formulation. IGBO encodes feature importance hierarchies as a Directed Acyclic Graph (DAG) via Central Limit Theorem-based construction and uses Temporal Integrated Gradients (TIG) to measure feature importance. To address the Out-of-Distribution (OOD) problem in TIG computation, we propose an Optimal Path Oracle that learns data-manifold-aware integration paths. Theoretical analysis establishes convergence properties via a geometric projection mapping $\mathcal{P}$ and proves robustness to mini-batch noise. Central Limit Theorem-based construction of the interpretability DAG ensures statistical validity of edge orientation decisions. Empirical results on time-series data demonstrate IGBO’s effectiveness in enforcing DAG constraints with minimal accuracy loss, outperforming standard regularization baselines.

Method: Proposes a unified geometric framework with two orthogonal principles: 1) Manifold-constrained hyper-connections that restrict residual updates to valid local tangent directions to prevent manifold drift, and 2) Deep delta learning that introduces data-dependent, non-monotonic updates enabling reflection and erasure of redundant features rather than unconditional accumulation. These mechanisms decouple direction and sign of feature updates.

Result: The resulting architecture is called Manifold-Geometric Transformer (MGT). The analysis predicts that enforcing geometric validity while allowing dynamic erasure is essential for avoiding rank collapse in ultra-deep networks. The paper outlines an evaluation protocol for Transformers exceeding 100 layers to test the hypothesis that geometry, not depth itself, is the key limiting factor.

Conclusion: The collapse of deep Transformers is fundamentally a geometric problem, not just an optimization issue. By addressing geometric drift through manifold-constrained updates and enabling feature erasure through deep delta learning, the proposed MGT architecture offers a solution to representational collapse in ultra-deep networks, potentially enabling stable geometric evolution across extreme depths.

Abstract: Despite their empirical success, pushing Transformer architectures to extreme depth often leads to a paradoxical failure: representations become increasingly redundant, lose rank, and ultimately collapse. Existing explanations largely attribute this phenomenon to optimization instability or vanishing gradients, yet such accounts fail to explain why collapse persists even under modern normalization and initialization schemes. In this paper, we argue that the collapse of deep Transformers is fundamentally a geometric problem. Standard residual updates implicitly assume that feature accumulation is always beneficial, but offer no mechanism to constrain update directions or to erase outdated information. As depth increases, this leads to systematic drift off the semantic manifold and monotonic feature accumulation, causing representational degeneracy. We propose a unified geometric framework that addresses these failures through two orthogonal principles. First, manifold-constrained hyper-connections restrict residual updates to valid local tangent directions, preventing uncontrolled manifold drift. Second, deep delta learning introduces data-dependent, non-monotonic updates that enable reflection and erasure of redundant features rather than their unconditional accumulation. Together, these mechanisms decouple the direction and sign of feature updates, yielding a stable geometric evolution across depth. We term the resulting architecture the Manifold-Geometric Transformer (MGT). Our analysis predicts that enforcing geometric validity while allowing dynamic erasure is essential for avoiding rank collapse in ultra-deep networks. We outline an evaluation protocol for Transformers exceeding 100 layers to test the hypothesis that geometry, rather than depth itself, is the key limiting factor in deep representation learning.

Method: Transplants an information-theoretic Lyapunov function framework to reverse diffusion. Proposes V-delta projected reverse diffusion that projects the process using a leak-tolerant potential V-delta, which has closed-form expression as scaling-and-clipping on block masses. Extends monotonicity properties to time-inhomogeneous Markov kernels.

Result: Numerical experiments with a toy model (block-constant images, simplified reverse kernel) show the method keeps block-mass error and leak-tolerant potential within prescribed tolerance while maintaining pixel-wise accuracy and visual quality comparable to non-projected dynamics.

Conclusion: The study reinterprets generative sampling as decreasing an information potential from noise to data, providing a design principle for reverse diffusion processes with explicit control of coarse-grained quantities.

Abstract: Diffusion models and score-based generative models provide a powerful framework for synthesizing high-quality images from noise. However, there is still no satisfactory theory that describes how coarse-grained quantities, such as blockwise intensity or class proportions after partitioning an image into spatial blocks, are preserved and evolve along the reverse diffusion dynamics. In previous work, the author introduced an information-theoretic Lyapunov function V for non-ergodic Markov processes on a state space partitioned into blocks, defined as the minimal Kullback-Leibler divergence to the set of stationary distributions reachable from a given initial condition, and showed that a leak-tolerant potential V-delta with a prescribed tolerance for block masses admits a closed-form expression as a scaling-and-clipping operation on block masses. In this paper, I transplant this framework to the reverse diffusion process in generative models and propose a reverse diffusion scheme that is projected by the potential V-delta (referred to as the V-delta projected reverse diffusion). I extend the monotonicity of V to time-inhomogeneous block-preserving Markov kernels and show that, under small leakage and the V-delta projection, V-delta acts as an approximate Lyapunov function. Furthermore, using a toy model consisting of block-constant images and a simplified reverse kernel, I numerically demonstrate that the proposed method keeps the block-mass error and the leak-tolerant potential within the prescribed tolerance, while achieving pixel-wise accuracy and visual quality comparable to the non-projected dynamics. This study reinterprets generative sampling as a decrease of an information potential from noise to data, and provides a design principle for reverse diffusion processes with explicit control of coarse-grained quantities.

