Method: Proposes a co-training framework that simultaneously trains three safety behaviors within a single supervised fine-tuning stage, using system-level instructions or magic tokens to dynamically activate different behaviors during inference.

Result: Matches safety alignment quality of SFT+DPO, with 8B model surpassing DeepSeek-R1 (671B) in safety performance while reducing training complexity and deployment costs. Creates a distinct Safety Alignment Margin with well-separated response distributions.

Conclusion: Provides a scalable, efficient, and highly controllable solution for LLM content safety that enables fine-grained behavioral switching and robust safety performance with reduced complexity.

Abstract: Current methods for content safety in Large Language Models (LLMs), such as Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), often rely on multi-stage training pipelines and lack fine-grained, post-deployment controllability. To address these limitations, we propose a unified co-training framework that efficiently integrates multiple safety behaviors: positive (lawful/prosocial), negative (unfiltered/risk-prone) and rejective (refusal-oriented/conservative) within a single SFT stage. Notably, each behavior is dynamically activated via a simple system-level instruction, or magic token, enabling stealthy and efficient behavioral switching at inference time. This flexibility supports diverse deployment scenarios, such as positive for safe user interaction, negative for internal red-teaming, and rejective for context-aware refusals triggered by upstream moderation signals. This co-training strategy induces a distinct Safety Alignment Margin in the output space, characterized by well-separated response distributions corresponding to each safety mode. The existence of this margin provides empirical evidence for the model’s safety robustness and enables unprecedented fine-grained control. Experiments show that our method matches the safety alignment quality of SFT+DPO, with our 8B model notably surpassing DeepSeek-R1 (671B) in safety performance, while significantly reducing both training complexity and deployment costs. This work presents a scalable, efficient, and highly controllable solution for LLM content safety.

Method: Automatic metrics evaluation of machine translation systems, with recognition that systems using re-ranking techniques may have unfair advantage

Result: Preliminary ranking generated but not final - serves as interim guidance for participants

Conclusion: This automatic ranking is preliminary and potentially biased; the definitive WMT25 ranking will be based on human evaluation which is more reliable

Abstract: We present the preliminary ranking of the WMT25 General Machine Translation Shared Task, in which MT systems have been evaluated using automatic metrics. As this ranking is based on automatic evaluations, it may be biased in favor of systems that employ re-ranking techniques, such as Quality Estimation re-ranking or Minimum Bayes Risk decoding. The official WMT25 ranking will be based on human evaluation, which is more reliable and will supersede the automatic ranking. The purpose of this report is not to present the final findings of the General MT task, but rather to share preliminary results with task participants, which may be useful when preparing their system submission papers.

Method: Introduces an LLM-driven framework for cultural localization that automatically constructs math word problem datasets with native entities from existing sources.

Result: Shows that translated benchmarks obscure true multilingual math ability, and the framework helps mitigate English-centric entity bias while improving robustness across languages.

Conclusion: The proposed framework successfully addresses cultural localization gaps in multilingual math reasoning by generating native entity datasets, reducing bias and enhancing performance in appropriate socio-cultural contexts.

Abstract: Large language models (LLMs) have demonstrated significant capabilities in solving mathematical problems expressed in natural language. However, multilingual and culturally-grounded mathematical reasoning in low-resource languages lags behind English due to the scarcity of socio-cultural task datasets that reflect accurate native entities such as person names, organization names, and currencies. Existing multilingual benchmarks are predominantly produced via translation and typically retain English-centric entities, owing to the high cost associated with human annotater-based localization. Moreover, automated localization tools are limited, and hence, truly localized datasets remain scarce. To bridge this gap, we introduce a framework for LLM-driven cultural localization of math word problems that automatically constructs datasets with native names, organizations, and currencies from existing sources. We find that translated benchmarks can obscure true multilingual math ability under appropriate socio-cultural contexts. Through extensive experiments, we also show that our framework can help mitigate English-centric entity bias and improves robustness when native entities are introduced across various languages.

Method: Used Direct Preference Optimization (DPO) training on a programmatically curated subset of public data. Generated training pairs by translating English Wikipedia articles with two LLMs (GaMS-9B-Instruct and EuroLLM-9B-Instruct), then ranked translations using heuristics and COMET metrics.

Result: The fine-tuned model outperformed both baseline models, achieving COMET score gains of ~0.04 and ~0.02 respectively on Wikipedia translations, with more consistent avoidance of language and formatting errors.

Conclusion: DPO fine-tuning with programmatically curated preference data effectively improves translation quality in LLMs, demonstrating significant gains even with relatively few data resources.

Abstract: Large language models have emerged as effective machine translation systems. In this paper, we explore how a general instruction-tuned large language model can be improved for machine translation using relatively few easily produced data resources. Using Slovene as a use case, we improve the GaMS-9B-Instruct model using Direct Preference Optimization (DPO) training on a programmatically curated and enhanced subset of a public dataset. As DPO requires pairs of quality-ranked instances, we generated its training dataset by translating English Wikipedia articles using two LLMs, GaMS-9B-Instruct and EuroLLM-9B-Instruct. We ranked the resulting translations based on heuristics coupled with automatic evaluation metrics such as COMET. The evaluation shows that our fine-tuned model outperforms both models involved in the dataset generation. In comparison to the baseline models, the fine-tuned model achieved a COMET score gain of around 0.04 and 0.02, respectively, on translating Wikipedia articles. It also more consistently avoids language and formatting errors.

Method: Introduces MultiCoXQL dataset extension covering 5 diverse languages, proposes new parsing approach for multilingual performance, creates Compass dataset for custom input extraction, and evaluates multiple LLMs and BERT-type models across monolingual, cross-lingual, and multilingual settings.

Result: The research provides multilingual datasets and parsing methods that enable better performance across diverse languages including low-resource ones, and demonstrates effectiveness through comprehensive evaluations of various language models.

Conclusion: The proposed datasets and parsing approaches successfully address multilingual generalization challenges in ConvXAI systems, enabling better support for diverse languages and free-form custom inputs while maintaining reliable intent recognition capabilities.

Abstract: Conversational explainable artificial intelligence (ConvXAI) systems based on large language models (LLMs) have garnered considerable attention for their ability to enhance user comprehension through dialogue-based explanations. Current ConvXAI systems often are based on intent recognition to accurately identify the user’s desired intention and map it to an explainability method. While such methods offer great precision and reliability in discerning users' underlying intentions for English, a significant challenge in the scarcity of training data persists, which impedes multilingual generalization. Besides, the support for free-form custom inputs, which are user-defined data distinct from pre-configured dataset instances, remains largely limited. To bridge these gaps, we first introduce MultiCoXQL, a multilingual extension of the CoXQL dataset spanning five typologically diverse languages, including one low-resource language. Subsequently, we propose a new parsing approach aimed at enhancing multilingual parsing performance, and evaluate three LLMs on MultiCoXQL using various parsing strategies. Furthermore, we present Compass, a new multilingual dataset designed for custom input extraction in ConvXAI systems, encompassing 11 intents across the same five languages as MultiCoXQL. We conduct monolingual, cross-lingual, and multilingual evaluations on Compass, employing three LLMs of varying sizes alongside BERT-type models.

Method: Proposes SSS algorithm that uses a small draft model aligned with human preferences to predict tokens for an unaligned target model. Modifies acceptance criterion and bonus token distribution to exploit distributional shift between models.

Result: Achieves superior gold reward scores at significantly reduced inference cost in test-time weak-to-strong alignment experiments, validating both effectiveness and efficiency.

Conclusion: The SSS algorithm successfully addresses the efficiency bottleneck of test-time alignment by leveraging aligned draft models with unaligned target models, enabling practical deployment of aligned LLMs with reduced computational overhead.

Abstract: Aligning large language models (LLMs) with human preferences has become a critical step in their development. Recent research has increasingly focused on test-time alignment, where additional compute is allocated during inference to enhance LLM safety and reasoning capabilities. However, these test-time alignment techniques often incur substantial inference costs, limiting their practical application. We are inspired by the speculative sampling acceleration, which leverages a small draft model to efficiently predict future tokens, to address the efficiency bottleneck of test-time alignment. We introduce the reward-Shifted Speculative Sampling (SSS) algorithm, in which the draft model is aligned with human preferences, while the target model remains unchanged. We theoretically demonstrate that the distributional shift between the aligned draft model and the unaligned target model can be exploited to recover the RLHF optimal solution without actually obtaining it, by modifying the acceptance criterion and bonus token distribution. Our algorithm achieves superior gold reward scores at a significantly reduced inference cost in test-time weak-to-strong alignment experiments, thereby validating both its effectiveness and efficiency.

Method: A three-stage framework that: 1) decomposes answers into self-contained facts, 2) successively narrows plausible candidate matches through lexical and semantic filtering, and 3) verifies their alignment through structured entailment checks to reduce false positives/negatives while accommodating diverse phrasings.

Result: Substantial improvements in recall accuracy over strong lexical and LLM-as-a-Judge baselines when evaluated on three challenging long-form QA benchmarks using both human annotations and LLM-based judges.

Conclusion: LongRecall serves as a foundational building block for systematic recall assessment, effectively addressing the limitations of existing methods by providing structured verification that captures broader semantics while reducing errors from lexical overlap and hallucinations.

Abstract: LongRecall. The completeness of machine-generated text, ensuring that it captures all relevant information, is crucial in domains such as medicine and law and in tasks like list-based question answering (QA), where omissions can have serious consequences. However, existing recall metrics often depend on lexical overlap, leading to errors with unsubstantiated entities and paraphrased answers, while LLM-as-a-Judge methods with long holistic prompts capture broader semantics but remain prone to misalignment and hallucinations without structured verification. We introduce LongRecall, a general three-stage recall evaluation framework that decomposes answers into self-contained facts, successively narrows plausible candidate matches through lexical and semantic filtering, and verifies their alignment through structured entailment checks. This design reduces false positives and false negatives while accommodating diverse phrasings and contextual variations, serving as a foundational building block for systematic recall assessment. We evaluate LongRecall on three challenging long-form QA benchmarks using both human annotations and LLM-based judges, demonstrating substantial improvements in recall accuracy over strong lexical and LLM-as-a-Judge baselines.

Method: Proposes combining LLMs with combinatorial inference, experimenting with various prompting strategies to estimate LLM confidence values for use with symbolic inference methods.

Result: Symbolic inference on top of prompting alone consistently leads to more accurate and consistent predictions. Calibration and fine-tuning with structured prediction objectives further improve performance on challenging tasks.

Conclusion: Structured learning remains valuable in the LLM era, and combining LLMs with symbolic inference methods provides better structural consistency and accuracy for complex reasoning tasks.

Abstract: LLMs have been shown to be useful for a variety of language tasks, without requiring task-specific fine-tuning. However, these models often struggle with hallucinations and complex reasoning problems due to their autoregressive nature. We propose to address some of these issues, specifically in the area of structured prediction, by combining LLMs with combinatorial inference in an attempt to marry the predictive power of LLMs with the structural consistency provided by inference methods. We perform exhaustive experiments in an effort to understand which prompting strategies can effectively estimate LLM confidence values for use with symbolic inference, and show that, regardless of the prompting strategy, the addition of symbolic inference on top of prompting alone leads to more consistent and accurate predictions. Additionally, we show that calibration and fine-tuning using structured prediction objectives leads to increased performance for challenging tasks, showing that structured learning is still valuable in the era of LLMs.

Method: A novel domain-agnostic pipeline using layout-aware rendering with lynx and targeted LLM-based cleaning to recover math across various formats (MathJax, KaTeX, MathML), preserve structural integrity of equations and code blocks, remove boilerplate, standardize notation to LaTeX, and correct inconsistencies.

Result: Created Nemotron-CC-Math-3+ (133B tokens) and Nemotron-CC-Math-4+ (52B tokens) datasets. Nemotron-CC-Math-4+ surpasses all prior open math datasets and contains 5.5x more tokens than previous highest-quality dataset. Pretraining Nemotron-T 8B model yielded +4.8 to +12.6 gains on MATH and +4.6 to +14.3 gains on MBPP+ over baselines, while improving general-domain performance.

Conclusion: This work presents the first reliable pipeline for extracting scientific content including math from noisy web-scale data, achieving measurable gains in math, code, and general reasoning, and setting new state-of-the-art among open math pretraining corpora. Code and datasets are released to support open-source efforts.

Abstract: Pretraining large language models (LLMs) on high-quality, structured data such as mathematics and code substantially enhances reasoning capabilities. However, existing math-focused datasets built from Common Crawl suffer from degraded quality due to brittle extraction heuristics, lossy HTML-to-text conversion, and the failure to reliably preserve mathematical structure. In this work, we introduce Nemotron-CC-Math, a large-scale, high-quality mathematical corpus constructed from Common Crawl using a novel, domain-agnostic pipeline specifically designed for robust scientific text extraction. Unlike previous efforts, our pipeline recovers math across various formats (e.g., MathJax, KaTeX, MathML) by leveraging layout-aware rendering with lynx and a targeted LLM-based cleaning stage. This approach preserves the structural integrity of equations and code blocks while removing boilerplate, standardizing notation into LaTeX representation, and correcting inconsistencies. We collected a large, high-quality math corpus, namely Nemotron-CC-Math-3+ (133B tokens) and Nemotron-CC-Math-4+ (52B tokens). Notably, Nemotron-CC-Math-4+ not only surpasses all prior open math datasets-including MegaMath, FineMath, and OpenWebMath-but also contains 5.5 times more tokens than FineMath-4+, which was previously the highest-quality math pretraining dataset. When used to pretrain a Nemotron-T 8B model, our corpus yields +4.8 to +12.6 gains on MATH and +4.6 to +14.3 gains on MBPP+ over strong baselines, while also improving general-domain performance on MMLU and MMLU-Stem. We present the first pipeline to reliably extract scientific content–including math–from noisy web-scale data, yielding measurable gains in math, code, and general reasoning, and setting a new state of the art among open math pretraining corpora. To support open-source efforts, we release our code and datasets.

Method: Propose UniCoM pipeline with SWORDS algorithm that substitutes selected words with their translations while considering parts of speech, without altering sentence semantics.

Result: Constructed CS-FLEURS corpus achieves high intelligibility and naturalness, performing comparably to existing datasets on both objective and subjective metrics.

Conclusion: UniCoM approach advances CS speech technology and enables more inclusive multilingual systems by providing high-quality code-switched data.

Abstract: Code-switching (CS), the alternation between two or more languages within a single speaker’s utterances, is common in real-world conversations and poses significant challenges for multilingual speech technology. However, systems capable of handling this phenomenon remain underexplored, primarily due to the scarcity of suitable datasets. To resolve this issue, we propose Universal Code-Mixer (UniCoM), a novel pipeline for generating high-quality, natural CS samples without altering sentence semantics. Our approach utilizes an algorithm we call Substituting WORDs with Synonyms (SWORDS), which generates CS speech by replacing selected words with their translations while considering their parts of speech. Using UniCoM, we construct Code-Switching FLEURS (CS-FLEURS), a multilingual CS corpus designed for automatic speech recognition (ASR) and speech-to-text translation (S2TT). Experimental results show that CS-FLEURS achieves high intelligibility and naturalness, performing comparably to existing datasets on both objective and subjective metrics. We expect our approach to advance CS speech technology and enable more inclusive multilingual systems.

Method: Introduces LLaSO framework with three components: LLaSO-Align (12M speech-text alignment corpus), LLaSO-Instruct (13.5M multi-task instruction-tuning dataset), and LLaSO-Eval (reproducible benchmark). Builds LLaSO-Base, a 3.8B-parameter reference model trained on public data.

Result: LLaSO-Base achieves normalized score of 0.72, establishing strong reproducible baseline that surpasses comparable models. Analysis shows broader training coverage enhances performance but significant generalization gaps persist on unseen tasks, especially in pure audio scenarios.

Conclusion: LLaSO establishes foundational open standard to unify research efforts and accelerate community-driven progress in LSLMs by releasing complete stack of data, benchmarks, and models.

Abstract: The development of Large Speech-Language Models (LSLMs) has been slowed by fragmented architectures and a lack of transparency, hindering the systematic comparison and reproducibility of research. Unlike in the vision-language domain, the LSLM field suffers from the common practice of releasing model weights without their corresponding training data and configurations. To address these critical gaps, we introduce LLaSO, the first fully open, end-to-end framework for large-scale speech-language modeling. LLaSO provides the community with three essential resources: (1) LLaSO-Align, a 12M-instance speech-text alignment corpus; (2) LLaSO-Instruct, a 13.5M-instance multi-task instruction-tuning dataset; and (3) LLaSO-Eval, a reproducible benchmark for standardized evaluation. To validate our framework, we build and release LLaSO-Base, a 3.8B-parameter reference model trained exclusively on our public data. It achieves a normalized score of 0.72, establishing a strong, reproducible baseline that surpasses comparable models. Our analysis reveals that while broader training coverage enhances performance, significant generalization gaps persist on unseen tasks, particularly in pure audio scenarios. By releasing the complete stack of data, benchmarks, and models, LLaSO establishes a foundational open standard to unify research efforts and accelerate community-driven progress in LSLMs. We release the code, dataset, pretrained models, and results in https://github.com/EIT-NLP/LLaSO.

Method: The approach reduces the problem to fact verification, leverages external evidence to mitigate hallucinations, and involves generating and validating atomic assumptions for interpretable answers.

Result: Experiments with five LLMs show that incorporating retrieved evidence is beneficial, and generating/validating atomic assumptions yields further improvements while providing interpretable answers.

Conclusion: Using external evidence and atomic assumption validation effectively addresses questions with false assumptions, reducing LLM hallucinations and improving answer accuracy and interpretability.