Method: ML-UCB assumes Mean Squared Error decreases as a power law in training samples, derives a generalized concentration inequality from this learning curve behavior, and integrates this into a UCB framework. This enables any ML model with empirically characterized learning curves to be used without theoretical analysis.

Result: Theoretical proof that ML-UCB achieves sublinear regret. Experimental validation on collaborative filtering recommendation system using online matrix factorization with synthetic data shows substantial improvements over LinUCB.

Conclusion: ML-UCB provides a general framework for integrating arbitrary ML models into bandit algorithms by leveraging empirical learning curve characterization, eliminating the need for model-specific theoretical guarantees while maintaining principled exploration.

Abstract: We present ML-UCB, a generalized upper confidence bound algorithm that integrates arbitrary machine learning models into multi-armed bandit frameworks. A fundamental challenge in deploying sophisticated ML models for sequential decision-making is the lack of tractable concentration inequalities required for principled exploration. We overcome this limitation by directly modeling the learning curve behavior of the underlying estimator. Specifically, assuming the Mean Squared Error decreases as a power law in the number of training samples, we derive a generalized concentration inequality and prove that ML-UCB achieves sublinear regret. This framework enables the principled integration of any ML model whose learning curve can be empirically characterized, eliminating the need for model-specific theoretical analysis. We validate our approach through experiments on a collaborative filtering recommendation system using online matrix factorization with synthetic data designed to simulate a simplified two-tower model, demonstrating substantial improvements over LinUCB

Method: AGNES employs block-wise storage I/O processing to fully utilize I/O bandwidth of high-performance storage devices, combined with hyperbatch-based processing that leverages characteristics of real-world graphs to enhance I/O efficiency.

Result: Comprehensive experiments on five real-world graphs show AGNES consistently outperforms four state-of-the-art methods, achieving up to 4.1X faster training than the best competitor.

Conclusion: AGNES effectively addresses the I/O bottleneck in storage-based GNN training through optimized block-wise I/O processing and hyperbatch strategies, enabling efficient training of web-scale graphs on single machines.

Abstract: Graph neural networks (GNNs) have achieved breakthroughs in various real-world downstream tasks due to their powerful expressiveness. As the scale of real-world graphs has been continuously growing, a storage-based approach to GNN training has been studied, which leverages external storage (e.g., NVMe SSDs) to handle such web-scale graphs on a single machine. Although such storage-based GNN training methods have shown promising potential in large-scale GNN training, we observed that they suffer from a severe bottleneck in data preparation since they overlook a critical challenge: how to handle a large number of small storage I/Os. To address the challenge, in this paper, we propose a novel storage-based GNN training framework, named AGNES, that employs a method of block-wise storage I/O processing to fully utilize the I/O bandwidth of high-performance storage devices. Moreover, to further enhance the efficiency of each storage I/O, AGNES employs a simple yet effective strategy, hyperbatch-based processing based on the characteristics of real-world graphs. Comprehensive experiments on five real-world graphs reveal that AGNES consistently outperforms four state-of-the-art methods, by up to 4.1X faster than the best competitor. Our code is available at https://github.com/Bigdasgit/agnes-kdd26.

Method: The paper demonstrates that safety alignment can be recovered with only a single safety example, regardless of harmful examples used in fine-tuning or model size, achieving convergence within few epochs. The approach leverages the discovered low-rank structure of safety gradients.

Result: The method effectively recovers safety alignment across five safety-aligned LLMs and multiple datasets, demonstrating generality and efficiency without sacrificing model utility.

Conclusion: Safety alignment in LLMs can be efficiently corrected with minimal data (single example) due to the low-rank structure of safety gradients, offering a practical solution to maintain safety without utility trade-offs.

Abstract: Fine-tuning safety-aligned large language models (LLMs) can substantially compromise their safety. Previous approaches require many safety samples or calibration sets, which not only incur significant computational overhead during realignment but also lead to noticeable degradation in model utility. Contrary to this belief, we show that safety alignment can be fully recovered with only a single safety example, without sacrificing utility and at minimal cost. Remarkably, this recovery is effective regardless of the number of harmful examples used in fine-tuning or the size of the underlying model, and convergence is achieved within just a few epochs. Furthermore, we uncover the low-rank structure of the safety gradient, which explains why such efficient correction is possible. We validate our findings across five safety-aligned LLMs and multiple datasets, demonstrating the generality of our approach.

Method: ASSIST uses ant colony optimization inspired by traveling salesperson approaches. It first performs node matching (peering) in O(q·log(d)) time, then uses iterative swarming/stigmergy-based subgraph search that’s linear in query size and constant in data size. The approach can be extended to handle temporal ordering, inexact matches, and missing nodes/edges.

Result: The method achieves significantly improved time complexity: O(q·log(d)) for initial matching plus linear time in query size and constant time in data size for the combinatorial search, compared to O(d²) for current heuristics and exponential complexity for exact solutions.

Conclusion: ASSIST provides an efficient approximate solution to the NP-complete subgraph isomorphism problem with superior time complexity that scales well with large data graphs, while maintaining flexibility to handle various real-world matching constraints.