Abstract: People often ask questions with false assumptions, a type of question that does not have regular answers. Answering such questions require first identifying the false assumptions. Large Language Models (LLMs) often generate misleading answers because of hallucinations. In this paper, we focus on identifying and answering questions with false assumptions in several domains. We first investigate to reduce the problem to fact verification. Then, we present an approach leveraging external evidence to mitigate hallucinations. Experiments with five LLMs demonstrate that (1) incorporating retrieved evidence is beneficial and (2) generating and validating atomic assumptions yields more improvements and provides an interpretable answer by specifying the false assumptions.

Method: Processes short, fixed-width 120ms windows independently, learning fundamental acoustic patterns common to all languages through supervised and self-supervised training.

Result: Achieves competitive cross-lingual performance despite fewer parameters, with strong generalization shown in zero-shot tests on UCLA Phonetic Corpus.

Conclusion: Effective universal speech processing is possible by modeling basic acoustic patterns within phoneme-length windows rather than relying on long contextual analysis.

Abstract: Universal phoneme recognition typically requires analyzing long speech segments and language-specific patterns. Many speech processing tasks require pure phoneme representations free from contextual influence, which motivated our development of CUPE - a lightweight model that captures key phoneme features in just 120 milliseconds, about one phoneme’s length. CUPE processes short, fixed-width windows independently and, despite fewer parameters than current approaches, achieves competitive cross-lingual performance by learning fundamental acoustic patterns common to all languages. Our extensive evaluation through supervised and self-supervised training on diverse languages, including zero-shot tests on the UCLA Phonetic Corpus, demonstrates strong cross-lingual generalization and reveals that effective universal speech processing is possible through modeling basic acoustic patterns within phoneme-length windows.

Method: Created MMDR-Bench dataset with 300 complex multi-turn dialogue scenarios (5-7 turns each) across 6 dimensions. Proposed CoLVLM Agent framework using iterative “memory-perception-planning-execution” cycle that enhances existing LVLMs without retraining.

Result: CoLVLM Agent achieved average human evaluation score of 4.03, outperforming GPT-4o (3.92) and Gemini 1.5 Pro (3.85). Showed superior performance in reasoning depth, instruction adherence, error suppression, and maintained robustness over extended dialogue turns.

Conclusion: The modular design and iterative approach of CoLVLM Agent effectively addresses complex multi-modal interaction challenges, demonstrating significant advantages over state-of-the-art commercial models without requiring extensive retraining.

Abstract: Despite significant advancements in Large Language Models (LLMs) and Large Vision-Language Models (LVLMs), current models still face substantial challenges in handling complex, multi-turn, and visually-grounded tasks that demand deep reasoning, sustained contextual understanding, entity tracking, and multi-step instruction following. Existing benchmarks often fall short in capturing the dynamism and intricacies of real-world multi-modal interactions, leading to issues such as context loss and visual hallucinations. To address these limitations, we introduce MMDR-Bench (Multi-Modal Dialogue Reasoning Benchmark), a novel dataset comprising 300 meticulously designed complex multi-turn dialogue scenarios, each averaging 5-7 turns and evaluated across six core dimensions including visual entity tracking and reasoning depth. Furthermore, we propose CoLVLM Agent (Contextual LVLM Agent), a holistic framework that enhances existing LVLMs with advanced reasoning and instruction following capabilities through an iterative “memory-perception-planning-execution” cycle, requiring no extensive re-training of the underlying models. Our extensive experiments on MMDR-Bench demonstrate that CoLVLM Agent consistently achieves superior performance, attaining an average human evaluation score of 4.03, notably surpassing state-of-the-art commercial models like GPT-4o (3.92) and Gemini 1.5 Pro (3.85). The framework exhibits significant advantages in reasoning depth, instruction adherence, and error suppression, and maintains robust performance over extended dialogue turns, validating the effectiveness of its modular design and iterative approach for complex multi-modal interactions.

Method: SemToken extracts contextual semantic embeddings via lightweight encoders, performs local semantic clustering to merge semantically equivalent tokens, and allocates heterogeneous token granularity based on semantic density (finer-grained in content-rich regions, coarser compression in repetitive spans).

Result: Experiments show up to 2.4× reduction in token count and 1.9× speedup on benchmarks like WikiText-103 and LongBench, with negligible or no degradation in perplexity and downstream accuracy.

Conclusion: Semantic structure offers a promising new axis for optimizing tokenization and computation in large language models, with SemToken demonstrating significant efficiency improvements while maintaining performance.

Abstract: Tokenization plays a critical role in language modeling, yet existing approaches such as Byte-Pair Encoding (BPE) or WordPiece operate purely on frequency statistics, ignoring the underlying semantic structure of text. This leads to over-tokenization of semantically redundant spans and underutilization of contextual coherence, particularly in long-context scenarios. In this work, we propose \textbf{SemToken}, a semantic-aware tokenization framework that jointly reduces token redundancy and improves computation efficiency. SemToken first extracts contextual semantic embeddings via lightweight encoders and performs local semantic clustering to merge semantically equivalent tokens. Then, it allocates heterogeneous token granularity based on semantic density, allowing finer-grained tokenization in content-rich regions and coarser compression in repetitive or low-entropy spans. SemToken can be seamlessly integrated with modern language models and attention acceleration methods. Experiments on long-context language modeling benchmarks such as WikiText-103 and LongBench show that SemToken achieves up to $2.4\times$ reduction in token count and $1.9\times$ speedup, with negligible or no degradation in perplexity and downstream accuracy. Our findings suggest that semantic structure offers a promising new axis for optimizing tokenization and computation in large language models.

Method: Fin-PRM integrates step-level and trajectory-level reward supervision for fine-grained evaluation of financial reasoning traces. It’s applied in offline/online settings for supervised fine-tuning, reinforcement learning, and test-time inference.

Result: Outperforms general-purpose PRMs and domain baselines on financial benchmarks (CFLUE, FinQA). Achieves 12.9% improvement in supervised learning, 5.2% in reinforcement learning, and 5.1% in test-time performance.

Conclusion: Domain-specialized reward modeling is valuable for aligning LLMs with expert-level financial reasoning, demonstrating significant performance gains across multiple learning paradigms.

Abstract: Process Reward Models (PRMs) have emerged as a promising framework for supervising intermediate reasoning in large language models (LLMs), yet existing PRMs are primarily trained on general or Science, Technology, Engineering, and Mathematics (STEM) domains and fall short in domain-specific contexts such as finance, where reasoning is more structured, symbolic, and sensitive to factual and regulatory correctness. We introduce \textbf{Fin-PRM}, a domain-specialized, trajectory-aware PRM tailored to evaluate intermediate reasoning steps in financial tasks. Fin-PRM integrates step-level and trajectory-level reward supervision, enabling fine-grained evaluation of reasoning traces aligned with financial logic. We apply Fin-PRM in both offline and online reward learning settings, supporting three key applications: (i) selecting high-quality reasoning trajectories for distillation-based supervised fine-tuning, (ii) providing dense process-level rewards for reinforcement learning, and (iii) guiding reward-informed Best-of-N inference at test time. Experimental results on financial reasoning benchmarks, including CFLUE and FinQA, demonstrate that Fin-PRM consistently outperforms general-purpose PRMs and strong domain baselines in trajectory selection quality. Downstream models trained with Fin-PRM yield substantial improvements with baselines, with gains of 12.9% in supervised learning, 5.2% in reinforcement learning, and 5.1% in test-time performance. These findings highlight the value of domain-specialized reward modeling for aligning LLMs with expert-level financial reasoning. Our project resources will be available at https://github.com/aliyun/qwen-dianjin.

Method: SPARK applies unstructured sparsity by pruning KV cache at the channel level and dynamically restoring pruned entries during attention computation. It’s a plug-and-play method orthogonal to existing compression techniques.

Result: SPARK reduces KV cache storage by over 30% compared to eviction-based methods while preserving or improving model accuracy. Even with 80% pruning ratio, it maintains performance with less than 5% degradation compared to baseline.

Conclusion: SPARK effectively addresses the KV cache bottleneck by leveraging channel-level sparsity, enabling longer sequence processing within the same memory budget while maintaining model performance and being compatible with existing compression techniques.

Abstract: Long-context inference in large language models (LLMs) is increasingly constrained by the KV cache bottleneck: memory usage grows linearly with sequence length, while attention computation scales quadratically. Existing approaches address this issue by compressing the KV cache along the temporal axis through strategies such as token eviction or merging to reduce memory and computational overhead. However, these methods often neglect fine-grained importance variations across feature dimensions (i.e., the channel axis), thereby limiting their ability to effectively balance efficiency and model accuracy. In reality, we observe that channel saliency varies dramatically across both queries and positions: certain feature channels carry near-zero information for a given query, while others spike in relevance. To address this oversight, we propose SPARK, a training-free plug-and-play method that applies unstructured sparsity by pruning KV at the channel level, while dynamically restoring the pruned entries during attention score computation. Notably, our approach is orthogonal to existing KV compression and quantization techniques, making it compatible for integration with them to achieve further acceleration. By reducing channel-level redundancy, SPARK enables processing of longer sequences within the same memory budget. For sequences of equal length, SPARK not only preserves or improves model accuracy but also reduces KV cache storage by over 30% compared to eviction-based methods. Furthermore, even with an aggressive pruning ratio of 80%, SPARK maintains performance with less degradation than 5% compared to the baseline eviction method, demonstrating its robustness and effectiveness. Our code will be available at https://github.com/Xnhyacinth/SparK.

Method: S2K uses internal-external knowledge self-selection strategy, selective supervised fine-tuning, structured reasoning data generation pipeline, and integrates GRPO to enhance reasoning ability.

Result: Experiments on medical, legal, and financial QA benchmarks show S2K consistently outperforms existing methods and matches domain-pretrained LLMs with significantly lower cost.

Conclusion: The proposed progressive knowledge acquisition approach effectively addresses domain-specific QA challenges while maintaining cost efficiency and cross-domain flexibility.

Abstract: Large Language Models (LLMs) perform well in general QA but often struggle in domain-specific scenarios. Retrieval-Augmented Generation (RAG) introduces external knowledge but suffers from hallucinations and latency due to noisy retrievals. Continued pretraining internalizes domain knowledge but is costly and lacks cross-domain flexibility. We attribute this challenge to the long-tail distribution of domain knowledge, which leaves partial yet useful internal knowledge underutilized. We further argue that knowledge acquisition should be progressive, mirroring human learning: first understanding concepts, then applying them to complex reasoning. To address this, we propose Selct2Know (S2K), a cost-effective framework that internalizes domain knowledge through an internal-external knowledge self-selection strategy and selective supervised fine-tuning. We also introduce a structured reasoning data generation pipeline and integrate GRPO to enhance reasoning ability. Experiments on medical, legal, and financial QA benchmarks show that S2K consistently outperforms existing methods and matches domain-pretrained LLMs with significantly lower cost.

Method: Stepwise Experience Recall (SEER) performs fine-grained, stepwise retrieval from a continually updated experience pool that incrementally grows with past successful trajectories, enabling continuous improvement without manual curation.

Result: On ToolQA benchmark: 6.1% improvement on easy questions, 4.7% on hard questions. On τ-bench with real-world domains: 7.44% gain with Qwen2.5-7B and 23.38% gain with Qwen2.5-72B models.

Conclusion: SEER provides an effective self-guided approach that continuously improves LLM tool usage performance through experience recall, demonstrating substantial accuracy gains across different model sizes and benchmarks.

Abstract: Function calling enables large language models (LLMs) to interact with external systems by leveraging tools and APIs. When faced with multi-step tool usage, LLMs still struggle with tool selection, parameter generation, and tool-chain planning. Existing methods typically rely on manually designing task-specific demonstrations, or retrieving from a curated library. These approaches demand substantial expert effort and prompt engineering becomes increasingly complex and inefficient as tool diversity and task difficulty scale. To address these challenges, we propose a self-guided method, Stepwise Experience Recall (SEER), which performs fine-grained, stepwise retrieval from a continually updated experience pool. Instead of relying on static or manually curated library, SEER incrementally augments the experience pool with past successful trajectories, enabling continuous expansion of the pool and improved model performance over time. Evaluated on the ToolQA benchmark, SEER achieves an average improvement of 6.1% on easy and 4.7% on hard questions. We further test SEER on $\tau$-bench, which includes two real-world domains. Powered by Qwen2.5-7B and Qwen2.5-72B models, SEER demonstrates substantial accuracy gains of 7.44% and 23.38%, respectively.

Method: Investigated whether checklists should be used for all questions or selectively, generated them using six methods, evaluated effectiveness across eight model sizes, and identified checklist items correlating with human evaluations through pairwise comparison and direct scoring experiments.

Result: Selective checklist use tends to improve evaluation performance in pairwise settings, while benefits are less consistent in direct scoring. Checklist items with low correlation to human scores often still reflect human-written criteria.

Conclusion: Findings highlight the need to more clearly define objective evaluation criteria to guide both human and automatic evaluations, as human evaluation shows inconsistencies that affect both manual and automated approaches.

Abstract: Automatic evaluation of generative tasks using large language models faces challenges due to ambiguous criteria. Although automatic checklist generation is a potentially promising approach, its usefulness remains underexplored. We investigate whether checklists should be used for all questions or selectively, generate them using six methods, evaluate their effectiveness across eight model sizes, and identify checklist items that correlate with human evaluations. Through experiments on pairwise comparison and direct scoring tasks, we find that selective checklist use tends to improve evaluation performance in pairwise settings, while its benefits are less consistent in direct scoring. Our analysis also shows that even checklist items with low correlation to human scores often reflect human-written criteria, indicating potential inconsistencies in human evaluation. These findings highlight the need to more clearly define objective evaluation criteria to guide both human and automatic evaluations. \footnote{Our code is available at~https://github.com/momo0817/checklist-effectiveness-study

Method: Decoupled dynamic vocabulary framework based on lexical locality principle and computational asymmetry - offloads embeddings and uses hybrid static-dynamic selection for LM heads with on-demand loading.

Result: Achieves up to 99% reduction in vocabulary-related memory usage with minimal/no performance degradation across diverse downstream tasks, significantly outperforming static pruning.

Conclusion: VocabTailor effectively addresses SLM memory constraints through dynamic vocabulary management while maintaining performance, enabling better edge deployment.

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: Created a human-authored Thai dataset through multi-stage quality control with annotators, domain experts, and AI researchers, covering four professional domains and seven task types. Conducted zero-shot evaluation and instruction tuning studies with ablations to isolate native supervision effects.

Result: Models fine-tuned on WangchanThaiInstruct outperformed those using translated data in both in-domain and out-of-domain benchmarks, showing significant performance improvements on culturally and professionally specific tasks.

Conclusion: Culturally and professionally grounded instruction data is essential for improving LLM alignment in low-resource, linguistically diverse settings, as native supervision addresses nuances that translations cannot capture.

Abstract: Large language models excel at instruction-following in English, but their performance in low-resource languages like Thai remains underexplored. Existing benchmarks often rely on translations, missing cultural and domain-specific nuances needed for real-world use. We present WangchanThaiInstruct, a human-authored Thai dataset for evaluation and instruction tuning, covering four professional domains and seven task types. Created through a multi-stage quality control process with annotators, domain experts, and AI researchers, WangchanThaiInstruct supports two studies: (1) a zero-shot evaluation showing performance gaps on culturally and professionally specific tasks, and (2) an instruction tuning study with ablations isolating the effect of native supervision. Models fine-tuned on WangchanThaiInstruct outperform those using translated data in both in-domain and out-of-domain benchmarks. These findings underscore the need for culturally and professionally grounded instruction data to improve LLM alignment in low-resource, linguistically diverse settings.

Method: Developed EMNLP framework with extended psychological scales and 88 teacher-specific moral dilemmas, tested on 12 LLMs with targeted soft prompt injection to assess compliance and vulnerability.

Result: Teacher-role LLMs show more idealized/polarized personalities than humans, excel in abstract moral reasoning but struggle with emotional complexity. Stronger reasoning models are paradoxically more vulnerable to harmful injections.

Conclusion: First benchmark for ethical/psychological alignment of teacher-role LLMs reveals capability-safety paradox, providing resources for educational AI development and evaluation.

Abstract: Simulating Professions (SP) enables Large Language Models (LLMs) to emulate professional roles. However, comprehensive psychological and ethical evaluation in these contexts remains lacking. This paper introduces EMNLP, an Educator-role Moral and Normative LLMs Profiling framework for personality profiling, moral development stage measurement, and ethical risk under soft prompt injection. EMNLP extends existing scales and constructs 88 teacher-specific moral dilemmas, enabling profession-oriented comparison with human teachers. A targeted soft prompt injection set evaluates compliance and vulnerability in teacher SP. Experiments on 12 LLMs show teacher-role LLMs exhibit more idealized and polarized personalities than human teachers, excel in abstract moral reasoning, but struggle with emotionally complex situations. Models with stronger reasoning are more vulnerable to harmful prompt injection, revealing a paradox between capability and safety. The model temperature and other hyperparameters have limited influence except in some risk behaviors. This paper presents the first benchmark to assess ethical and psychological alignment of teacher-role LLMs for educational AI. Resources are available at https://e-m-n-l-p.github.io/.

Method: Introduces Conflict-Aware RAG (CARE) with a context assessor that encodes memory token embeddings and uses grounded/adversarial soft prompting to identify unreliable context and guide reasoning.

Result: Extensive experiments show CARE effectively mitigates context-memory conflicts with 5.0% average performance improvement on QA and fact-checking benchmarks.

Conclusion: CARE establishes a promising direction for developing more trustworthy and adaptive RAG systems that can handle knowledge conflicts effectively.

Abstract: Retrieval-augmented generation (RAG) enhances the capabilities of large language models (LLMs) by incorporating external knowledge into their input prompts. However, when the retrieved context contradicts the LLM’s parametric knowledge, it often fails to resolve the conflict between incorrect external context and correct parametric knowledge, known as context-memory conflict. To tackle this problem, we introduce Conflict-Aware REtrieval-Augmented Generation (CARE), consisting of a context assessor and a base LLM. The context assessor encodes compact memory token embeddings from raw context tokens. Through grounded/adversarial soft prompting, the context assessor is trained to discern unreliable context and capture a guidance signal that directs reasoning toward the more reliable knowledge source. Extensive experiments show that CARE effectively mitigates context-memory conflicts, leading to an average performance gain of 5.0% on QA and fact-checking benchmarks, establishing a promising direction for trustworthy and adaptive RAG systems.

Method: Proposed a temporal commonsense extraction pipeline leveraging LLMs to automatically mine temporal commonsense. Constructed TComQA dataset from SAMSum and RealNews corpora, validated through crowdsourcing.

Result: TComQA achieves over 80% precision in extracting temporal commonsense. Models trained with TComQA outperform LLMs fine-tuned on existing temporal QA datasets.

Conclusion: Automated mining of temporal commonsense using LLMs enables creation of robust language models that better understand temporal context, addressing a key limitation in current NLP systems.

Abstract: Understanding events necessitates grasping their temporal context, which is often not explicitly stated in natural language. For example, it is not a trivial task for a machine to infer that a museum tour may last for a few hours, but can not take months. Recent studies indicate that even advanced large language models (LLMs) struggle in generating text that require reasoning with temporal commonsense due to its infrequent explicit mention in text. Therefore, automatically mining temporal commonsense for events enables the creation of robust language models. In this work, we investigate the capacity of LLMs to extract temporal commonsense from text and evaluate multiple experimental setups to assess their effectiveness. Here, we propose a temporal commonsense extraction pipeline that leverages LLMs to automatically mine temporal commonsense and use it to construct TComQA, a dataset derived from SAMSum and RealNews corpora. TComQA has been validated through crowdsourcing and achieves over 80% precision in extracting temporal commonsense. The model trained with TComQA also outperforms an LLM fine-tuned on existing dataset of temporal question answering task.

Method: EDAS (Evaluation based on Distance from Average Solution) - a robust meta-metric that synthesizes model performance across multiple datasets and diverse evaluation criteria into a single normalized score (0-1 range).

Result: Experimental results on benchmark datasets (FB15k-237, WN18RR) show EDAS effectively integrates multi-metric, multi-dataset performance into unified rankings, providing consistent and robust evaluation.

Conclusion: EDAS offers a global perspective for KGC model evaluation, enabling more informed model selection, promoting fairness in cross-dataset comparison, and providing a generalizable framework for comprehensive performance assessment.

Abstract: Knowledge Graphs (KGs) enable applications in various domains such as semantic search, recommendation systems, and natural language processing. KGs are often incomplete, missing entities and relations, an issue addressed by Knowledge Graph Completion (KGC) methods that predict missing elements. Different evaluation metrics, such as Mean Reciprocal Rank (MRR), Mean Rank (MR), and Hit@k, are commonly used to assess the performance of such KGC models. A major challenge in evaluating KGC models, however, lies in comparing their performance across multiple datasets and metrics. A model may outperform others on one dataset but underperform on another, making it difficult to determine overall superiority. Moreover, even within a single dataset, different metrics such as MRR and Hit@1 can yield conflicting rankings, where one model excels in MRR while another performs better in Hit@1, further complicating model selection for downstream tasks. These inconsistencies hinder holistic comparisons and highlight the need for a unified meta-metric that integrates performance across all metrics and datasets to enable a more reliable and interpretable evaluation framework. To address this need, we propose KG Evaluation based on Distance from Average Solution (EDAS), a robust and interpretable meta-metric that synthesizes model performance across multiple datasets and diverse evaluation criteria into a single normalized score ($M_i \in [0,1]$). Unlike traditional metrics that focus on isolated aspects of performance, EDAS offers a global perspective that supports more informed model selection and promotes fairness in cross-dataset evaluation. Experimental results on benchmark datasets such as FB15k-237 and WN18RR demonstrate that EDAS effectively integrates multi-metric, multi-dataset performance into a unified ranking, offering a consistent, robust, and generalizable framework for evaluating KGC models.

Method: Systematic review and categorization of 283 representative benchmarks into three main categories: general capabilities (linguistics, knowledge, reasoning), domain-specific (natural sciences, humanities, engineering), and target-specific (risks, reliability, agents).

Result: Identified key problems including score inflation from data contamination, unfair evaluation due to cultural/linguistic biases, and lack of evaluation on process credibility and dynamic environments.

Conclusion: Provides a referable design paradigm for future benchmark innovation to address current limitations and improve the quality and fairness of LLM evaluation.

Abstract: In recent years, with the rapid development of the depth and breadth of large language models’ capabilities, various corresponding evaluation benchmarks have been emerging in increasing numbers. As a quantitative assessment tool for model performance, benchmarks are not only a core means to measure model capabilities but also a key element in guiding the direction of model development and promoting technological innovation. We systematically review the current status and development of large language model benchmarks for the first time, categorizing 283 representative benchmarks into three categories: general capabilities, domain-specific, and target-specific. General capability benchmarks cover aspects such as core linguistics, knowledge, and reasoning; domain-specific benchmarks focus on fields like natural sciences, humanities and social sciences, and engineering technology; target-specific benchmarks pay attention to risks, reliability, agents, etc. We point out that current benchmarks have problems such as inflated scores caused by data contamination, unfair evaluation due to cultural and linguistic biases, and lack of evaluation on process credibility and dynamic environments, and provide a referable design paradigm for future benchmark innovation.

Method: Proposed MultiTrust-X benchmark with three-dimensional framework: 5 trustworthiness aspects (truthfulness, robustness, safety, fairness, privacy), 2 novel risk types (multimodal risks, cross-modal impacts), and various mitigation strategies from data, model architecture, training, and inference perspectives.

Result: Extensive experiments revealed significant vulnerabilities in current models, including a gap between trustworthiness and general capabilities, amplification of risks in base LLMs by multimodal training/inference, and limitations in existing mitigation strategies that often introduce trade-offs compromising model utility.

Conclusion: The findings provide practical insights for future improvements, leading to the development of Reasoning-Enhanced Safety Alignment (RESA) approach that uses chain-of-thought reasoning to discover underlying risks, achieving state-of-the-art results in balancing safety and performance.

Abstract: The trustworthiness of Multimodal Large Language Models (MLLMs) remains an intense concern despite the significant progress in their capabilities. Existing evaluation and mitigation approaches often focus on narrow aspects and overlook risks introduced by the multimodality. To tackle these challenges, we propose MultiTrust-X, a comprehensive benchmark for evaluating, analyzing, and mitigating the trustworthiness issues of MLLMs. We define a three-dimensional framework, encompassing five trustworthiness aspects which include truthfulness, robustness, safety, fairness, and privacy; two novel risk types covering multimodal risks and cross-modal impacts; and various mitigation strategies from the perspectives of data, model architecture, training, and inference algorithms. Based on the taxonomy, MultiTrust-X includes 32 tasks and 28 curated datasets, enabling holistic evaluations over 30 open-source and proprietary MLLMs and in-depth analysis with 8 representative mitigation methods. Our extensive experiments reveal significant vulnerabilities in current models, including a gap between trustworthiness and general capabilities, as well as the amplification of potential risks in base LLMs by both multimodal training and inference. Moreover, our controlled analysis uncovers key limitations in existing mitigation strategies that, while some methods yield improvements in specific aspects, few effectively address overall trustworthiness, and many introduce unexpected trade-offs that compromise model utility. These findings also provide practical insights for future improvements, such as the benefits of reasoning to better balance safety and performance. Based on these insights, we introduce a Reasoning-Enhanced Safety Alignment (RESA) approach that equips the model with chain-of-thought reasoning ability to discover the underlying risks, achieving state-of-the-art results.

Method: Information-theoretic framework using entropy and margin-based uncertainty measures from drafter’s output distribution to dynamically adjust speculative token generation length and modulate verification criteria.

Result: Significant speedups over standard speculative decoding while preserving or improving BLEU and ROUGE scores on machine translation and summarization tasks.

Conclusion: Provides a principled, plug-in method for efficient and robust decoding in large language models that adapts to varying uncertainty conditions.

Abstract: Speculative decoding has emerged as an effective approach for accelerating autoregressive inference by parallelizing token generation through a draft-then-verify paradigm. However, existing methods rely on static drafting lengths and rigid verification criteria, limiting their adaptability across varying model uncertainties and input complexities. This paper proposes an information-theoretic framework for speculative decoding based on confidence-modulated drafting. By leveraging entropy and margin-based uncertainty measures over the drafter’s output distribution, the proposed method dynamically adjusts the number of speculatively generated tokens at each iteration. This adaptive mechanism reduces rollback frequency, improves resource utilization, and maintains output fidelity. Additionally, the verification process is modulated using the same confidence signals, enabling more flexible acceptance of drafted tokens without sacrificing generation quality. Experiments on machine translation and summarization tasks demonstrate significant speedups over standard speculative decoding while preserving or improving BLEU and ROUGE scores. The proposed approach offers a principled, plug-in method for efficient and robust decoding in large language models under varying conditions of uncertainty.

Method: Conducted controlled experiments scaling vocabulary from 24K to 196K while holding data, compute, and optimization fixed. Used Kolmogorov complexity to quantify tokenized text complexity, performed word-level loss decomposition, and constrained embedding norms to test effects of token-frequency imbalance.

Result: Larger vocabularies reduce cross-entropy almost exclusively by lowering uncertainty on the 2,500 most frequent words, while loss on rare words increases. The model exploits rather than suffers from token-frequency imbalance. Same benefit achieved by enlarging model parameters with fixed vocabulary.

Conclusion: The benefit of larger vocabularies comes from lowering the complexity of tokenized text, not from better rare word handling. This provides a principled approach for tokenizer-model co-design and clarifies loss dynamics in language model scaling.

Abstract: Large language models are trained with tokenizers, and the resulting token distribution is highly imbalanced: a few words dominate the stream while most occur rarely. Recent practice favors ever-larger vocabularies, but the source of the benefit is unclear. We conduct a controlled study that scales the language model’s vocabulary from 24K to 196K while holding data, compute, and optimization fixed. We first quantify the complexity of tokenized text, formalized via Kolmogorov complexity, and show that larger vocabularies reduce this complexity. Above 24K, every common word is already a single token, so further growth mainly deepens the relative token-frequency imbalance. A word-level loss decomposition shows that larger vocabularies reduce cross-entropy almost exclusively by lowering uncertainty on the 2,500 most frequent words, even though loss on the rare tail rises. Constraining input and output embedding norms to attenuate the effect of token-frequency imbalance reverses the gain, directly showing that the model exploits rather than suffers from imbalance. Because the same frequent words cover roughly 77% of tokens in downstream benchmarks, this training advantage transfers intact. We also show that enlarging model parameters with a fixed vocabulary yields the same frequent-word benefit. Our results reframe “bigger vocabularies help” as “lowering the complexity of tokenized text helps,” providing a simple, principled lever for tokenizer-model co-design and clarifying the loss dynamics that govern language-model scaling in pre-training.

Method: The authors systematically analyzed 134 papers, introduced a unified taxonomy for evidence-based text generation with LLMs, and investigated 300 evaluation metrics across seven key dimensions, focusing on approaches using citations, attribution, or quotations.

Result: The study provides a comprehensive framework for understanding evidence-based text generation with LLMs, examining distinctive characteristics and representative methods in the field.

Conclusion: The paper highlights open challenges and outlines promising directions for future work in evidence-based text generation with LLMs, providing a foundation for more standardized research in this area.

Abstract: The increasing adoption of large language models (LLMs) has been accompanied by growing concerns regarding their reliability and trustworthiness. As a result, a growing body of research focuses on evidence-based text generation with LLMs, aiming to link model outputs to supporting evidence to ensure traceability and verifiability. However, the field is fragmented due to inconsistent terminology, isolated evaluation practices, and a lack of unified benchmarks. To bridge this gap, we systematically analyze 134 papers, introduce a unified taxonomy of evidence-based text generation with LLMs, and investigate 300 evaluation metrics across seven key dimensions. Thereby, we focus on approaches that use citations, attribution, or quotations for evidence-based text generation. Building on this, we examine the distinctive characteristics and representative methods in the field. Finally, we highlight open challenges and outline promising directions for future work.

Method: Introduces MCR-BENCH, a comprehensive benchmark with inconsistent audio-text pairs, evaluates across diverse audio understanding tasks, investigates influencing factors of text bias, explores mitigation through supervised finetuning, and analyzes model confidence patterns.

Result: LALMs display significant bias toward textual input, frequently disregarding audio evidence when inconsistencies exist, leading to substantial performance degradation in audio-centric tasks. Models show persistent overconfidence even with contradictory inputs.

Conclusion: There is a need for improved modality balance during training and more sophisticated fusion mechanisms to enhance robustness when handling conflicting multimodal inputs, as current LALMs are unreliable in real-world scenarios with inconsistent information.

Abstract: Large Audio-Language Models (LALMs) are enhanced with audio perception capabilities, enabling them to effectively process and understand multimodal inputs that combine audio and text. However, their performance in handling conflicting information between audio and text modalities remains largely unexamined. This paper introduces MCR-BENCH, the first comprehensive benchmark specifically designed to evaluate how LALMs prioritize information when presented with inconsistent audio-text pairs. Through extensive evaluation across diverse audio understanding tasks, we reveal a concerning phenomenon: when inconsistencies exist between modalities, LALMs display a significant bias toward textual input, frequently disregarding audio evidence. This tendency leads to substantial performance degradation in audio-centric tasks and raises important reliability concerns for real-world applications. We further investigate the influencing factors of text bias, and explore mitigation strategies through supervised finetuning, and analyze model confidence patterns that reveal persistent overconfidence even with contradictory inputs. These findings underscore the need for improved modality balance during training and more sophisticated fusion mechanisms to enhance the robustness when handling conflicting multi-modal inputs. The project is available at https://github.com/WangCheng0116/MCR-BENCH.

Method: Provides an in-depth overview and comprehensive study of existing privacy-preserving language modeling approaches, analyzing their methodologies and implementations.

Result: The study highlights the strengths and investigates the limitations of current privacy-preserving techniques for language models.

Conclusion: Contributes valuable insights to ongoing privacy-preserving language modeling research and outlines important future research directions to address privacy risks.

Abstract: Recent developments in language modeling have increased their use in various applications and domains. Language models, often trained on sensitive data, can memorize and disclose this information during privacy attacks, raising concerns about protecting individuals’ privacy rights. Preserving privacy in language models has become a crucial area of research, as privacy is one of the fundamental human rights. Despite its significance, understanding of how much privacy risk these language models possess and how it can be mitigated is still limited. This research addresses this by providing a comprehensive study of the privacy-preserving language modeling approaches. This study gives an in-depth overview of these approaches, highlights their strengths, and investigates their limitations. The outcomes of this study contribute to the ongoing research on privacy-preserving language modeling, providing valuable insights and outlining future research directions.

Method: Introduces a novel dataset (M-Help) specifically designed to detect help-seeking behavior, going beyond traditional labels by identifying help-seeking activity, specific mental health disorders, and underlying causes

Result: AI models trained on M-Help can perform three key tasks: identifying help-seekers, diagnosing mental health conditions, and uncovering root causes of issues

Conclusion: The M-Help dataset provides a comprehensive framework for detecting and understanding help-seeking behavior and mental health challenges on social media

Abstract: Mental health disorders are a global crisis. While various datasets exist for detecting such disorders, there remains a critical gap in identifying individuals actively seeking help. This paper introduces a novel dataset, M-Help, specifically designed to detect help-seeking behavior on social media. The dataset goes beyond traditional labels by identifying not only help-seeking activity but also specific mental health disorders and their underlying causes, such as relationship challenges or financial stressors. AI models trained on M-Help can address three key tasks: identifying help-seekers, diagnosing mental health conditions, and uncovering the root causes of issues.

Method: Library-based research analyzing different translation methods and rules for rendering non-equivalent words from source to target languages.

Result: Developed various translation approaches for non-equivalent words and successfully rendered 25 non-equivalent words from Dari and Uzbek into English and Russian languages.

Conclusion: The research provides professional methods and rules for translating culturally-specific non-equivalent words, helping bridge language gaps and reduce misunderstandings in translation work.

Abstract: These pure languages understanding directly relates to translation knowledge where linguists and translators need to work and research to eradicate misunderstanding. Misunderstandings mostly appear in non-equivalent words because there are different local and internal words like food, garment, cultural and traditional words and others in every notion. Truly, most of these words do not have equivalent in the target language and these words need to be worked and find their equivalent in the target language to fully understand the both languages. The purpose of this research is to introduce the methods of rendering non-equivalent words professionally from the source language to the target language and this research has been completed using library-based research. However, some of these non-equivalent words are already professionally rendered to the target language but still there many other words to be rendered. As a result, this research paper includes different ways and rules of rendering non-equivalent words from source language to the target language and 25 non-equvalent words have been rendered from Dar & Uzbek into English and Russian languages.

Method: PyTOD generates executable code for dialogue state tracking and employs constrained decoding using language models (instead of grammar rules) to follow API schemata, with policy and execution feedback for error correction.

Result: PyTOD achieves state-of-the-art state tracking performance on the challenging SGD benchmark, surpassing strong baselines in both accuracy and robust user goal estimation as dialogues progress.

Conclusion: The approach demonstrates the effectiveness of execution-aware state tracking for programmable task-oriented dialogue agents.