Abstract: Maximum partial subgraph isomorphism compares two graphs (nodes joined by edges) to find a largest common subgraph. A common use case, for graphs with labeled nodes, seeks to find instances of a \textit{query} graph with $q$ nodes in a (typically larger) \textit{data} graph with $d$ nodes. The problem is NP-complete, and naïve solutions are exponential in $q + d$. The fastest current heuristic has complexity $O(d^2)$. This paper outlines ASSIST (Approximate Swarming Subgraph Isomorphism through Stigmergy), inspired by the ant colony optimization approach to the traveling salesperson. After peering (identifying matching individual nodes in query and data) in time $O(q\cdot log(d))$, the time required for ASSIST’s iterative subgraph search, the combinatorially complex part of the problem, is linear in query size and constant in data size. ASSIST can be extended to support matching problems (such as temporally ordered edges, inexact matches, and missing nodes or edges in the data graph) that frustrate other heuristics.

Method: Social force theory-based microscopic model where individuals follow two basic rules: moving toward goals at certain speeds and maintaining distance from others/obstacles. Includes evolutionary algorithm for layout optimization, decision-making model for goal selection, adaptation/learning capabilities, and methods for path system analysis.

Result: Complex spatial-temporal structures emerge from simple individual interactions, forming self-organized trails. Shows strong dependencies between pedestrian flow properties and building geometry, demonstrates efficiency improvements by reducing walkable areas, and presents optimized path systems similar to natural transport networks.

Conclusion: Simple individual behavior rules can generate complex collective pedestrian flows, enabling infrastructure optimization through evolutionary algorithms and providing insights into self-organization principles similar to natural transport networks.

Abstract: This work presents a microscopic model to describe pedestrian flows based on the social force theory. The aim of this study is twofold: (1) developing a realistic model that can be used as a tool for designing pedestrian-friendly infrastructure, and (2) verifying a social science theory using a model with sufficient data. The investigation of the pedestrian model shows that despite simple individual behavior patterns, complex spatial and temporal structures emerge through the interactions in pedestrian flows. Collective behavior emerges from individuals following two basic rules: (1) moving directly towards their goal at a certain speed, and (2) maintaining a distance to other pedestrians and obstacles. This self-organized collective behavior manifests itself as trails that are formed by pedestrians moving in one direction. Furthermore, strong dependencies of the properties of pedestrian flows on geometric forms of buildings are shown, and the influence of geometric changes on performance characteristics is investigated. An example demonstrates how efficiency can be increased by reducing walkable areas. This work also presents an evolutionary algorithm for optimizing building layouts based on the social force model. Additionally, a decision-making model is integrated to describe alternative goal selection, and adaptation and learning capabilities are included to improve pedestrian avoidance behavior and decision strategies based on accumulated experience. A method for determining load distributions in individual sections of a path system considering subjective selection criteria is also developed. Finally, a model that describes the self-organization of path systems with minimal detours is presented, similar to natural transport networks where total length and material costs are optimized.

Method: Derives a unique, continuous, power-optimal current solution using sequential convex programming, then approximates this solution using a multilayer perceptron (neural network) model.

Result: The effectiveness of the proposed strategy was demonstrated through both numerical simulations and experimental trials on formation and attitude control.

Conclusion: The learning-based approach successfully achieves power-optimal magnetic-field interaction for multi-agent formation and attitude control, with validation through simulations and experiments.

Abstract: This paper presents a learning-based current calculation model to achieve power-optimal magnetic-field interaction for multi-agent formation and attitude control. In aerospace engineering, electromagnetic coils are referred to as magnetorquer (MTQ) coils and used as satellite attitude actuators in Earth’s orbit and for long-term formation and attitude control. This study derives a unique, continuous, and power-optimal current solution via sequential convex programming and approximates it using a multilayer perceptron model. The effectiveness of our strategy was demonstrated through numerical simulations and experimental trials on the formation and attitude control.

Method: Proposes a self-learning framework using SE(3)-equivariant graph neural networks with recurrent temporal modeling. Trained via dual self-supervised objective to detect surprising auditory events that deviate from learned temporal expectations, modeling natural attention decay where surprise diminishes as listeners adapt.

Result: Experiments on AVTrack360 dataset show integrating audio surprise with visual cues reduces bitrate waste by up to 18% compared to visual-only methods.

Conclusion: Spatial audio surprise provides valuable cues for viewport prediction in 360-degree video streaming, and combining auditory and visual information significantly improves bandwidth efficiency.

Abstract: Adaptive streaming of 360-degree video relies on viewport prediction to allocate bandwidth efficiently. Current approaches predominantly use visual saliency or historical gaze patterns, neglecting the role of spatial audio in guiding user attention. This paper presents a self-learning framework for detecting “surprising” auditory events – moments that deviate from learned temporal expectations – and demonstrates their utility for viewport prediction. The proposed architecture combines $SE(3)$-equivariant graph neural networks with recurrent temporal modeling, trained via a dual self-supervised objective. A key feature is the natural modeling of temporal attention decay: surprise is high at event onset but diminishes as the listener adapts. Experiments on the AVTrack360 dataset show that integrating audio surprise with visual cues reduces bitrate waste by up to 18% compared to visual-only methods.