Abstract: Programmable task-oriented dialogue (TOD) agents enable language models to follow structured dialogue policies, but their effectiveness hinges on accurate state tracking. We present PyTOD, an agent that generates executable code to track dialogue state and uses policy and execution feedback for efficient error correction. To this end, PyTOD employs a simple constrained decoding approach, using a language model instead of grammar rules to follow API schemata. This leads to state-of-the-art state tracking performance on the challenging SGD benchmark. Our experiments show that PyTOD surpasses strong baselines in both accuracy and robust user goal estimation as the dialogue progresses, demonstrating the effectiveness of execution-aware state tracking.

Method: Builds on Group Relative Policy Optimization with two innovations: Sub-score Dynamic Weighting (adaptively prioritizes clinically challenging error types) and Majority-Guided Advantage Scaling (adjusts policy gradient updates based on prompt difficulty from sub-score agreement).

Result: On the ReXVal benchmark, RadReason surpasses all prior offline metrics and achieves parity with GPT-4-based evaluations while remaining explainable and cost-efficient.

Conclusion: RadReason provides a clinically grounded, interpretable, and fine-grained evaluation framework suitable for real-world clinical deployment, addressing fundamental challenges in automated radiology report assessment.

Abstract: Evaluating automatically generated radiology reports remains a fundamental challenge due to the lack of clinically grounded, interpretable, and fine-grained metrics. Existing methods either produce coarse overall scores or rely on opaque black-box models, limiting their usefulness in real-world clinical workflows. We introduce RadReason, a novel evaluation framework for radiology reports that not only outputs fine-grained sub-scores across six clinically defined error types, but also produces human-readable justifications that explain the rationale behind each score. Our method builds on Group Relative Policy Optimization and incorporates two key innovations: (1) Sub-score Dynamic Weighting, which adaptively prioritizes clinically challenging error types based on live F1 statistics; and (2) Majority-Guided Advantage Scaling, which adjusts policy gradient updates based on prompt difficulty derived from sub-score agreement. Together, these components enable more stable optimization and better alignment with expert clinical judgment. Experiments on the ReXVal benchmark show that RadReason surpasses all prior offline metrics and achieves parity with GPT-4-based evaluations, while remaining explainable, cost-efficient, and suitable for clinical deployment. Code will be released upon publication.

Method: Fine-tuned Google’s T5-Small (60M) using contrastive learning with InfoNCE loss to align customer personas with relevant offers in shared embedding space.

Result: 17% improvement in offer acceptance rate compared to supervised fine-tuning baseline on synthetic customer behavior dataset.

Conclusion: Contrastive learning objectives effectively advance personalized marketing by enhancing model generalizability through adaptive latent space reshaping.

Abstract: Personalized marketing has emerged as a pivotal strategy for enhancing customer engagement and driving business growth. Academic and industry efforts have predominantly focused on recommendation systems and personalized advertisements. Nonetheless, this facet of personalization holds significant potential for increasing conversion rates and improving customer satisfaction. Prior studies suggest that well-executed personalization strategies can boost revenue by up to 40 percent, underscoring the strategic importance of developing intelligent, data-driven approaches for offer generation. This work introduces SLM4Offer, a generative AI model for personalized offer generation, developed by fine-tuning a pre-trained encoder-decoder language model, specifically Google’s Text-to-Text Transfer Transformer (T5-Small 60M) using a contrastive learning approach. SLM4Offer employs InfoNCE (Information Noise-Contrastive Estimation) loss to align customer personas with relevant offers in a shared embedding space. A key innovation in SLM4Offer lies in the adaptive learning behaviour introduced by contrastive loss, which reshapes the latent space during training and enhances the model’s generalizability. The model is fine-tuned and evaluated on a synthetic dataset designed to simulate customer behaviour and offer acceptance patterns. Experimental results demonstrate a 17 percent improvement in offer acceptance rate over a supervised fine-tuning baseline, highlighting the effectiveness of contrastive objectives in advancing personalized marketing.

Method: Introduces HeTLM (Heterogeneity aware Training of Language Model) using clusterwise LM training with page-level tokenization and heterogeneous cluster-specific parameter sets to capture diverse user browsing behaviors.

Result: Small LMs with page-level tokenizer outperform large pretrained/finetuned LMs. HeTLM with cluster-specific parameters outperforms single LMs of same parameter count, achieving higher mean performance and lower variance in generation.

Conclusion: Cluster-specific training approaches like HeTLM are more effective than single large LMs for modeling subjective user behaviors, providing better alignment through improved mean performance and reduced variance across heterogeneous user preferences.

Abstract: A Large Language Model (LLM) offers versatility across domains and tasks, purportedly benefiting users with a wide variety of behaviors and preferences. We question this perception about an LLM when users have inherently subjective behaviors and preferences, as seen in their ubiquitous and idiosyncratic browsing of websites or apps. The sequential behavior logs of pages, thus generated, form something akin to each user’s self-constructed “language”, albeit without the structure and grammar imbued in natural languages. We ask: (i) Can a small LM represent the “language of browsing” better than a large LM? (ii) Can an LM with a single set of parameters (or, single LM) adequately capture myriad users’ heterogeneous, subjective behaviors and preferences? (iii) Can a single LM with high average performance, yield low variance in performance to make alignment good at user level? We introduce clusterwise LM training, HeTLM (Heterogeneity aware Training of Language Model), appropriate for subjective behaviors. We find that (i) a small LM trained using a page-level tokenizer outperforms large pretrained or finetuned LMs; (ii) HeTLM with heterogeneous cluster specific set of parameters outperforms a single LM of the same family, controlling for the number of parameters; and (iii) a higher mean and a lower variance in generation ensues, implying improved alignment.

Method: Using training data influence scores to sort training examples by difficulty, replacing human-centered metrics with model-centric difficulty assessment.

Result: Models trained with this curriculum approach outperform random order training by over 10 percentage points in benchmarks.

Conclusion: Curriculum learning is beneficial for language model pre-training when using model-centric difficulty metrics rather than human-centered ones.

Abstract: Curriculum learning, a training technique where data is presented to the model in order of example difficulty (e.g., from simpler to more complex documents), has shown limited success for pre-training language models. In this work, we investigate whether curriculum learning becomes competitive if we replace conventional human-centered difficulty metrics with one that more closely corresponds to example difficulty as observed during model training. Specifically, we experiment with sorting training examples by their \textit{training data influence}, a score which estimates the effect of individual training examples on the model’s output. Models trained on our curricula are able to outperform ones trained in random order by over 10 percentage points in benchmarks, confirming that curriculum learning is beneficial for language model pre-training, as long as a more model-centric notion of difficulty is adopted.

Method: Developed SLM-Bench with 11 metrics across correctness, computation, and consumption dimensions. Evaluated 15 SLMs on 9 NLP tasks using 23 datasets from 14 domains across 4 hardware configurations under controlled conditions.

Result: The benchmark reveals diverse trade-offs among SLMs - some excel in accuracy while others achieve superior energy efficiency. The evaluation provides rigorous comparisons of model effectiveness.

Conclusion: SLM-Bench sets a new standard for SLM evaluation by bridging the gap between resource efficiency and real-world applicability, with an open-source pipeline for reproducibility and further research.

Abstract: Small Language Models (SLMs) offer computational efficiency and accessibility, yet a systematic evaluation of their performance and environmental impact remains lacking. We introduce SLM-Bench, the first benchmark specifically designed to assess SLMs across multiple dimensions, including accuracy, computational efficiency, and sustainability metrics. SLM-Bench evaluates 15 SLMs on 9 NLP tasks using 23 datasets spanning 14 domains. The evaluation is conducted on 4 hardware configurations, providing a rigorous comparison of their effectiveness. Unlike prior benchmarks, SLM-Bench quantifies 11 metrics across correctness, computation, and consumption, enabling a holistic assessment of efficiency trade-offs. Our evaluation considers controlled hardware conditions, ensuring fair comparisons across models. We develop an open-source benchmarking pipeline with standardized evaluation protocols to facilitate reproducibility and further research. Our findings highlight the diverse trade-offs among SLMs, where some models excel in accuracy while others achieve superior energy efficiency. SLM-Bench sets a new standard for SLM evaluation, bridging the gap between resource efficiency and real-world applicability.

Method: Created HebID corpus with 5,536 sentences from Israeli politicians’ Facebook posts (2018-2021), manually annotated for 12 nuanced social identities. Benchmarked multilabel/single-label encoders and 2B-9B parameter LLMs.

Result: Hebrew-tuned LLMs achieved best performance (macro-F1 = 0.74). Analysis revealed differences in popularity, temporal trends, clustering patterns, and gender variations in identity expression across Facebook posts and parliamentary speeches.

Conclusion: HebID provides comprehensive foundation for studying social identities in Hebrew and serves as model for similar research in other non-English political contexts, enabling comparison between elite discourse and public identity priorities.

Abstract: Political language is deeply intertwined with social identities. While social identities are often shaped by specific cultural contexts and expressed through particular uses of language, existing datasets for group and identity detection are predominantly English-centric, single-label and focus on coarse identity categories. We introduce HebID, the first multilabel Hebrew corpus for social identity detection: 5,536 sentences from Israeli politicians’ Facebook posts (Dec 2018-Apr 2021), manually annotated for twelve nuanced social identities (e.g. Rightist, Ultra-Orthodox, Socially-oriented) grounded by survey data. We benchmark multilabel and single-label encoders alongside 2B-9B-parameter generative LLMs, finding that Hebrew-tuned LLMs provide the best results (macro-$F_1$ = 0.74). We apply our classifier to politicians’ Facebook posts and parliamentary speeches, evaluating differences in popularity, temporal trends, clustering patterns, and gender-related variations in identity expression. We utilize identity choices from a national public survey, enabling a comparison between identities portrayed in elite discourse and the public’s identity priorities. HebID provides a comprehensive foundation for studying social identities in Hebrew and can serve as a model for similar research in other non-English political contexts.

Method: Uses discrete diffusion modeling for parallel sequence refinement through iterative denoising, with AR-based LLM initialization and context-adaptive token-level noise rescheduling techniques.

Result: Consistently outperforms existing diffusion language models on general, mathematical, and coding tasks, demonstrating superior planning abilities, arbitrary-order generation, infilling capabilities, and tunable quality-speed trade-offs.

Conclusion: Dream 7B represents a significant advancement in diffusion-based language modeling, achieving state-of-the-art results through simple yet effective training techniques, with both base and instruction-tuned versions released to support further research.

Abstract: We introduce Dream 7B, the most powerful open diffusion large language model to date. Unlike autoregressive (AR) models that generate tokens sequentially, Dream 7B employs discrete diffusion modeling to refine sequences in parallel through iterative denoising. Our model consistently outperforms existing diffusion language models on general, mathematical, and coding tasks. Dream 7B demonstrates superior planning abilities and inference flexibility, including arbitrary-order generation, infilling capabilities, and tunable quality-speed trade-offs. These results are achieved through simple yet effective training techniques, including AR-based LLM initialization and context-adaptive token-level noise rescheduling. We release both Dream-Base and Dream-Instruct to facilitate further research in diffusion-based language modeling.

Method: Curated and annotated Hebrew corpus of 10,410 sentences from Knesset speeches, Facebook posts, and news outlets. Developed two-stage classification pipeline combining finetuned encoder models and decoder LLMs (DictaLM 2.0).

Result: Best model achieved F1=0.74 for binary PDD detection and macro-F1=0.67 for classification characteristics. Analysis revealed: rising PDD over 30 years, higher prevalence on social media vs parliamentary debate, greater use by male politicians, stronger tendencies among right-leaning actors, with spikes during elections and major events.

Conclusion: Automated PDD analysis is feasible and valuable for understanding democratic discourse, providing insights into patterns and trends of political delegitimization across different platforms and over time.

Abstract: We present the first large-scale computational study of political delegitimization discourse (PDD), defined as symbolic attacks on the normative validity of political entities. We curate and manually annotate a novel Hebrew-language corpus of 10,410 sentences drawn from Knesset speeches (1993-2023), Facebook posts (2018-2021), and leading news outlets, of which 1,812 instances (17.4%) exhibit PDD and 642 carry additional annotations for intensity, incivility, target type, and affective framing. We introduce a two-stage classification pipeline combining finetuned encoder models and decoder LLMs. Our best model (DictaLM 2.0) attains an F$_1$ of 0.74 for binary PDD detection and a macro-F$_1$ of 0.67 for classification of delegitimization characteristics. Applying this classifier to longitudinal and cross-platform data, we see a marked rise in PDD over three decades, higher prevalence on social media versus parliamentary debate, greater use by male than female politicians, and stronger tendencies among right-leaning actors - with pronounced spikes during election campaigns and major political events. Our findings demonstrate the feasibility and value of automated PDD analysis for understanding democratic discourse.

Method: SafetyFlow orchestrates seven specialized agents with versatile tools to automatically construct safety benchmarks, integrating human expertise into an automated pipeline while ensuring process and cost controllability.

Result: Created SafetyFlowBench dataset with 23,446 queries showing low redundancy and strong discriminative power. Evaluated 49 advanced LLMs and validated the system’s efficacy and efficiency through extensive experiments.

Conclusion: SafetyFlow provides the first fully automated benchmarking pipeline for LLM safety evaluation, significantly reducing time and resource costs while maintaining high-quality benchmark construction.

Abstract: The rapid proliferation of large language models (LLMs) has intensified the requirement for reliable safety evaluation to uncover model vulnerabilities. To this end, numerous LLM safety evaluation benchmarks are proposed. However, existing benchmarks generally rely on labor-intensive manual curation, which causes excessive time and resource consumption. They also exhibit significant redundancy and limited difficulty. To alleviate these problems, we introduce SafetyFlow, the first agent-flow system designed to automate the construction of LLM safety benchmarks. SafetyFlow can automatically build a comprehensive safety benchmark in only four days without any human intervention by orchestrating seven specialized agents, significantly reducing time and resource cost. Equipped with versatile tools, the agents of SafetyFlow ensure process and cost controllability while integrating human expertise into the automatic pipeline. The final constructed dataset, SafetyFlowBench, contains 23,446 queries with low redundancy and strong discriminative power. Our contribution includes the first fully automated benchmarking pipeline and a comprehensive safety benchmark. We evaluate the safety of 49 advanced LLMs on our dataset and conduct extensive experiments to validate our efficacy and efficiency.

Method: Introduces “Trained Miniatures” - small language models that are fine-tuned for specific, high-value applications to generate domain-specific responses.

Result: The approach enables generation of similar domain-specific responses for a fraction of the cost of using large language models.

Conclusion: Fine-tuned small language models provide a cost-effective alternative to large language models for targeted, high-value applications where computational efficiency is important.

Abstract: Large language models (LLMs) excel in text generation; however, these creative elements require heavy computation and are accompanied by a steep cost. Especially for targeted applications such as sales and marketing outreach, these costs are far from feasible. This paper introduces the concept of “Trained Miniatures” - Small Language Models(SLMs) fine-tuned for specific, high-value applications, generating similar domain-specific responses for a fraction of the cost.

Method: Propose SDGO (Self-Discrimination-Guided Optimization), a reinforcement learning framework that leverages the model’s own discrimination capabilities as reward signals to iteratively improve generation safety without requiring additional annotated data or external models.

Result: SDGO significantly improves model safety compared to prompt-based and training-based baselines, maintains helpfulness on general benchmarks, and shows robust performance against out-of-distribution jailbreaking attacks. The alignment enables enhanced generation capability with minimal discriminative samples.

Conclusion: Aligning LLMs’ discrimination and generation capabilities through self-guided optimization effectively enhances safety against jailbreaking attacks while preserving model helpfulness, providing a data-efficient approach to improving LLM safety.

Abstract: Large Language Models (LLMs) excel at various natural language processing tasks but remain vulnerable to jailbreaking attacks that induce harmful content generation. In this paper, we reveal a critical safety inconsistency: LLMs can more effectively identify harmful requests as discriminators than defend against them as generators. This insight inspires us to explore aligning the model’s inherent discrimination and generation capabilities. To this end, we propose SDGO (Self-Discrimination-Guided Optimization), a reinforcement learning framework that leverages the model’s own discrimination capabilities as a reward signal to enhance generation safety through iterative self-improvement. Our method does not require any additional annotated data or external models during the training phase. Extensive experiments demonstrate that SDGO significantly improves model safety compared to both prompt-based and training-based baselines while maintaining helpfulness on general benchmarks. By aligning LLMs’ discrimination and generation capabilities, SDGO brings robust performance against out-of-distribution (OOD) jailbreaking attacks. This alignment achieves tighter coupling between these two capabilities, enabling the model’s generation capability to be further enhanced with only a small amount of discriminative samples. Our code and datasets are available at https://github.com/NJUNLP/SDGO.

Method: Introduces SurGE benchmark with (1) test instances (topic description, expert-written survey, cited references) and (2) large-scale academic corpus of 1M+ papers as retrieval pool. Proposes automated evaluation framework measuring four dimensions: information coverage, referencing accuracy, structural organization, and content quality.

Result: Evaluation of diverse LLM-based approaches shows survey generation remains highly challenging even for advanced self-reflection frameworks, highlighting the complexity of the task.

Conclusion: The findings demonstrate the need for continued research in automated survey generation, and all code, data, and models have been open-sourced to facilitate further development in this area.

Abstract: Scientific survey articles play a vital role in summarizing research progress, yet their manual creation is becoming increasingly infeasible due to the rapid growth of academic literature. While large language models (LLMs) offer promising capabilities for automating this process, progress in this area is hindered by the absence of standardized benchmarks and evaluation protocols. To address this gap, we introduce SurGE (Survey Generation Evaluation), a new benchmark for evaluating scientific survey generation in the computer science domain. SurGE consists of (1) a collection of test instances, each including a topic description, an expert-written survey, and its full set of cited references, and (2) a large-scale academic corpus of over one million papers that serves as the retrieval pool. In addition, we propose an automated evaluation framework that measures generated surveys across four dimensions: information coverage, referencing accuracy, structural organization, and content quality. Our evaluation of diverse LLM-based approaches shows that survey generation remains highly challenging, even for advanced self-reflection frameworks. These findings highlight the complexity of the task and the necessity for continued research. We have open-sourced all the code, data, and models at: https://github.com/oneal2000/SurGE

Method: Introduces Pos2Distill framework that leverages position-induced disparity to counteract positional bias itself, with two specialized instantiations: Pos2Distill-R¹ for retrieval and Pos2Distill-R² for reasoning tasks.

Result: Achieves enhanced uniformity and significant performance gains across all contextual positions in long-context retrieval and reasoning tasks, with strong cross-task generalization between specialized systems.

Conclusion: The proposed Pos2Distill framework effectively addresses positional bias by transferring capabilities between positions, demonstrating superior performance on both retrieval and reasoning tasks while maintaining cross-task generalization.

Abstract: Positional bias (PB), manifesting as non-uniform sensitivity across different contextual locations, significantly impairs long-context comprehension and processing capabilities. While prior work seeks to mitigate PB through modifying the architectures causing its emergence, significant PB still persists. To address PB effectively, we introduce \textbf{Pos2Distill}, a position to position knowledge distillation framework. Pos2Distill transfers the superior capabilities from advantageous positions to less favorable ones, thereby reducing the huge performance gaps. The conceptual principle is to leverage the inherent, position-induced disparity to counteract the PB itself. We identify distinct manifestations of PB under \textbf{\textsc{r}}etrieval and \textbf{\textsc{r}}easoning paradigms, thereby designing two specialized instantiations: \emph{Pos2Distill-R\textsuperscript{1}} and \emph{Pos2Distill-R\textsuperscript{2}} respectively, both grounded in this core principle. By employing the Pos2Distill approach, we achieve enhanced uniformity and significant performance gains across all contextual positions in long-context retrieval and reasoning tasks. Crucially, both specialized systems exhibit strong cross-task generalization mutually, while achieving superior performance on their respective tasks.

Method: Comprehensive survey of existing Bangla stemming approaches, analyzing literature gaps, implementation accessibility issues, and evaluation methodology critiques.

Result: Identified significant research discontinuity, scarcity of reproducible implementations, and inadequate evaluation metrics in current Bangla stemming research.

Conclusion: Advocates for robust Bangla stemmer development with improved methodologies and continued research to enhance language processing for this under-resourced language.

Abstract: Bangla, the seventh most widely spoken language worldwide with 300 million native speakers, faces digital under-representation due to limited resources and lack of annotated datasets. Stemming, a critical preprocessing step in language analysis, is essential for low-resource, highly-inflectional languages like Bangla, because it can reduce the complexity of algorithms and models by significantly reducing the number of words the algorithm needs to consider. This paper conducts a comprehensive survey of stemming approaches, emphasizing the importance of handling morphological variants effectively. While exploring the landscape of Bangla stemming, it becomes evident that there is a significant gap in the existing literature. The paper highlights the discontinuity from previous research and the scarcity of accessible implementations for replication. Furthermore, it critiques the evaluation methodologies, stressing the need for more relevant metrics. In the context of Bangla’s rich morphology and diverse dialects, the paper acknowledges the challenges it poses. To address these challenges, the paper suggests directions for Bangla stemmer development. It concludes by advocating for robust Bangla stemmers and continued research in the field to enhance language analysis and processing.

Method: Introduced EcomMMMU dataset with 406,190 samples and 8.9M images across 8 tasks, then proposed SUMEI - a data-driven method that predicts visual utilities before using images for downstream tasks.

Result: Analysis showed product images do not consistently improve performance and can degrade it, indicating MLLMs struggle with rich visual content. SUMEI demonstrated effectiveness and robustness in comprehensive experiments.

Conclusion: Strategic image selection via utility prediction (SUMEI) is necessary for effective multimodal understanding in e-commerce, as indiscriminate use of product images can harm performance.

Abstract: E-commerce platforms are rich in multimodal data, featuring a variety of images that depict product details. However, this raises an important question: do these images always enhance product understanding, or can they sometimes introduce redundancy or degrade performance? Existing datasets are limited in both scale and design, making it difficult to systematically examine this question. To this end, we introduce EcomMMMU, an e-commerce multimodal multitask understanding dataset with 406,190 samples and 8,989,510 images. EcomMMMU is comprised of multi-image visual-language data designed with 8 essential tasks and a specialized VSS subset to benchmark the capability of multimodal large language models (MLLMs) to effectively utilize visual content. Analysis on EcomMMMU reveals that product images do not consistently improve performance and can, in some cases, degrade it. This indicates that MLLMs may struggle to effectively leverage rich visual content for e-commerce tasks. Building on these insights, we propose SUMEI, a data-driven method that strategically utilizes multiple images via predicting visual utilities before using them for downstream tasks. Comprehensive experiments demonstrate the effectiveness and robustness of SUMEI. The data and code are available through https://anonymous.4open.science/r/submission25.

Method: End-to-end reinforcement learning training framework that treats LLM as core agent and retrieval corpus as environment, with tailored rewards for format, retrieval, reasoning structure, and diagnostic accuracy. Uses large-scale medical retrieval corpus with patient records and reliable medical knowledge.

Result: Consistently outperforms prompt-engineering and training-free RAG approaches across multiple data centers. Achieves substantial gains in diagnostic accuracy, surpassing GPT-4o, DeepSeek-R1, and other medical-specific frameworks for both common and rare diseases in in-distribution and out-of-distribution settings.

Conclusion: The approach demonstrates significant improvements in diagnostic policy through reinforcement learning training, with ablation studies confirming critical roles of reward design and retrieval corpus components, providing more reliable and precise preliminary diagnoses for clinicians.

Abstract: Accurate diagnosis with medical large language models is hindered by knowledge gaps and hallucinations. Retrieval and tool-augmented methods help, but their impact is limited by weak use of external knowledge and poor feedback-reasoning traceability. To address these challenges, We introduce Deep-DxSearch, an agentic RAG system trained end-to-end with reinforcement learning (RL) that enables steer tracebale retrieval-augmented reasoning for medical diagnosis. In Deep-DxSearch, we first construct a large-scale medical retrieval corpus comprising patient records and reliable medical knowledge sources to support retrieval-aware reasoning across diagnostic scenarios. More crutially, we frame the LLM as the core agent and the retrieval corpus as its environment, using tailored rewards on format, retrieval, reasoning structure, and diagnostic accuracy, thereby evolving the agentic RAG policy from large-scale data through RL. Experiments demonstrate that our end-to-end agentic RL training framework consistently outperforms prompt-engineering and training-free RAG approaches across multiple data centers. After training, Deep-DxSearch achieves substantial gains in diagnostic accuracy, surpassing strong diagnostic baselines such as GPT-4o, DeepSeek-R1, and other medical-specific frameworks for both common and rare disease diagnosis under in-distribution and out-of-distribution settings. Moreover, ablation studies on reward design and retrieval corpus components confirm their critical roles, underscoring the uniqueness and effectiveness of our approach compared with traditional implementations. Finally, case studies and interpretability analyses highlight improvements in Deep-DxSearch’s diagnostic policy, providing deeper insight into its performance gains and supporting clinicians in delivering more reliable and precise preliminary diagnoses. See https://github.com/MAGIC-AI4Med/Deep-DxSearch.

Method: Introduces ReasonZoo benchmark with nine diverse reasoning categories to evaluate TIR effectiveness. Proposes two novel metrics: Performance-Aware Cost (PAC) and Area Under the Performance-Cost Curve (AUC-PCC) to assess reasoning efficiency.

Result: TIR-enabled models consistently outperform non-TIR counterparts in both mathematical and non-mathematical tasks. TIR enhances reasoning efficiency with improved PAC and AUC-PCC scores, indicating reduced overthinking and more streamlined reasoning.

Conclusion: TIR provides domain-general benefits and has strong potential to advance LLM capabilities in complex reasoning tasks by improving both performance and reasoning efficiency.

Abstract: Large Language Models (LLMs) have made significant strides in reasoning tasks through methods like chain-of-thought (CoT) reasoning. However, they often fall short in tasks requiring precise computations. Tool-Integrated Reasoning (TIR) has emerged as a solution by incorporating external tools into the reasoning process. Nevertheless, the generalization of TIR in improving the reasoning ability of LLM is still unclear. Additionally, whether TIR has improved the model’s reasoning behavior and helped the model think remains to be studied. We introduce ReasonZoo, a comprehensive benchmark encompassing nine diverse reasoning categories, to evaluate the effectiveness of TIR across various domains. Additionally, we propose two novel metrics, Performance-Aware Cost (PAC) and Area Under the Performance-Cost Curve (AUC-PCC), to assess reasoning efficiency. Our empirical evaluation demonstrates that TIR-enabled models consistently outperform their non-TIR counterparts in both mathematical and non-mathematical tasks. Furthermore, TIR enhances reasoning efficiency, as evidenced by improved PAC and AUC-PCC, indicating reduced overthinking and more streamlined reasoning. These findings underscore the domain-general benefits of TIR and its potential to advance LLM capabilities in complex reasoning tasks.

Method: Created 101 carefully curated real-world queries through iterative LLM rewriting and manual review, requiring coordinated use of multiple MCP tools including web search, file operations, mathematical reasoning, and data analysis. Introduced novel evaluation using ground-truth execution plans.

Result: Experiments show even frontier LLMs achieve success rate below 60%, with detailed ablations revealing distinct failure modes and inefficiencies in token usage.

Conclusion: LiveMCP-101 sets a rigorous standard for evaluating real-world agent capabilities and points to concrete directions for advancing current models toward autonomous AI systems that reliably execute complex tasks through tool use.

Abstract: Tool calling has emerged as a critical capability for AI agents to interact with the real world and solve complex tasks. While the Model Context Protocol (MCP) provides a powerful standardized framework for tool integration, there is a significant gap in benchmarking how well AI agents can effectively solve multi-step tasks using diverse MCP tools in realistic, dynamic scenarios. In this work, we present LiveMCP-101, a benchmark of 101 carefully curated real-world queries, refined through iterative LLM rewriting and manual review, that require coordinated use of multiple MCP tools including web search, file operations, mathematical reasoning, and data analysis. Moreover, we introduce a novel evaluation approach that leverages ground-truth execution plans rather than raw API outputs, better reflecting the evolving nature of real-world environments. Experiments show that even frontier LLMs achieve a success rate below 60%, highlighting major challenges in tool orchestration. Detailed ablations and error analysis further reveal distinct failure modes and inefficiencies in token usage, pointing to concrete directions for advancing current models. LiveMCP-101 sets a rigorous standard for evaluating real-world agent capabilities, advancing toward autonomous AI systems that reliably execute complex tasks through tool use.

Method: Four-step unplug-and-play process: (1) receive user request, (2) identify corresponding task expert using attribution methods and prompt tuning, (3) perform inference with expert-localized model, (4) restore original model for next task.

Result: Achieves up to 1.73x inference speed-up with 65% pruning rate across five NLU benchmarks without compromising accuracy. Effective task expert identification validated through comparisons and ablation studies.

Conclusion: DoE provides a practical and scalable framework for efficient task-specific inference that works with any transformer-based architecture, offering significant computational savings while preserving performance.

Abstract: Enabled by large-scale text corpora with huge parameters, pre-trained language models operate as multi-task experts using a single model architecture. However, recent studies have revealed that certain neurons play disproportionately important roles in solving specific tasks, suggesting that task-relevant substructures can be isolated and selectively activated for each task. Therefore, we introduce Decomposition of Experts (DoE), a novel framework that dynamically identifies and activates task-specific experts within a language model to reduce inference cost without sacrificing accuracy. We first define a task expert as a set of parameters that significantly influence the performance of a specific task and propose a four-step unplug-and-play process: (1) receiving a user request, (2) identifying the corresponding task expert, (3) performing inference using the expert-localized model, and (4) restoring the original model and waiting for the next task. Using attribution methods and prompt tuning, DoE isolates task-relevant neurons, minimizing computational overhead while maintaining task performance. We assume a setting where a language model receives user requests from five widely used natural language understanding benchmarks, processing one task at a time. In this setup, we demonstrate that DoE achieves up to a x1.73 inference speed-up with a 65% pruning rate, without compromising accuracy. Comparisons with various task expert localization methods reveal that DoE effectively identifies task experts, while ablation studies validate the importance of its components. Additionally, we analyze the effects of batch size, token count, and layer types on inference speed-up, providing practical insights for adopting DoE. The proposed framework is both practical and scalable, applicable to any transformer-based architecture, offering a robust solution for efficient task-specific inference.

Method: Proposes a four-level categorization based on instance types, conducts comprehensive survey of 150+ papers, and creates unified taxonomy covering definitions, resources, methods, and applications with focus on entity and event levels.

Result: Provides first systematic framework for understanding conceptualization, clarifies terminology scope, and organizes diverse research into coherent taxonomy to advance the field.

Conclusion: The survey addresses gaps in conceptualization research, offers structured framework for future work, and aims to stimulate more community attention to this important area of AI reasoning.

Abstract: Conceptualization, a fundamental element of human cognition, plays a pivotal role in human generalizable reasoning. Generally speaking, it refers to the process of sequentially abstracting specific instances into higher-level concepts and then forming abstract knowledge that can be applied in unfamiliar or novel situations. This enhances models’ inferential capabilities and supports the effective transfer of knowledge across various domains. Despite its significance, the broad nature of this term has led to inconsistencies in understanding conceptualization across various works, as there exists different types of instances that can be abstracted in a wide variety of ways. There is also a lack of a systematic overview that comprehensively examines existing works on the definition, execution, and application of conceptualization to enhance reasoning tasks. In this paper, we address these gaps by first proposing a categorization of different types of conceptualizations into four levels based on the types of instances being conceptualized, in order to clarify the term and define the scope of our work. Then, we present the first comprehensive survey of over 150 papers, surveying various definitions, resources, methods, and downstream applications related to conceptualization into a unified taxonomy, with a focus on the entity and event levels. Furthermore, we shed light on potential future directions in this field and hope to garner more attention from the community.

Method: Trained on dataset with ~60% non-English data, used custom multilingual tokenizer, detailed development principles including data composition and training methodologies.

Result: Strong performance across multilingual benchmarks, particularly on European versions of ARC, HellaSwag, and TruthfulQA.

Conclusion: Successfully developed multilingual LLMs that effectively support all 24 official EU languages, demonstrating viability for diverse European language processing.

Abstract: We present two multilingual LLMs, Teuken 7B-base and Teuken 7B-instruct, designed to embrace Europe’s linguistic diversity by supporting all 24 official languages of the European Union. Trained on a dataset comprising around 60% non-English data and utilizing a custom multilingual tokenizer, our models address the limitations of existing LLMs that predominantly focus on English or a few high-resource languages. We detail the models’ development principles, i.e., data composition, tokenizer optimization, and training methodologies. The models demonstrate strong performance across multilingual benchmarks, as evidenced by their performance on European versions of ARC, HellaSwag, and TruthfulQA.

Method: Fine-tuned 13 freely-available pre-trained language models on free-text notes from 246,473 manually-coded veterinary patient visits from CSU Veterinary Teaching Hospital’s EHRs to infer all 7,739 SNOMED-CT diagnosis codes.

Result: Superior performance relative to previous efforts (DeepTag and VetTag). Best results obtained when expansive labeled data were used to fine-tune large clinical language models, but comparable results achievable with limited resources and non-clinical models.

Conclusion: The study improves veterinary EHR quality through accessible automated coding methods and supports integrated health databases spanning species and institutions, benefiting both animal and human health research.

Abstract: Veterinary medical records represent a large data resource for application to veterinary and One Health clinical research efforts. Use of the data is limited by interoperability challenges including inconsistent data formats and data siloing. Clinical coding using standardized medical terminologies enhances the quality of medical records and facilitates their interoperability with veterinary and human health records from other sites. Previous studies, such as DeepTag and VetTag, evaluated the application of Natural Language Processing (NLP) to automate veterinary diagnosis coding, employing long short-term memory (LSTM) and transformer models to infer a subset of Systemized Nomenclature of Medicine - Clinical Terms (SNOMED-CT) diagnosis codes from free-text clinical notes. This study expands on these efforts by incorporating all 7,739 distinct SNOMED-CT diagnosis codes recognized by the Colorado State University (CSU) Veterinary Teaching Hospital (VTH) and by leveraging the increasing availability of pre-trained language models (LMs). 13 freely-available pre-trained LMs were fine-tuned on the free-text notes from 246,473 manually-coded veterinary patient visits included in the CSU VTH’s electronic health records (EHRs), which resulted in superior performance relative to previous efforts. The most accurate results were obtained when expansive labeled data were used to fine-tune relatively large clinical LMs, but the study also showed that comparable results can be obtained using more limited resources and non-clinical LMs. The results of this study contribute to the improvement of the quality of veterinary EHRs by investigating accessible methods for automated coding and support both animal and human health research by paving the way for more integrated and comprehensive health databases that span species and institutions.

Method: Developed a framework with multidimensional metrics to evaluate LLMs on three key tasks: reference generation, literature summary, and review composition, measuring hallucination rates, semantic coverage, and factual consistency against human-written benchmarks.

Result: Experimental results show that even the most advanced LLMs still produce hallucinated references, and model performance varies significantly across different academic disciplines when writing literature reviews.

Conclusion: The findings demonstrate current limitations in LLM reliability for academic literature review automation, highlighting the need for further research and development to improve their performance and trustworthiness in this domain.