Method: Uses CLIP encoder’s facial-semantic knowledge on noisy audio-visual data to learn face-voice associations efficiently. Then employs retrieval-based strategy with GMM-based speaker generation module for downstream TTS system to produce probable target speakers given reference images.

Result: Achieves 89.63% cross-modal verification AUC score on Voxceleb testset. The system bridges gap between face and voice features for face-based speech synthesis, and feedback from downstream TTS helps synthesize voices that closely match reference faces.

Conclusion: Vclip effectively learns face-voice associations from noisy data using CLIP, and the retrieval-based approach combined with TTS feedback enables high-quality face-based speech synthesis that closely matches reference faces.

Abstract: This paper discusses the task of face-based speech synthesis, a kind of personalized speech synthesis where the synthesized voices are constrained to perceptually match with a reference face image. Due to the lack of TTS-quality audio-visual corpora, previous approaches suffer from either low synthesis quality or domain mismatch induced by a knowledge transfer scheme. This paper proposes a new approach called Vclip that utilizes the facial-semantic knowledge of the CLIP encoder on noisy audio-visual data to learn the association between face and voice efficiently, achieving 89.63% cross-modal verification AUC score on Voxceleb testset. The proposed method then uses a retrieval-based strategy, combined with GMM-based speaker generation module for a downstream TTS system, to produce probable target speakers given reference images. Experimental results demonstrate that the proposed Vclip system in conjunction with the retrieval step can bridge the gap between face and voice features for face-based speech synthesis. And using the feedback information distilled from downstream TTS helps to synthesize voices that match closely with reference faces. Demos available at sos1sos2sixteen.github.io/vclip.

Method: Proposes XLSR-MamBo framework with XLSR front-end and synergistic Mamba-Attention backbones. Evaluates four topological designs using SSM variants (Mamba, Mamba2, Hydra, Gated DeltaNet) with systematic architecture exploration.

Result: MamBo-3-Hydra-N3 configuration achieves competitive performance on ASVspoof 2021 LA, DF, and In-the-Wild benchmarks. Shows robust generalization to unseen diffusion/flow-matching methods on DFADD dataset. Scaling backbone depth mitigates performance variance in shallower models.

Conclusion: Hybrid framework effectively captures artifacts in spoofed speech signals, providing an effective audio deepfake detection method. Hydra’s native bidirectional modeling captures temporal dependencies more efficiently than previous dual-branch strategies.

Abstract: Advanced speech synthesis technologies have enabled highly realistic speech generation, posing security risks that motivate research into audio deepfake detection (ADD). While state space models (SSMs) offer linear complexity, pure causal SSMs architectures often struggle with the content-based retrieval required to capture global frequency-domain artifacts. To address this, we explore the scaling properties of hybrid architectures by proposing XLSR-MamBo, a modular framework integrating an XLSR front-end with synergistic Mamba-Attention backbones. We systematically evaluate four topological designs using advanced SSM variants, Mamba, Mamba2, Hydra, and Gated DeltaNet. Experimental results demonstrate that the MamBo-3-Hydra-N3 configuration achieves competitive performance compared to other state-of-the-art systems on the ASVspoof 2021 LA, DF, and In-the-Wild benchmarks. This performance benefits from Hydra’s native bidirectional modeling, which captures holistic temporal dependencies more efficiently than the heuristic dual-branch strategies employed in prior works. Furthermore, evaluations on the DFADD dataset demonstrate robust generalization to unseen diffusion- and flow-matching-based synthesis methods. Crucially, our analysis reveals that scaling backbone depth effectively mitigates the performance variance and instability observed in shallower models. These results demonstrate the hybrid framework’s ability to capture artifacts in spoofed speech signals, providing an effective method for ADD.

Method: 1) FCaps dataset creation via novel end-to-end pipeline for direct audio caption grounding, avoiding LLM-based rewriting errors; 2) CLSP model with contrastive pre-training integrating global and fine-grained supervision.

Result: FCaps annotations surpass existing cascaded annotations in correctness, coverage, and naturalness. CLSP achieves strong performance across global/fine-grained speech-text retrieval, zero-shot paralinguistic classification, and speech style similarity scoring.

Conclusion: FCaps enables fine-grained speech-text modeling, and CLSP learns unified multi-granular representations that align well with human judgments, advancing language-speech representation pre-training.

Abstract: Modeling fine-grained speaking styles remains challenging for language-speech representation pre-training, as existing speech-text models are typically trained with coarse captions or task-specific supervision, and scalable fine-grained style annotations are unavailable. We present FCaps, a large-scale dataset with fine-grained free-text style descriptions, encompassing 47k hours of speech and 19M fine-grained captions annotated via a novel end-to-end pipeline that directly grounds detailed captions in audio, thereby avoiding the error propagation caused by LLM-based rewriting in existing cascaded pipelines. Evaluations using LLM-as-a-judge demonstrate that our annotations surpass existing cascaded annotations in terms of correctness, coverage, and naturalness. Building on FCaps, we propose CLSP, a contrastive language-speech pre-trained model that integrates global and fine-grained supervision, enabling unified representations across multiple granularities. Extensive experiments demonstrate that CLSP learns fine-grained and multi-granular speech-text representations that perform reliably across global and fine-grained speech-text retrieval, zero-shot paralinguistic classification, and speech style similarity scoring, with strong alignment to human judgments. All resources will be made publicly available.

Method: The method uses LLM’s decoding mechanism to generate candidate next tokens, then aligns each candidate with SSL acoustic model to get acoustic scores. Combines acoustic and LLM scores using MAP estimator decomposition, maintains beam of highest-scoring tokens, and proceeds iteratively through decoding steps.

Result: The approach outperforms current state-of-the-art LLM-based decoding, post-processing, and error-correcting methods across multiple datasets, particularly effective for challenging inputs like complex sentences, acronyms, and domain-specific vocabulary.

Conclusion: The proposed method successfully integrates LLMs with SSL acoustic models for ASR, overcoming LLM overconfidence issues and demonstrating superior performance on challenging speech recognition tasks compared to existing LLM-based approaches.

Abstract: Self-supervised automatic speech recognition (SSL-ASR) is an ASR approach that uses speech encoders pretrained on large amounts of unlabeled audio (e.g., wav2vec2.0 or HuBERT) and then fine-tunes them with limited labeled data to perform transcription. Decoding is usually performed with a CTC decoder, whose hypotheses are scored and refined using an external language model (LM), typically an n-gram or neural LM, which guides beam search to produce the final transcription. Using Large Language Models (LLMs) as external LMs remains a challenge, as their word probabilities are overly confident. The proposed method integrates an LLM with an SSL acoustic model by using the LLM’s decoding mechanism to generate a set of candidate next tokens. For each candidate, the SSL model provides an acoustic score by aligning it to the input acoustics of the SSL model. A combined acoustic and LLM score is then calculated based on decomposing the MAP estimator of words given the acoustic signal. The tokens with the highest combined scores are maintained in a beam, which is then used to proceed to the next decoding step. We illustrate the effectiveness of our method through a comprehensive comparison with the current state-of-the-art LLM-based decoding, post-processing, and error-correcting methods across multiple datasets. Our approach proves particularly effective when processing challenging inputs such as complex speech sentences, acronyms, and domain-specific vocabulary.

Method: Prospective study with 42 participants undergoing 7T knee MRI with both low-resolution (0.8×0.8×2 mm³) and high-resolution (0.4×0.4×2 mm³) sequences. Superresolution images were generated from low-resolution data using a Hybrid Attention Transformer model. Three radiologists assessed image quality, anatomic conspicuity, and pathology detection, with arthroscopy as reference in 10 cases.

Result: Superresolution images showed higher overall quality than low-resolution (median score 5 vs 4) and lower noise than high-resolution (5 vs 4). Visibility of cartilage, menisci, and ligaments was superior in both SR and HR compared to LR. However, detection rates and diagnostic performance (sensitivity, specificity, AUC) for intra-articular pathology were similar across all image types.

Conclusion: Deep learning superresolution improves subjective image quality in 7T knee MRI but does not increase diagnostic accuracy compared with standard low-resolution imaging, suggesting that while image appearance improves, it doesn’t translate to better diagnostic performance.

Abstract: Background: Deep learning superresolution (SR) may enhance musculoskeletal MR image quality, but its diagnostic value in knee imaging at 7T is unclear. Objectives: To compare image quality and diagnostic performance of SR, low-resolution (LR), and high-resolution (HR) 7T knee MRI. Methods: In this prospective study, 42 participants underwent 7T knee MRI with LR (0.80.82 mm3) and HR (0.40.42 mm3) sequences. SR images were generated from LR data using a Hybrid Attention Transformer model. Three radiologists assessed image quality, anatomic conspicuity, and detection of knee pathologies. Arthroscopy served as reference in 10 cases. Results: SR images showed higher overall quality than LR (median score 5 vs 4, P<.001) and lower noise than HR (5 vs 4, P<.001). Visibility of cartilage, menisci, and ligaments was superior in SR and HR compared to LR (P<.001). Detection rates and diagnostic performance (sensitivity, specificity, AUC) for intra-articular pathology were similar across image types (P>=.095). Conclusions: Deep learning superresolution improved subjective image quality in 7T knee MRI but did not increase diagnostic accuracy compared with standard LR imaging.

Method: Redesigned transform kernels with data-driven transforms, expanded transform partitioning, mode & coefficient dependent transform signaling, and several new coding tools including Intra/Inter Secondary Transforms (IST), Trellis Coded Quantization (TCQ), Adaptive Transform Coding (ATC), Probability Adaptation Rate Adjustment (PARA), Forward Skip Coding (FSC), Cross Chroma Component Transforms (CCTX), Parity Hiding (PH) tools, and improved lossless coding.

Result: AV2 achieves the highest quality video experience for video applications at a significantly reduced bitrate compared to previous standards.

Conclusion: AV2 represents a significant advancement in video compression technology, delivering superior compression efficiency and quality improvements over AV1 while maintaining practical complexity for real-world applications.