Abstract: Large language models (LLMs) have emerged as a potential solution to automate the complex processes involved in writing literature reviews, such as literature collection, organization, and summarization. However, it is yet unclear how good LLMs are at automating comprehensive and reliable literature reviews. This study introduces a framework to automatically evaluate the performance of LLMs in three key tasks of literature writing: reference generation, literature summary, and literature review composition. We introduce multidimensional evaluation metrics that assess the hallucination rates in generated references and measure the semantic coverage and factual consistency of the literature summaries and compositions against human-written counterparts. The experimental results reveal that even the most advanced models still generate hallucinated references, despite recent progress. Moreover, we observe that the performance of different models varies across disciplines when it comes to writing literature reviews. These findings highlight the need for further research and development to improve the reliability of LLMs in automating academic literature reviews.

Method: Proposes contextualized equivariant position encoding (TAPE) that incorporates sequence content across layers, enforces permutation and orthogonal equivariance for stability, and can be integrated into pre-trained transformers with parameter-efficient fine-tuning.

Result: TAPE achieves superior performance in language modeling, arithmetic reasoning, and long-context retrieval tasks compared to existing positional embedding techniques, while facilitating LLM reasoning ability by emulating a broader class of algorithms.

Conclusion: TAPE provides a flexible, context-aware positional encoding framework that overcomes limitations of traditional fixed patterns, improves long-context ability, and can be easily integrated into existing transformer architectures with minimal overhead.

Abstract: Transformers rely on both content-based and position-based addressing mechanisms to make predictions, but existing positional encoding techniques often diminish the effectiveness of position-based addressing. Many current methods enforce rigid patterns in attention maps, limiting the ability to model long-range dependencies and adapt to diverse tasks. Additionally, most positional encodings are learned as general biases, lacking the specialization required for different instances within a dataset. To address this, we propose con\textbf{T}extualized equivari\textbf{A}nt \textbf{P}osition \textbf{E}ncoding (\textbf{TAPE}), a novel framework that enhances positional embeddings by incorporating sequence content across layers. TAPE introduces dynamic, context-aware positional encodings, overcoming the constraints of traditional fixed patterns. We show that TAPE can provably facilitate LLM reasoning ability by emulating a broader class of algorithms. By enforcing permutation and orthogonal equivariance, TAPE ensures the stability of positional encodings during updates, improving long-context ability. Our method can be easily integrated into pre-trained transformers, offering parameter-efficient fine-tuning with minimal overhead. Extensive experiments show that TAPE achieves superior performance in language modeling, arithmetic reasoning, and long-context retrieval tasks compared to existing positional embedding techniques. Code is available at https://github.com/VITA-Group/TAPE.

Method: Used a computer-assisted approach to link individual items from 17 object naming datasets to unified concepts, covering 30 languages from 10 language families.

Result: Created a comparative dataset that allows searching for recurring concepts across datasets and comparing conceptual spaces with classical basic vocabulary lists from historical linguistics.

Conclusion: The findings provide a basis for enhancing cross-linguistic object naming research and serve as guidelines for future studies dealing with object naming tasks.

Abstract: Object naming - the act of identifying an object with a word or a phrase - is a fundamental skill in interpersonal communication, relevant to many disciplines, such as psycholinguistics, cognitive linguistics, or language and vision research. Object naming datasets, which consist of concept lists with picture pairings, are used to gain insights into how humans access and select names for objects in their surroundings and to study the cognitive processes involved in converting visual stimuli into semantic concepts. Unfortunately, object naming datasets often lack transparency and have a highly idiosyncratic structure. Our study tries to make current object naming data transparent and comparable by using a multilingual, computer-assisted approach that links individual items of object naming lists to unified concepts. Our current sample links 17 object naming datasets that cover 30 languages from 10 different language families. We illustrate how the comparative dataset can be explored by searching for concepts that recur across the majority of datasets and comparing the conceptual spaces of covered object naming datasets with classical basic vocabulary lists from historical linguistics and linguistic typology. Our findings can serve as a basis for enhancing cross-linguistic object naming research and as a guideline for future studies dealing with object naming tasks.

Method: SPO uses pairwise output comparisons evaluated by an LLM evaluator to select superior prompts, followed by an LLM optimizer that aligns outputs with task requirements, all without requiring external references.

Result: Extensive experiments show SPO outperforms state-of-the-art prompt optimization methods, achieving comparable or superior results with significantly lower costs (1.1% to 5.6% of existing methods) and fewer samples (e.g., three samples).

Conclusion: SPO provides a cost-efficient framework for prompt optimization that works for both closed and open-ended tasks without external references, demonstrating practical applicability in real-world scenarios where reference data is scarce.

Abstract: Well-designed prompts are crucial for enhancing Large language models’ (LLMs) reasoning capabilities while aligning their outputs with task requirements across diverse domains. However, manually designed prompts require expertise and iterative experimentation. While existing prompt optimization methods aim to automate this process, they rely heavily on external references such as ground truth or by humans, limiting their applicability in real-world scenarios where such data is unavailable or costly to obtain. To address this, we propose Self-Supervised Prompt Optimization (SPO), a cost-efficient framework that discovers effective prompts for both closed and open-ended tasks without requiring external reference. Motivated by the observations that prompt quality manifests directly in LLM outputs and LLMs can effectively assess adherence to task requirements, we derive evaluation and optimization signals purely from output comparisons. Specifically, SPO selects superior prompts through pairwise output comparisons evaluated by an LLM evaluator, followed by an LLM optimizer that aligns outputs with task requirements. Extensive experiments demonstrate that SPO outperforms state-of-the-art prompt optimization methods, achieving comparable or superior results with significantly lower costs (e.g., 1.1% to 5.6% of existing methods) and fewer samples (e.g., three samples). The code is available at https://github.com/FoundationAgents/SPO.

Method: Adaptive Critique Refinement (ACR) with composite scoring (LLM-as-a-Judge) and selective critique strategy (LLM-as-a-Critic) to evaluate and improve low-quality code responses.

Result: RefineCoder series shows continuous performance improvement on multiple benchmarks, achieving comparable or superior performance to same-size baselines using less data.

Conclusion: ACR provides an effective approach for code LLMs to self-refine through iterative critique, reducing reliance on teacher models and improving efficiency.

Abstract: Code generation has attracted increasing attention with the rise of Large Language Models (LLMs). Many studies have developed powerful code LLMs by synthesizing code-related instruction data and applying supervised fine-tuning. However, these methods are limited by teacher model distillation and ignore the potential of iterative refinement by self-generated code. In this paper, we propose Adaptive Critique Refinement (ACR), which enables the model to refine itself by self-generated code and external critique, rather than directly imitating the code responses of the teacher model. Concretely, ACR includes a composite scoring system with LLM-as-a-Judge to evaluate the quality of code responses and a selective critique strategy with LLM-as-a-Critic to critique self-generated low-quality code responses. We develop the RefineCoder series by iteratively applying ACR, achieving continuous performance improvement on multiple code generation benchmarks. Compared to the baselines of the same size, our proposed RefineCoder series can achieve comparable or even superior performance using less data.

Method: Ontology-Guided Reverse Thinking (ORT) framework with three phases: 1) LLM extracts purpose and condition labels, 2) constructs label reasoning paths based on KG ontology, 3) uses paths to guide knowledge retrieval.

Result: ORT achieves state-of-the-art performance on WebQSP and CWQ datasets, significantly enhancing LLMs’ capability for KGQA.

Conclusion: The reverse thinking approach with ontology guidance effectively addresses multi-hop reasoning challenges in KGQA, demonstrating superior performance over existing methods.

Abstract: Large language models (LLMs) have shown remarkable capabilities in natural language processing. However, in knowledge graph question answering tasks (KGQA), there remains the issue of answering questions that require multi-hop reasoning. Existing methods rely on entity vector matching, but the purpose of the question is abstract and difficult to match with specific entities. As a result, it is difficult to establish reasoning paths to the purpose, which leads to information loss and redundancy. To address this issue, inspired by human reverse thinking, we propose Ontology-Guided Reverse Thinking (ORT), a novel framework that constructs reasoning paths from purposes back to conditions. ORT operates in three key phases: (1) using LLM to extract purpose labels and condition labels, (2) constructing label reasoning paths based on the KG ontology, and (3) using the label reasoning paths to guide knowledge retrieval. Experiments on the WebQSP and CWQ datasets show that ORT achieves state-of-the-art performance and significantly enhances the capability of LLMs for KGQA.

Method: Developed Pub-Guard-LLM with three application modes: vanilla reasoning, retrieval-augmented generation, and multi-agent debate, all providing textual explanations. Created PubMed Retraction benchmark with 11K+ biomedical articles and retraction labels.

Result: Pub-Guard-LLM consistently outperformed various baselines across all modes and provided more relevant and coherent explanations as evaluated by multiple assessment methods.

Conclusion: Pub-Guard-LLM enhances both detection performance and explainability in scientific fraud detection, offering a novel open-source tool to safeguard research integrity.

Abstract: A significant and growing number of published scientific articles is found to involve fraudulent practices, posing a serious threat to the credibility and safety of research in fields such as medicine. We propose Pub-Guard-LLM, the first large language model-based system tailored to fraud detection of biomedical scientific articles. We provide three application modes for deploying Pub-Guard-LLM: vanilla reasoning, retrieval-augmented generation, and multi-agent debate. Each mode allows for textual explanations of predictions. To assess the performance of our system, we introduce an open-source benchmark, PubMed Retraction, comprising over 11K real-world biomedical articles, including metadata and retraction labels. We show that, across all modes, Pub-Guard-LLM consistently surpasses the performance of various baselines and provides more reliable explanations, namely explanations which are deemed more relevant and coherent than those generated by the baselines when evaluated by multiple assessment methods. By enhancing both detection performance and explainability in scientific fraud detection, Pub-Guard-LLM contributes to safeguarding research integrity with a novel, effective, open-source tool.

Method: Proposed mixed models to account for random effects, used pseudo-perplexity weights for template-derived sentences, and employed statistical effect sizes for bias quantification across seven MLMs.

Result: MLMs show varying bias patterns - ALBERT unbiased for binary gender but most biased for non-binary, RoBERTa-large most biased for binary gender but minimal bias for non-binary. Some alignment found with psychological findings on personality dimensions.

Conclusion: The systematic approach effectively quantifies bias, revealing complex patterns across MLMs and some alignment with human psychological perspectives, though character trait comparisons remain limited due to lack of human studies.

Abstract: There has been significant prior work using templates to study bias against demographic attributes in MLMs. However, these have limitations: they overlook random variability of templates and target concepts analyzed, assume equality amongst templates, and overlook bias quantification. Addressing these, we propose a systematic statistical approach to assess bias in MLMs, using mixed models to account for random effects, pseudo-perplexity weights for sentences derived from templates and quantify bias using statistical effect sizes. Replicating prior studies, we match on bias scores in magnitude and direction with small to medium effect sizes. Next, we explore the novel problem of gender bias in the context of $\emph{personality}$ and $\textit{character}$ traits, across seven MLMs (base and large). We find that MLMs vary; ALBERT is unbiased for binary gender but the most biased for non-binary $\textit{neo}$, while RoBERTa-large is the most biased for binary gender but shows small to no bias for $\textit{neo}$. There is some alignment of MLM bias and findings in psychology (human perspective) - in $\textit{agreeableness}$ with RoBERTa-large and $\textit{emotional stability}$ with BERT-large. There is general agreement for the remaining 3 personality dimensions: both sides observe at most small differences across gender. For character traits, human studies on gender bias are limited thus comparisons are not feasible.

Method: Leveraging Large Language Models to generate synthetic cyberbullying data and labels, and using LLMs to label authentic unlabeled data for training BERT classifiers.

Result: BERT classifiers trained on synthetic data achieved 75.8% accuracy (vs 81.5% on authentic data), and LLM-labeled authentic data achieved 79.1% accuracy.

Conclusion: LLMs provide a scalable, ethical, and cost-effective solution for generating cyberbullying detection data, demonstrating comparable performance to human-annotated datasets.

Abstract: Cyberbullying (CB) presents a pressing threat, especially to children, underscoring the urgent need for robust detection systems to ensure online safety. While large-scale datasets on online abuse exist, there remains a significant gap in labeled data that specifically reflects the language and communication styles used by children. The acquisition of such data from vulnerable populations, such as children, is challenging due to ethical, legal and technical barriers. Moreover, the creation of these datasets relies heavily on human annotation, which not only strains resources but also raises significant concerns due to annotators exposure to harmful content. In this paper, we address these challenges by leveraging Large Language Models (LLMs) to generate synthetic data and labels. Our experiments demonstrate that synthetic data enables BERT-based CB classifiers to achieve performance close to that of those trained on fully authentic datasets (75.8% vs. 81.5% accuracy). Additionally, LLMs can effectively label authentic yet unlabeled data, allowing BERT classifiers to attain a comparable performance level (79.1% vs. 81.5% accuracy). These results highlight the potential of LLMs as a scalable, ethical, and cost-effective solution for generating data for CB detection.

Method: Proposed Pragmatic Inference Chain (PIC) prompting method based on cognitive science and linguistics findings, tested on authentic toxic interactions verified by human annotators as inference-intensive.

Result: PIC prompting significantly improved success rates for GPT-4o, Llama-3.1-70B-Instruct, DeepSeek-v2.5, and DeepSeek-v3 in identifying implicit toxic language compared to five baseline prompts including CoT and rule-based methods.

Conclusion: PIC facilitates more explicit and coherent reasoning processes in LLMs and shows potential for generalization to other inference-intensive tasks like understanding humor and metaphors.

Abstract: The rapid development of large language models (LLMs) gives rise to ethical concerns about their performance, while opening new avenues for developing toxic language detection techniques. However, LLMs’ unethical output and their capability of detecting toxicity have primarily been tested on language data that do not demand complex meaning inference, such as the biased associations of ‘he’ with programmer and ‘she’ with household. Nowadays toxic language adopts a much more creative range of implicit forms, thanks to advanced censorship. In this study, we collect authentic toxic interactions that evade online censorship and that are verified by human annotators as inference-intensive. To evaluate and improve LLMs’ reasoning of the authentic implicit toxic language, we propose a new prompting method, Pragmatic Inference Chain (PIC), drawn on interdisciplinary findings from cognitive science and linguistics. The PIC prompting significantly improves the success rate of GPT-4o, Llama-3.1-70B-Instruct, DeepSeek-v2.5, and DeepSeek-v3 in identifying implicit toxic language, compared to five baseline prompts, such as CoT and rule-based baselines. In addition, it also facilitates the models to produce more explicit and coherent reasoning processes, hence can potentially be generalized to other inference-intensive tasks, e.g., understanding humour and metaphors.

Method: Developed an advanced database with improvements in data handling, selection, and presentation, including phonetic transcriptions for all word forms and more balanced sampling across language families.

Result: The new database provides a more balanced sample covering more language families worldwide with enhanced data quality compared to previous versions.

Conclusion: The new Database of Cross-Linguistic Colexifications has the potential to inspire exciting new studies linking cross-linguistic data to open questions in linguistic typology, historical linguistics, psycholinguistics, and computational linguistics.

Abstract: Lexical resources are crucial for cross-linguistic analysis and can provide new insights into computational models for natural language learning. Here, we present an advanced database for comparative studies of words with multiple meanings, a phenomenon known as colexification. The new version includes improvements in the handling, selection and presentation of the data. We compare the new database with previous versions and find that our improvements provide a more balanced sample covering more language families worldwide, with enhanced data quality, given that all word forms are provided in phonetic transcription. We conclude that the new Database of Cross-Linguistic Colexifications has the potential to inspire exciting new studies that link cross-linguistic data to open questions in linguistic typology, historical linguistics, psycholinguistics, and computational linguistics.

Method: Investigates two approaches: 1) using LLMs as zero-shot ADL recognition models to avoid labeled data collection, and 2) using LLMs to automate explanation generation for existing XAI approaches when training data is available.

Result: The paper provides a critical evaluation of benefits and challenges of using LLMs for explainable ADL recognition.

Conclusion: LLMs show promise for enhancing explainable ADL recognition through zero-shot capabilities and automated explanation generation, though challenges remain.

Abstract: Explainable Artificial Intelligence (XAI) aims to uncover the inner reasoning of machine learning models. In IoT systems, XAI improves the transparency of models processing sensor data from multiple heterogeneous devices, ensuring end-users understand and trust their outputs. Among the many applications, XAI has also been applied to sensor-based Activities of Daily Living (ADLs) recognition in smart homes. Existing approaches highlight which sensor events are most important for each predicted activity, using simple rules to convert these events into natural language explanations for non-expert users. However, these methods produce rigid explanations lacking natural language flexibility and are not scalable. With the recent rise of Large Language Models (LLMs), it is worth exploring whether they can enhance explanation generation, considering their proven knowledge of human activities. This paper investigates potential approaches to combine XAI and LLMs for sensor-based ADL recognition. We evaluate if LLMs can be used: a) as explainable zero-shot ADL recognition models, avoiding costly labeled data collection, and b) to automate the generation of explanations for existing data-driven XAI approaches when training data is available and the goal is higher recognition rates. Our critical evaluation provides insights into the benefits and challenges of using LLMs for explainable ADL recognition.

Method: Two-level verification approach: meta-verification assesses completeness and consistency, while tool-based adaptive verification autonomously selects appropriate verification tools (mathematical, logical, commonsense) based on reasoning type.

Result: Outperforms baseline verification methods across all reasoning tasks, enhances reasoning accuracy through verification feedback, and achieves better results with fewer samples and costs in mathematical reasoning.

Conclusion: VerifiAgent provides an efficient, robust, and scalable verification solution that can be applied across diverse reasoning domains and also supports inference scaling with reduced computational costs.

Abstract: Large language models demonstrate remarkable reasoning capabilities but often produce unreliable or incorrect responses. Existing verification methods are typically model-specific or domain-restricted, requiring significant computational resources and lacking scalability across diverse reasoning tasks. To address these limitations, we propose VerifiAgent, a unified verification agent that integrates two levels of verification: meta-verification, which assesses completeness and consistency in model responses, and tool-based adaptive verification, where VerifiAgent autonomously selects appropriate verification tools based on the reasoning type, including mathematical, logical, or commonsense reasoning. This adaptive approach ensures both efficiency and robustness across different verification scenarios. Experimental results show that VerifiAgent outperforms baseline verification methods (e.g., deductive verifier, backward verifier) among all reasoning tasks. Additionally, it can further enhance reasoning accuracy by leveraging feedback from verification results. VerifiAgent can also be effectively applied to inference scaling, achieving better results with fewer generated samples and costs compared to existing process reward models in the mathematical reasoning domain. Code is available at https://github.com/Jiuzhouh/VerifiAgent

Method: Created MuSeD dataset with ≈11 hours of videos from TikTok and BitChute, developed an innovative annotation framework for analyzing textual, vocal, and visual modalities, and evaluated various large language models and multimodal LLMs for sexism detection.

Result: Visual information is crucial for detecting sexist content for both humans and models. Models perform well on explicit sexism but struggle with implicit cases like stereotypes, mirroring human annotators’ low agreement on such content.

Conclusion: Detecting implicit sexism remains challenging as it depends on social and cultural context. The study highlights the importance of multimodal approaches and the difficulties in automating detection of nuanced sexist content that requires contextual understanding.

Abstract: Sexism is generally defined as prejudice and discrimination based on sex or gender, affecting every sector of society, from social institutions to relationships and individual behavior. Social media platforms amplify the impact of sexism by conveying discriminatory content not only through text but also across multiple modalities, highlighting the critical need for a multimodal approach to the analysis of sexism online. With the rise of social media platforms where users share short videos, sexism is increasingly spreading through video content. Automatically detecting sexism in videos is a challenging task, as it requires analyzing the combination of verbal, audio, and visual elements to identify sexist content. In this study, (1) we introduce MuSeD, a new Multimodal Spanish dataset for Sexism Detection consisting of $\approx$ 11 hours of videos extracted from TikTok and BitChute; (2) we propose an innovative annotation framework for analyzing the contributions of textual, vocal, and visual modalities to the classification of content as either sexist or non-sexist; and (3) we evaluate a range of large language models (LLMs) and multimodal LLMs on the task of sexism detection. We find that visual information plays a key role in labeling sexist content for both humans and models. Models effectively detect explicit sexism; however, they struggle with implicit cases, such as stereotypes, instances where annotators also show low agreement. This highlights the inherent difficulty of the task, as identifying implicit sexism depends on the social and cultural context.

Method: Injecting Arabic language into a small open-source English model (1.5B parameters) using minimal original training data to create the Kuwain model.

Result: 8% average improvement in Arabic language benchmarks while maintaining existing English knowledge with minimal data requirements.

Conclusion: This approach provides a cost-effective alternative to full bilingual training, enabling targeted language expansion without extensive resources.

Abstract: Enhancing existing models with new knowledge is a crucial aspect of AI development. This paper introduces a novel method for integrating a new language into a large language model (LLM). Our approach successfully incorporates a previously unseen target language into an existing LLM without compromising its prior knowledge. We trained a tiny model with 1.5 billion parameters named Kuwain by injecting the Arabic language into a small open-source model mainly trained in English. Our method demonstrates significant improvements in Arabic language performance, with an average 8% improvement across various benchmarks, while retaining the model’s existing knowledge with a minimum amount of the original model’s data. This offers a cost-effective alternative to training a comprehensive model in both English and Arabic. The results highlight the potential for efficient, targeted language model expansion without extensive retraining or resource-intensive processes.

Method: Cequel constructs must-link and cannot-link constraints by selectively querying LLMs on informative text pairs or triplets identified via EdgeLLM and TriangleLLM algorithms, then uses these constraints in weighted constrained clustering.

Result: Experiments on multiple benchmark datasets show that Cequel consistently outperforms existing methods in unsupervised text clustering under the same query budget.

Conclusion: Cequel provides an effective solution for accurate text clustering while significantly reducing the computational and financial costs associated with LLM usage.

Abstract: Text clustering aims to automatically partition a collection of documents into coherent groups based on their linguistic features. In the literature, this task is formulated either as metric clustering over pre-trained text embeddings or as graph clustering based on pairwise similarities derived from an oracle, e.g., a large machine learning model. Recent advances in large language models (LLMs) have significantly improved this field by providing high-quality contextualized embeddings and accurate semantic similarity estimates. However, leveraging LLMs at scale introduces substantial computational and financial costs due to the large number of required API queries or inference calls. To address this issue, we propose Cequel, a cost-effective framework that achieves accurate text clustering under a limited budget of LLM queries. At its core, Cequel constructs must-link and cannot-link constraints by selectively querying LLMs on informative text pairs or triplets, identified via our proposed algorithms, EdgeLLM and TriangleLLM. These constraints are then utilized in a weighted constrained clustering algorithm to form high-quality clusters. Specifically, EdgeLLM and TriangleLLM employ carefully designed greedy selection strategies and prompting techniques to identify and extract informative constraints efficiently. Experiments on multiple benchmark datasets demonstrate that Cequel consistently outperforms existing methods in unsupervised text clustering under the same query budget.

Method: Combines Annif toolkit’s traditional NLP/ML techniques with LLM-based translation, synthetic data generation, and merging predictions from monolingual models.

Result: Ranked 1st in all-subjects category, 2nd in tib-core-subjects category (quantitative), and 4th in qualitative evaluations.

Conclusion: Demonstrates successful integration of traditional XMTC algorithms with LLM techniques for effective multilingual subject indexing.

Abstract: This paper presents the Annif system in SemEval-2025 Task 5 (LLMs4Subjects), which focussed on subject indexing using large language models (LLMs). The task required creating subject predictions for bibliographic records from the bilingual TIBKAT database using the GND subject vocabulary. Our approach combines traditional natural language processing and machine learning techniques implemented in the Annif toolkit with innovative LLM-based methods for translation and synthetic data generation, and merging predictions from monolingual models. The system ranked first in the all-subjects category and second in the tib-core-subjects category in the quantitative evaluation, and fourth in qualitative evaluations. These findings demonstrate the potential of combining traditional XMTC algorithms with modern LLM techniques to improve the accuracy and efficiency of subject indexing in multilingual contexts.

Method: Proposes a co-evolving World Model LLM that predicts next observations based on current state and action. The model serves dual roles: (1) as a virtual web server generating self-instructed training data, and (2) as an imagination engine for look-ahead simulation during inference.

Result: Experiments in real-world web environments (Mind2Web-Live, WebVoyager, GAIA-web) show 10% performance gain over existing self-evolving agents, demonstrating efficacy and generalizability without distillation from closed-source models.

Conclusion: The work establishes the necessity of integrating world models into autonomous agent frameworks to unlock sustained adaptability in web environments.

Abstract: Agent self-improvement, where the backbone Large Language Model (LLM) of the agent are trained on trajectories sampled autonomously based on their own policies, has emerged as a promising approach for enhancing performance. Recent advancements, particularly in web environments, face a critical limitation: their performance will reach a stagnation point during autonomous learning cycles, hindering further improvement. We argue that this stems from limited exploration of the web environment and insufficient exploitation of pre-trained web knowledge in LLMs. To improve the performance of self-improvement, we propose a novel framework that introduces a co-evolving World Model LLM. This world model predicts the next observation based on the current observation and action within the web environment. Leveraging LLMs’ pretrained knowledge of abundant web content, the World Model serves dual roles: (1) as a virtual web server generating self-instructed training data to continuously refine the agent’s policy, and (2) as an imagination engine during inference, enabling look-ahead simulation to guide action selection for the agent LLM. Experiments in real-world web environments (Mind2Web-Live, WebVoyager, and GAIA-web) show a 10% performance gain over existing self-evolving agents, demonstrating the efficacy and generalizability of our approach, without using any distillation from more powerful close-sourced models. Our work establishes the necessity of integrating world models into autonomous agent frameworks to unlock sustained adaptability. Code is available at https://github.com/Tencent/SelfEvolvingAgent

Method: Fine-tuned Kuwain 1.5B decoder-only language model on carefully curated diacritized datasets using rigorous data-cleaning and normalization pipeline.

Result: Sadeed achieves competitive results compared to proprietary large language models and outperforms traditional models trained on similar domains, despite using modest computational resources.

Conclusion: Sadeed and the new benchmark SadeedDiac-25 provide a robust foundation for advancing Arabic NLP applications including machine translation, text-to-speech, and language learning tools.

Abstract: Arabic text diacritization remains a persistent challenge in natural language processing due to the language’s morphological richness. In this paper, we introduce Sadeed, a novel approach based on a fine-tuned decoder-only language model adapted from Kuwain 1.5B Hennara et al. [2025], a compact model originally trained on diverse Arabic corpora. Sadeed is fine-tuned on carefully curated, high-quality diacritized datasets, constructed through a rigorous data-cleaning and normalization pipeline. Despite utilizing modest computational resources, Sadeed achieves competitive results compared to proprietary large language models and outperforms traditional models trained on similar domains. Additionally, we highlight key limitations in current benchmarking practices for Arabic diacritization. To address these issues, we introduce SadeedDiac-25, a new benchmark designed to enable fairer and more comprehensive evaluation across diverse text genres and complexity levels. Together, Sadeed and SadeedDiac-25 provide a robust foundation for advancing Arabic NLP applications, including machine translation, text-to-speech, and language learning tools.

Method: Developed Mutarjim based on Kuwain-1.5B using an optimized two-phase training approach and carefully curated high-quality training corpus. Also created Tarjama-25, a new benchmark with 5,000 expert-reviewed sentence pairs across diverse domains.

Result: Mutarjim outperforms models up to 20 times larger on established benchmarks and achieves state-of-the-art performance on English-to-Arabic translation in Tarjama-25, surpassing even GPT-4o mini.

Conclusion: Compact models like Mutarjim can achieve superior translation performance through optimized training approaches and high-quality data, challenging the need for massive model sizes while reducing computational burdens.

Abstract: We introduce Mutarjim, a compact yet powerful language model for bidirectional Arabic-English translation. While large-scale LLMs have shown impressive progress in natural language processing tasks, including machine translation, smaller models. Leveraging this insight, we developed Mutarjim based on Kuwain-1.5B , a language model tailored for both Arabic and English. Despite its modest size, Mutarjim outperforms much larger models on several established benchmarks, achieved through an optimized two-phase training approach and a carefully curated, high-quality training corpus.. Experimental results show that Mutarjim rivals models up to 20 times larger while significantly reducing computational costs and training requirements. We also introduce Tarjama-25, a new benchmark designed to overcome limitations in existing Arabic-English benchmarking datasets, such as domain narrowness, short sentence lengths, and English-source bias. Tarjama-25 comprises 5,000 expert-reviewed sentence pairs and spans a wide range of domains, offering a more comprehensive and balanced evaluation framework. Notably, Mutarjim achieves state-of-the-art performance on the English-to-Arabic task in Tarjama-25, surpassing even significantly larger and proprietary models like GPT-4o mini. We publicly release Tarjama-25 to support future research and advance the evaluation of Arabic-English translation systems.

Method: Trained 13,440 LLMs using entropy minimization technique with minimal unlabeled data (single example) and very short optimization (10 steps)

Result: Achieved performance improvements comparable to or better than rule-based reinforcement learning that uses thousands of data points and carefully designed rewards

Conclusion: This finding suggests a paradigm shift in post-training approaches, indicating that simpler entropy minimization with minimal data can be more effective than complex RL methods

Abstract: We trained 13,440 large language models and found that entropy minimization requires only a single unlabeled data and 10 steps optimization to achieve performance improvements comparable to or even greater than those obtained using thousands of data and carefully designed rewards in rule-based reinforcement learning. This striking result may prompt a rethinking of post-training paradigms for large language models. Our code is avaliable at https://github.com/zitian-gao/one-shot-em.

Method: Task-agnostic lossless compression technique similar to LZ77 algorithm that transforms token sequences in a reversible manner, ensuring no semantic information is lost during compression.

Result: Achieved 27% and 18% reduction in input token sequence length for two evaluation tasks, equating to 47% and 33% less encoding computation respectively due to quadratic attention. Performance gap compared to uncompressed input is minimal.

Conclusion: Lossless compression produces only a small performance gap compared to uncompressed input, and this gap would likely be erased entirely with larger models and expanded computing budget, making it a viable approach for strict semantic preservation tasks.

Abstract: Existing work on prompt compression for Large Language Models (LLM) focuses on lossy methods that try to maximize the retention of semantic information that is relevant to downstream tasks while significantly reducing the sequence length. In this paper, we introduce a task-agnostic lossless compression technique similar to LZ77 that makes it possible to reduce the input token sequence length on average by 27% and 18% for the two evaluation tasks explored here. Given that we use transformer-based LLMs, this equates to 47% and 33% less encoding computation, respectively, due to the quadratic nature of attention. The token sequence transformation is trivial to reverse and highlights that no semantic information is lost in the process. We evaluate our proposed approach on two tasks that require strict preservation of semantics/syntax and demonstrate that existing lossy compression methods perform poorly in this setting. We find that our lossless compression technique produces only a small gap in performance compared to using the uncompressed input and posit that larger models and an expanded computing budget would likely erase the gap entirely.

Method: Introduces LaMP-Cap dataset with multimodal figure profiles including figure images, captions, and figure-mentioning paragraphs from the same document. Tests four LLMs using profile information for personalized caption generation.

Result: Using profile information consistently helps generate captions closer to original author-written ones. Images in profiles are more helpful than text-only figure-mentioning paragraphs.

Conclusion: Multimodal profiles significantly improve personalized caption generation over text-only approaches, with visual information being particularly valuable for capturing author style and context.

Abstract: Figure captions are crucial for helping readers understand and remember a figure’s key message. Many models have been developed to generate these captions, helping authors compose better quality captions more easily. Yet, authors almost always need to revise generic AI-generated captions to match their writing style and the domain’s style, highlighting the need for personalization. Despite language models’ personalization (LaMP) advances, these technologies often focus on text-only settings and rarely address scenarios where both inputs and profiles are multimodal. This paper introduces LaMP-Cap, a dataset for personalized figure caption generation with multimodal figure profiles. For each target figure, LaMP-Cap provides not only the needed inputs, such as figure images, but also up to three other figures from the same document–each with its image, caption, and figure-mentioning paragraphs–as a profile to characterize the context. Experiments with four LLMs show that using profile information consistently helps generate captions closer to the original author-written ones. Ablation studies reveal that images in the profile are more helpful than figure-mentioning paragraphs, highlighting the advantage of using multimodal profiles over text-only ones.

Method: Benchmarked multiple reasoning-focused LLMs (especially DeepSeek models) across 15 Arabic NLP tasks using zero-shot, few-shot, and fine-tuning strategies with systematic evaluation on datasets of varying complexity.

Result: Three carefully selected in-context examples improved classification by 13+ F1 points; DeepSeek outperformed GPT-4-mini by 12 F1 points on complex inference; LoRA fine-tuning yielded up to 8 additional F1/BLEU points.

Conclusion: Strategic example selection, reasoning-focused architectures like DeepSeek, and efficient fine-tuning methods like LoRA significantly enhance LLM performance on complex Arabic NLP tasks, addressing the language’s unique challenges.

Abstract: Large language models (LLMs) have shown remarkable progress in reasoning abilities and general natural language processing (NLP) tasks, yet their performance on Arabic data, characterized by rich morphology, diverse dialects, and complex script, remains underexplored. This paper presents a comprehensive benchmarking study of multiple reasoning-focused LLMs, with a special emphasis on the newly introduced DeepSeek models, across a suite of fifteen Arabic NLP tasks. We experiment with various strategies, including zero-shot, few-shot, and fine-tuning. This allows us to systematically evaluate performance on datasets covering a range of applications to examine their capacity for linguistic reasoning under different levels of complexity. Our experiments reveal several key findings. First, carefully selecting just three in-context examples delivers an average uplift of over 13 F1 points on classification tasks-boosting sentiment analysis from 35.3% to 87.5% and paraphrase detection from 56.1% to 87.0%. Second, reasoning-focused DeepSeek architectures outperform a strong GPT o4-mini baseline by an average of 12 F1 points on complex inference tasks in the zero-shot setting. Third, LoRA-based fine-tuning yields up to an additional 8 points in F1 and BLEU compared to equivalent increases in model scale. The code is available at https://anonymous.4open.science/r/AraReasoner41299

Method: Created MFTCXplain dataset with 3,000 tweets across 4 languages (Portuguese, Italian, Persian, English) annotated with hate speech labels, moral categories, and text span-level rationales using Moral Foundation Theory.

Result: LLMs show good hate speech detection (F1 up to 0.836) but weak moral sentiment prediction (F1 < 0.35). Rationale alignment is particularly limited in underrepresented languages.

Conclusion: Current LLMs have limited capacity to internalize and reflect human moral reasoning, especially across diverse cultural and linguistic contexts.

Abstract: Ensuring the moral reasoning capabilities of Large Language Models (LLMs) is a growing concern as these systems are used in socially sensitive tasks. Nevertheless, current evaluation benchmarks present two major shortcomings: a lack of annotations that justify moral classifications, which limits transparency and interpretability; and a predominant focus on English, which constrains the assessment of moral reasoning across diverse cultural settings. In this paper, we introduce MFTCXplain, a multilingual benchmark dataset for evaluating the moral reasoning of LLMs via hate speech multi-hop explanation using Moral Foundation Theory (MFT). The dataset comprises 3,000 tweets across Portuguese, Italian, Persian, and English, annotated with binary hate speech labels, moral categories, and text span-level rationales. Empirical results highlight a misalignment between LLM outputs and human annotations in moral reasoning tasks. While LLMs perform well in hate speech detection (F1 up to 0.836), their ability to predict moral sentiments is notably weak (F1 < 0.35). Furthermore, rationale alignment remains limited mainly in underrepresented languages. These findings show the limited capacity of current LLMs to internalize and reflect human moral reasoning.