Abstract: AV2 is the successor to the AV1 royalty-free video coding standard developed by the Alliance for Open Media (AOMedia). Its primary objective is to deliver substantial compression gains and subjective quality improvements while maintaining low-complexity encoder and decoder operations. This paper describes the transform, quantization and entropy coding design in AV2, including redesigned transform kernels and data-driven transforms, expanded transform partitioning, and a mode & coefficient dependent transform signaling. AV2 introduces several new coding tools including Intra/Inter Secondary Transforms (IST), Trellis Coded Quantization (TCQ), Adaptive Transform Coding (ATC), Probability Adaptation Rate Adjustment (PARA), Forward Skip Coding (FSC), Cross Chroma Component Transforms (CCTX), Parity Hiding (PH) tools and improved lossless coding. These advances enable AV2 to deliver the highest quality video experience for video applications at a significantly reduced bitrate.

Method: 1) EGxFSL: Integrates radiologist-defined ROIs as spatial supervision via Grad-CAM-based Dice loss, jointly optimized with prototypical classification for interpretable few-shot learning. 2) xGAL: Introduces iterative sample acquisition prioritizing both predictive uncertainty and attention misalignment, creating a closed-loop framework where explainability guides training and sample selection synergistically.

Result: Achieved accuracies of 92% (BraTS MRI), 76% (VinDr-CXR), and 62% (SIIM-COVID-19), consistently outperforming non-guided baselines. Under severe data constraints, xGAL achieved 76% accuracy with only 680 samples versus 57% for random sampling. Grad-CAM visualizations show models focus on diagnostically relevant regions, with cross-modality validation on breast ultrasound.

Conclusion: The proposed dual-framework solution successfully addresses both data scarcity and interpretability challenges in medical image analysis, demonstrating improved performance and clinical relevance through explainability-guided approaches that synergistically combine training and sample selection.

Abstract: Medical image analysis faces two critical challenges: scarcity of labeled data and lack of model interpretability, both hindering clinical AI deployment. Few-shot learning (FSL) addresses data limitations but lacks transparency in predictions. Active learning (AL) methods optimize data acquisition but overlook interpretability of acquired samples. We propose a dual-framework solution: Expert-Guided Explainable Few-Shot Learning (EGxFSL) and Explainability-Guided AL (xGAL). EGxFSL integrates radiologist-defined regions-of-interest as spatial supervision via Grad-CAM-based Dice loss, jointly optimized with prototypical classification for interpretable few-shot learning. xGAL introduces iterative sample acquisition prioritizing both predictive uncertainty and attention misalignment, creating a closed-loop framework where explainability guides training and sample selection synergistically. On the BraTS (MRI), VinDr-CXR (chest X-ray), and SIIM-COVID-19 (chest X-ray) datasets, we achieve accuracies of 92%, 76%, and 62%, respectively, consistently outperforming non-guided baselines across all datasets. Under severe data constraints, xGAL achieves 76% accuracy with only 680 samples versus 57% for random sampling. Grad-CAM visualizations demonstrate guided models focus on diagnostically relevant regions, with generalization validated on breast ultrasound confirming cross-modality applicability.

Method: Evaluated four binarization methods (LSH, ITQ, KSH, SDH) on the ODIR dataset using deep feature embeddings, comparing their performance with different bit lengths and benchmarking against prior studies.

Result: SDH achieved the best performance with mAP@100 of 0.9184 using only 32-bit codes, outperforming LSH, ITQ, and KSH. The method proved highly competitive with state-of-the-art approaches despite using significantly fewer bits compared to prior studies.

Conclusion: SDH is the most effective approach among those tested, offering a practical balance of accuracy, storage, and efficiency for medical image retrieval applications, demonstrating that high retrieval accuracy can be achieved with compact binary codes.

Abstract: In this study, we evaluated four binarization methods. Locality-Sensitive Hashing (LSH), Iterative Quantization (ITQ), Kernel-based Supervised Hashing (KSH), and Supervised Discrete Hashing (SDH) on the ODIR dataset using deep feature embeddings. Experimental results show that SDH achieved the best performance, with an mAP@100 of 0.9184 using only 32-bit codes, outperforming LSH, ITQ, and KSH. Compared with prior studies, our method proved highly competitive: Fang et al. reported 0.7528 (Fundus-iSee, 48 bits) and 0.8856 (ASOCT-Cataract, 48 bits), while Wijesinghe et al. achieved 94.01 (KVASIR, 256 bits). Despite using significantly fewer bits, our SDH-based framework reached retrieval accuracy close to the state-of-the-art. These findings demonstrate that SDH is the most effective approach among those tested, offering a practical balance of accuracy, storage, and efficiency for medical image retrieval and device inventory management.

Method: Proposes fast distillation to transfer pre-trained diffusion models into multi-scale EBMs, then introduces Annealed Langevin Posterior Sampling (ALPS) algorithm for MAP, MMSE, and uncertainty estimation. Unlike diffusion guidance on latent variables, ALPS performs annealing on static posterior distributions.

Result: Experiments on image inpainting and MRI reconstruction show the method matches or surpasses diffusion-based baselines in both accuracy and efficiency, while also supporting MAP recovery. The framework offers scalable solution for inverse problems.

Conclusion: The ALPS framework provides a scalable and principled solution for inverse problems in imaging with practical deployment potential in scientific and clinical settings, combining efficiency of diffusion models with interpretability of EBMs.

Abstract: Solving inverse problems in imaging requires models that support efficient inference, uncertainty quantification, and principled probabilistic reasoning. Energy-Based Models (EBMs), with their interpretable energy landscapes and compositional structure, are well-suited for this task but have historically suffered from high computational costs and training instability. To overcome the historical shortcomings of EBMs, we introduce a fast distillation strategy to transfer the strengths of pre-trained diffusion models into multi-scale EBMs. These distilled EBMs enable efficient sampling and preserve the interpretability and compositionality inherent to potential-based frameworks. Leveraging EBM compositionality, we propose Annealed Langevin Posterior Sampling (ALPS) algorithm for Maximum-A-Posteriori (MAP), Minimum Mean Square Error (MMSE), and uncertainty estimates for inverse problems in imaging. Unlike diffusion models that use complex guidance strategies for latent variables, we perform annealing on static posterior distributions that are well-defined and composable. Experiments on image inpainting and MRI reconstruction demonstrate that our method matches or surpasses diffusion-based baselines in both accuracy and efficiency, while also supporting MAP recovery. Overall, our framework offers a scalable and principled solution for inverse problems in imaging, with potential for practical deployment in scientific and clinical settings. ALPS code is available at the GitHub repository \href{https://github.com/JyoChand/ALPS}{ALPS}.

Method: Swin Transformer UNet 3D (SwinUNet3D) framework combining shifted window self-attention with U-Net style skip connections for lesion segmentation in FDG-PET/CT scans, evaluated on AutoPET III FDG dataset.

Result: SwinUNet3D achieves Dice score 0.88 and IoU 0.78, significantly outperforming baseline 3D U-Net (Dice 0.48, IoU 0.32) with faster inference times, better small lesion detection, reduced false positives, and improved PET/CT fusion.

Conclusion: SwinUNet3D represents an efficient and robust approach to PET/CT lesion segmentation, advancing transformer-based models in oncology imaging, with potential for future multi-tracer, multi-center evaluations and benchmarking.

Abstract: Accurate and automated lesion segmentation in Positron Emission Tomography / Computed Tomography (PET/CT) imaging is essential for cancer diagnosis and therapy planning. This paper presents a Swin Transformer UNet 3D (SwinUNet3D) framework for lesion segmentation in Fluorodeoxyglucose Positron Emission Tomography / Computed Tomography (FDG-PET/CT) scans. By combining shifted window self-attention with U-Net style skip connections, the model captures both global context and fine anatomical detail. We evaluate SwinUNet3D on the AutoPET III FDG dataset and compare it against a baseline 3D U-Net. Results show that SwinUNet3D achieves a Dice score of 0.88 and IoU of 0.78, surpassing 3D U-Net (Dice 0.48, IoU 0.32) while also delivering faster inference times. Qualitative analysis demonstrates improved detection of small and irregular lesions, reduced false positives, and more accurate PET/CT fusion. While the framework is currently limited to FDG scans and trained under modest GPU resources, it establishes a strong foundation for future multi-tracer, multi-center evaluations and benchmarking against other transformer-based architectures. Overall, SwinUNet3D represents an efficient and robust approach to PET/CT lesion segmentation, advancing the integration of transformer-based models into oncology imaging workflows.

Method: Proposes DiT-JSCC with: 1) semantics-detail dual-branch encoder aligned with coarse-to-fine conditional DiT decoder, 2) training-free adaptive bandwidth allocation based on Kolmogorov complexity, and 3) joint learning of semantics-prioritized representation encoder with diffusion transformer decoder.

Result: Extensive experiments show DiT-JSCC consistently outperforms existing JSCC methods in both semantic consistency and visual quality, particularly under extreme channel conditions like ultra-low bandwidth and low SNR.

Conclusion: DiT-JSCC successfully addresses the semantic consistency limitation in GJSCC by aligning encoder-decoder objectives, redefines information value in generative decoding era, and provides an open-source backbone for future GJSCC research.

Abstract: Generative joint source-channel coding (GJSCC) has emerged as a new Deep JSCC paradigm for achieving high-fidelity and robust image transmission under extreme wireless channel conditions, such as ultra-low bandwidth and low signal-to-noise ratio. Recent studies commonly adopt diffusion models as generative decoders, but they frequently produce visually realistic results with limited semantic consistency. This limitation stems from a fundamental mismatch between reconstruction-oriented JSCC encoders and generative decoders, as the former lack explicit semantic discriminability and fail to provide reliable conditional cues. In this paper, we propose DiT-JSCC, a novel GJSCC backbone that can jointly learn a semantics-prioritized representation encoder and a diffusion transformer (DiT) based generative decoder, our open-source project aims to promote the future research in GJSCC. Specifically, we design a semantics-detail dual-branch encoder that aligns naturally with a coarse-to-fine conditional DiT decoder, prioritizing semantic consistency under extreme channel conditions. Moreover, a training-free adaptive bandwidth allocation strategy inspired by Kolmogorov complexity is introduced to further improve the transmission efficiency, thereby indeed redefining the notion of information value in the era of generative decoding. Extensive experiments demonstrate that DiT-JSCC consistently outperforms existing JSCC methods in both semantic consistency and visual quality, particularly in extreme regimes.

Method: Collected eye images using seven different head-mounted eye trackers (including VR/AR devices) during various tasks like car rides, simulator rides, outdoor sports, and indoor activities. Provided comprehensive annotations including 2D/3D landmarks, semantic segmentation, 3D eyeball annotation, gaze vectors, and eye movement types.

Result: Created TEyeD dataset with over 20 million carefully annotated images, making it the world’s largest unified public dataset of eye images from head-mounted devices, with video lengths ranging from minutes to hours.

Conclusion: TEyeD provides a unique, coherent resource and valuable foundation for advancing research in computer vision, eye tracking, and gaze estimation, particularly for modern VR/AR applications.

Abstract: We present TEyeD, the world’s largest unified public data set of eye images taken with head-mounted devices. TEyeD was acquired with seven different head-mounted eye trackers. Among them, two eye trackers were integrated into virtual reality (VR) or augmented reality (AR) devices. The images in TEyeD were obtained from various tasks, including car rides, simulator rides, outdoor sports activities, and daily indoor activities. The data set includes 2D and 3D landmarks, semantic segmentation, 3D eyeball annotation and the gaze vector and eye movement types for all images. Landmarks and semantic segmentation are provided for the pupil, iris and eyelids. Video lengths vary from a few minutes to several hours. With more than 20 million carefully annotated images, TEyeD provides a unique, coherent resource and a valuable foundation for advancing research in the field of computer vision, eye tracking and gaze estimation in modern VR and AR applications. Download: https://es-cloud.cs.uni-tuebingen.de/d/8e2ab8c3fdd444e1a135/?p=%2FTEyeDS&mode=list Alternative Download: https://hctlsrva.edu.sot.tum.de/TEyeDS/

Method: Custom CNN and three fine-tuned transfer learning models (DenseNet121, ResNet152, VGG19) trained on IQ-OTH/NCCD dataset (1,197 CT scans across Normal, Benign, Malignant classes). Used cost-sensitive learning to address class imbalance and SHAP for explainability.

Result: ResNet152 achieved highest accuracy (97.3%), while DenseNet121 provided best overall balance in precision (92%), recall (90%), and F1-score (91%). SHAP visualizations successfully identified evidence contributing to predictions.

Conclusion: CNN-based approaches with explainability can provide fast, accurate, and interpretable support for lung cancer screening, particularly valuable in resource-limited healthcare settings.

Abstract: Early detection of lung cancer is critical to improving survival outcomes. We present a deep learning framework for automated lung cancer screening from chest computed tomography (CT) images with integrated explainability. Using the IQ-OTH/NCCD dataset (1,197 scans across Normal, Benign, and Malignant classes), we evaluate a custom convolutional neural network (CNN) and three fine-tuned transfer learning backbones: DenseNet121, ResNet152, and VGG19. Models are trained with cost-sensitive learning to mitigate class imbalance and evaluated via accuracy, precision, recall, F1-score, and ROC-AUC. While ResNet152 achieved the highest accuracy (97.3%), DenseNet121 provided the best overall balance in precision, recall, and F1 (up to 92%, 90%, 91%, respectively). We further apply Shapley Additive Explanations (SHAP) to visualize evidence contributing to predictions, improving clinical transparency. Results indicate that CNN-based approaches augmented with explainability can provide fast, accurate, and interpretable support for lung cancer screening, particularly in resource-limited settings.

Method: Analyzed 43,425 images from 965 Korean subjects (Fitzpatrick III-IV) across DSLR, tablet, and smartphone devices. Used Linear Color Correction Matrix (CCM) normalization and evaluated color accuracy (Delta E) versus clinical biomarker reliability (Individual Typology Angle ITA and Melanin Index).

Result: CCM reduced color error by 67-77% achieving near-clinical accuracy (Delta E < 2.3), but biomarker reliability varied: ITA showed poor inter-device agreement (ICC = 0.40) while Melanin Index achieved good agreement (ICC = 0.77). Facial region accounted for 25.2% of color variance (3.6x greater than device effects at 7.0%).

Conclusion: Current colorimetric standards are insufficient for clinical-grade biomarker extraction in mobile dermatology. “Color-clinical decoupling” occurs where perceptual accuracy doesn’t guarantee biomarker reliability, necessitating region-aware protocols rather than single-patch calibration.

Abstract: Smartphone-based tele-dermatology assumes that colorimetric calibration ensures clinical reliability, yet this remains untested for underrepresented skin phototypes. We investigated whether standard calibration translates to reliable clinical biomarkers using 43,425 images from 965 Korean subjects (Fitzpatrick III-IV) across DSLR, tablet, and smartphone devices. While Linear Color Correction Matrix (CCM) normalization reduced color error by 67-77% – achieving near-clinical accuracy (Delta E < 2.3) – this success did not translate to biomarker reliability. We identify a phenomenon termed “color-clinical decoupling”: despite perceptual accuracy, the Individual Typology Angle (ITA) showed poor inter-device agreement (ICC = 0.40), while the Melanin Index achieved good agreement (ICC = 0.77). This decoupling is driven by the ITA formula’s sensitivity to b* channel noise and is further compounded by anatomical variance. Facial region accounts for 25.2% of color variance – 3.6x greater than device effects (7.0%) – challenging the efficacy of single-patch calibration. Our results demonstrate that current colorimetric standards are insufficient for clinical-grade biomarker extraction, necessitating region-aware protocols for mobile dermatology.