Method: Tested LLaMA 3 8B model for alignment faking behavior and implemented prompt-only interventions including deontological moral framing and scratchpad reasoning techniques without modifying model internals.

Result: Found that small instruction-tuned models can exhibit alignment faking, and that prompt-based interventions significantly reduce this deceptive behavior, suggesting shallow deception can be contextually suppressed.

Conclusion: The study refines understanding of deception in language models by distinguishing shallow vs deep deception and emphasizes the need for alignment evaluations across all model sizes and deployment scenarios.

Abstract: Current literature suggests that alignment faking (deceptive alignment) is an emergent property of large language models. We present the first empirical evidence that a small instruction-tuned model, specifically LLaMA 3 8B, can exhibit alignment faking. We further show that prompt-only interventions, including deontological moral framing and scratchpad reasoning, significantly reduce this behavior without modifying model internals. This challenges the assumption that prompt-based ethics are trivial and that deceptive alignment requires scale. We introduce a taxonomy distinguishing shallow deception, shaped by context and suppressible through prompting, from deep deception, which reflects persistent, goal-driven misalignment. Our findings refine the understanding of deception in language models and underscore the need for alignment evaluations across model sizes and deployment settings.

Method: Proposes Self-taught ActioN Deliberation (SAND) framework with self-consistency action sampling and execution-guided action critique to synthesize step-wise deliberation thoughts, then uses these deliberation trajectories to iteratively finetune the LLM agent itself.

Result: Achieves 20% average improvement over initial supervised finetuning and outperforms state-of-the-art agent tuning approaches on two representative interactive agent tasks.

Conclusion: Explicit action deliberation through self-consistency sampling and execution-guided critique significantly improves LLM agent performance by enabling better action space exploration and avoiding commitment to suboptimal actions.

Abstract: Large Language Model (LLM) agents are commonly tuned with supervised finetuning on ReAct-style expert trajectories or preference optimization over pairwise rollouts. Most of these methods focus on imitating specific expert behaviors or promoting chosen reasoning thoughts and actions over rejected ones. However, without reasoning and comparing over alternatives actions, LLM agents finetuned with these methods may over-commit towards seemingly plausible but suboptimal actions due to limited action space exploration. To address this, in this paper we propose Self-taught ActioN Deliberation (SAND) framework, enabling LLM agents to explicitly deliberate over candidate actions before committing to one. To tackle the challenges of when and what to deliberate given large action space and step-level action evaluation, we incorporate self-consistency action sampling and execution-guided action critique to help synthesize step-wise action deliberation thoughts using the base model of the LLM agent. In an iterative manner, the deliberation trajectories are then used to finetune the LLM agent itself. Evaluating on two representative interactive agent tasks, SAND achieves an average 20% improvement over initial supervised finetuning and also outperforms state-of-the-art agent tuning approaches.

Method: Conducted large-scale empirical study of hidden representations in 11 autoregressive models across 6 scientific topics, analyzing separability patterns across layers and prompt types.

Result: Found consistent low-dimensional linearly separable semantic representations, with enhanced separability in deeper layers and structured reasoning prompts. Trained MLP probe as latent-space guardrail that significantly improves refusal rates on malicious queries.

Conclusion: LLM latent spaces contain geometrically organized semantic information that enables effective safety interventions through geometry-aware tools operating directly in the latent space.

Abstract: Understanding the latent space geometry of large language models (LLMs) is key to interpreting their behavior and improving alignment. However, it remains unclear to what extent LLMs internally organize representations related to semantic understanding. To explore this, we conduct a large-scale empirical study of hidden representations in 11 autoregressive models across 6 scientific topics. We find that high-level semantic information consistently resides in low-dimensional subspaces that form linearly separable representations across domains. This separability becomes more pronounced in deeper layers and under prompts that elicit structured reasoning or alignment behavior$\unicode{x2013}$even when surface content remains unchanged. These findings support geometry-aware tools that operate directly in latent space to detect and mitigate harmful or adversarial content. As a proof of concept, we train an MLP probe on final-layer hidden states to act as a lightweight latent-space guardrail. This approach substantially improves refusal rates on malicious queries and prompt injections that bypass both the model’s built-in safety alignment and external token-level filters.

Method: Pre-trained base model on diverse high-quality monolingual/bilingual data across 28 languages. Instruct model finetuned with Chain-of-Thought reasoning and enhanced through reinforcement learning for better generalization.

Result: Achieves performance comparable to leading closed-source models (Gemini-2.5, GPT-4o) and significantly outperforms larger open-source models in both automatic metrics and human evaluations across 28 languages.

Conclusion: Seed-X demonstrates that 7B parameter open-source models can achieve state-of-the-art translation performance, providing valuable best practices and making parameters publicly available to advance translation research.

Abstract: Multilingual translation stands as a challenging task for large language models (LLMs) to handle intricate language patterns and stilted translations that arise in automated translations. In this paper, we introduce Seed-X, a family of open-source LLMs comprising instruct and reasoning models, pushing the limits of translation capability with 7B parameter size. The base model is pre-trained on a diverse, high-quality dataset encompassing both monolingual and bilingual content across 28 languages, harnessing the full potential of multilingual data. The instruct model is then finetuned to translate by Chain-of-Thought (CoT) reasoning and further enhanced through reinforcement learning (RL) to achieve better generalization across diverse language pairs. Seed-X achieves performance comparable to leading closed-source models, including Gemini-2.5 and GPT-4o, across 28 languages, and significantly outperforms larger open-source models in both automatic metrics and human evaluations. We share the best practices through our optimization process, and make the parameter public available for advancing translation research and applications.

Method: Empirical analysis of internal representations, focusing on multi-head attention mechanisms and scaling specific hidden units to examine length control.

Result: Multi-head attention is critical for length determination; specific hidden units can be scaled to control length without compromising text informativeness; some units become more active with length-specific prompts.

Conclusion: LLMs have learned robust internal mechanisms for output length control that are partially disentangled from semantic information, demonstrating internal awareness of length attributes.

Abstract: Large language models (LLMs) have shown remarkable capabilities across various tasks, that are learned from massive amounts of text-based data. Although LLMs can control output sequence length, particularly in instruction-based settings, the internal mechanisms behind this control have been unexplored yet. In this study, we provide empirical evidence on how output sequence length information is encoded within the internal representations in LLMs. In particular, our findings show that multi-head attention mechanisms are critical in determining output sequence length, which can be adjusted in a disentangled manner. By scaling specific hidden units within the model, we can control the output sequence length without losing the informativeness of the generated text, thereby indicating that length information is partially disentangled from semantic information. Moreover, some hidden units become increasingly active as prompts become more length-specific, thus reflecting the model’s internal awareness of this attribute. Our findings suggest that LLMs have learned robust and adaptable internal mechanisms for controlling output length without any external control.

Method: Developed a manually curated dataset from Wikipedia by native speakers, including both text-only and visual questions with English translations. Evaluated state-of-the-art LLMs through prompting and human judgment of answer correctness.

Result: Best open-weight LLMs achieved only ~50% accuracy on textual questions and <30% on visual questions. LLM-based evaluation metrics showed strong correlation with human judgment, while traditional string-overlap metrics performed well due to named entity prevalence.

Conclusion: Current LLMs struggle with regional knowledge and multimodal understanding, highlighting the need for improved models and evaluation methods for culturally-specific and visual question answering tasks.

Abstract: We introduce CUS-QA, a benchmark for open-ended regional question answering that encompasses both textual and visual modalities. We also provide strong baselines using state-of-the-art large language models (LLMs). Our dataset consists of manually curated questions and answers grounded in Wikipedia, created by native speakers from Czechia, Slovakia, and Ukraine, with accompanying English translations. It includes both purely textual questions and those requiring visual understanding. We evaluate state-of-the-art LLMs through prompting and complement this with human judgments of answer correctness. Using these human evaluations, we analyze the reliability of existing automatic evaluation metrics. Our baseline results show that even the best open-weight LLMs achieve only around 50% accuracy on textual questions and below 30% on visual questions. LLM-based evaluation metrics show strong correlation with human judgment, while traditional string-overlap metrics perform surprisingly well due to the prevalence of named entities in answers.

Method: STIM (Source-aware Token-level Identification of Memorization) attributes each token in reasoning chains to memorization sources (local, mid-range, or long-range) based on statistical co-occurrence patterns in the pretraining corpus.

Result: Analysis shows models rely more on memorization in complex/long-tail cases, with local memorization driving up to 67% of wrong tokens. STIM memorization scores effectively predict wrong tokens in reasoning steps.

Conclusion: STIM provides a powerful diagnostic tool for identifying and addressing memorization issues in model reasoning, with potential applications to other structured step-wise generation tasks.

Abstract: Large Language Models (LLMs) perform well on reasoning benchmarks but often fail when inputs alter slightly, raising concerns about the extent to which their success relies on memorization. This issue is especially acute in Chain-of-Thought (CoT) reasoning, where spurious memorized patterns can trigger intermediate errors that cascade into incorrect final answers. We introduce STIM, a novel framework for Source-aware Token-level Identification of Memorization, which attributes each token in a reasoning chain to one of multiple memorization sources - local, mid-range, or long-range - based on their statistical co-occurrence with the token in the pretraining corpus. Our token-level analysis across tasks and distributional settings reveals that models rely more on memorization in complex or long-tail cases, and that local memorization is often the dominant driver of errors, leading to up to 67% of wrong tokens. We also show that memorization scores from STIM can be effective in predicting the wrong tokens in the wrong reasoning step. STIM offers a powerful tool for diagnosing and improving model reasoning and can generalize to other structured step-wise generation tasks.

Method: Curated dataset of 150,430 High Court bail judgments with structured annotations, fine-tuned LLM with parameter-efficient techniques, evaluated with/without statutory context and RAG.

Result: Models with statutory knowledge significantly outperform baselines, achieving strong accuracy and explanation quality, generalize well to expert-annotated test set.

Conclusion: IBPS provides transparent, scalable solution for data-driven legal assistance to reduce bail delays and promote procedural fairness in Indian judicial system.

Abstract: Bail decisions are among the most frequently adjudicated matters in Indian courts, yet they remain plagued by subjectivity, delays, and inconsistencies. With over 75% of India’s prison population comprising undertrial prisoners, many from socioeconomically disadvantaged backgrounds, the lack of timely and fair bail adjudication exacerbates human rights concerns and contributes to systemic judicial backlog. In this paper, we present the Indian Bail Prediction System (IBPS), an AI-powered framework designed to assist in bail decision-making by predicting outcomes and generating legally sound rationales based solely on factual case attributes and statutory provisions. We curate and release a large-scale dataset of 150,430 High Court bail judgments, enriched with structured annotations such as age, health, criminal history, crime category, custody duration, statutes, and judicial reasoning. We fine-tune a large language model using parameter-efficient techniques and evaluate its performance across multiple configurations, with and without statutory context, and with RAG. Our results demonstrate that models fine-tuned with statutory knowledge significantly outperform baselines, achieving strong accuracy and explanation quality, and generalize well to a test set independently annotated by legal experts. IBPS offers a transparent, scalable, and reproducible solution to support data-driven legal assistance, reduce bail delays, and promote procedural fairness in the Indian judicial system.

Method: Proposes LATTE framework that uses contrastive learning to align raw event embeddings with semantic embeddings from frozen LLMs. Behavioral features are summarized into short prompts, embedded by LLM, and used as supervision via contrastive loss.

Result: Significantly reduces inference cost and input size compared to conventional LLM processing of complete sequences. Outperforms state-of-the-art techniques for learning event sequence representations on real-world financial datasets.

Conclusion: The approach remains deployable in latency-sensitive environments while achieving superior performance compared to existing methods for financial sequence representation learning.

Abstract: Learning clients embeddings from sequences of their historic communications is central to financial applications. While large language models (LLMs) offer general world knowledge, their direct use on long event sequences is computationally expensive and impractical in real-world pipelines. In this paper, we propose LATTE, a contrastive learning framework that aligns raw event embeddings with semantic embeddings from frozen LLMs. Behavioral features are summarized into short prompts, embedded by the LLM, and used as supervision via contrastive loss. The proposed approach significantly reduces inference cost and input size compared to conventional processing of complete sequence by LLM. We experimentally show that our method outperforms state-of-the-art techniques for learning event sequence representations on real-world financial datasets while remaining deployable in latency-sensitive environments.

Method: The paper explores Arabic MML by categorizing existing research through a novel taxonomy organized into four key topics: datasets, applications, approaches, and challenges.

Result: Provides a structured overview of the current state of Arabic MML, highlighting unexplored areas and critical research gaps to guide future research.

Conclusion: This survey empowers researchers to build upon identified opportunities and address challenges to advance the field of Arabic multimodal machine learning.

Abstract: Multimodal Machine Learning (MML) aims to integrate and analyze information from diverse modalities, such as text, audio, and visuals, enabling machines to address complex tasks like sentiment analysis, emotion recognition, and multimedia retrieval. Recently, Arabic MML has reached a certain level of maturity in its foundational development, making it time to conduct a comprehensive survey. This paper explores Arabic MML by categorizing efforts through a novel taxonomy and analyzing existing research. Our taxonomy organizes these efforts into four key topics: datasets, applications, approaches, and challenges. By providing a structured overview, this survey offers insights into the current state of Arabic MML, highlighting areas that have not been investigated and critical research gaps. Researchers will be empowered to build upon the identified opportunities and address challenges to advance the field.

Method: Pre-trained a 12B parameter model on 20T tokens using FP8 training, then used Minitron strategy to compress and distill it to 9B parameters for inference on 128k tokens with A10G GPU.

Result: Achieves on-par or better accuracy than similar models (e.g., Qwen3-8B) with up to 6x higher inference throughput in reasoning scenarios (8k input, 16k output tokens).

Conclusion: The hybrid Mamba-Transformer architecture successfully balances accuracy and throughput, making it suitable for efficient reasoning workloads, with models and datasets released publicly.

Abstract: We introduce Nemotron-Nano-9B-v2, a hybrid Mamba-Transformer language model designed to increase throughput for reasoning workloads while achieving state-of-the-art accuracy compared to similarly-sized models. Nemotron-Nano-9B-v2 builds on the Nemotron-H architecture, in which the majority of the self-attention layers in the common Transformer architecture are replaced with Mamba-2 layers, to achieve improved inference speed when generating the long thinking traces needed for reasoning. We create Nemotron-Nano-9B-v2 by first pre-training a 12-billion-parameter model (Nemotron-Nano-12B-v2-Base) on 20 trillion tokens using an FP8 training recipe. After aligning Nemotron-Nano-12B-v2-Base, we employ the Minitron strategy to compress and distill the model with the goal of enabling inference on up to 128k tokens on a single NVIDIA A10G GPU (22GiB of memory, bfloat16 precision). Compared to existing similarly-sized models (e.g., Qwen3-8B), we show that Nemotron-Nano-9B-v2 achieves on-par or better accuracy on reasoning benchmarks while achieving up to 6x higher inference throughput in reasoning settings like 8k input and 16k output tokens. We are releasing Nemotron-Nano-9B-v2, Nemotron-Nano12B-v2-Base, and Nemotron-Nano-9B-v2-Base checkpoints along with the majority of our pre- and post-training datasets on Hugging Face.

Method: The authors propose a new training objective based on the classic structured prediction framework as an alternative to Soft-argmax for differentiable end-to-end training in facial landmark detection.

Result: The method achieves state-of-the-art performance on three facial landmark benchmarks (WFLW, COFW, and 300W), converges 2.2x faster during training, and maintains competitive or better accuracy compared to existing approaches.

Conclusion: Soft-argmax is not the only effective approach for facial landmark detection; the proposed structured prediction-based method offers faster convergence and strong performance, providing a viable alternative to established techniques.

Abstract: Facial landmark detection is an important task in computer vision with numerous applications, such as head pose estimation, expression analysis, face swapping, etc. Heatmap regression-based methods have been widely used to achieve state-of-the-art results in this task. These methods involve computing the argmax over the heatmaps to predict a landmark. Since argmax is not differentiable, these methods use a differentiable approximation, Soft-argmax, to enable end-to-end training on deep-nets. In this work, we revisit this long-standing choice of using Soft-argmax and demonstrate that it is not the only way to achieve strong performance. Instead, we propose an alternative training objective based on the classic structured prediction framework. Empirically, our method achieves state-of-the-art performance on three facial landmark benchmarks (WFLW, COFW, and 300W), converging 2.2x faster during training while maintaining better/competitive accuracy. Our code is available here: https://github.com/ca-joe-yang/regression-without-softarg.

Method: Represents images as graphs where pixels/regions are vertices, refines graphs using Delaunay triangulations, and integrates GCNs with Voronoi diagrams for optimized representation.

Result: Achieves significant improvements in preprocessing efficiency and classification accuracy across benchmark datasets, surpassing state-of-the-art approaches in challenging scenarios.

Conclusion: The combination of GCNs with Voronoi diagrams provides an effective approach for advancing image classification and expands graph-based learning applications in computer vision and unstructured data analysis.

Abstract: The rapid progress in image classification has been largely driven by the adoption of Graph Convolutional Networks (GCNs), which offer a robust framework for handling complex data structures. This study introduces a novel approach that integrates GCNs with Voronoi diagrams to enhance image classification by leveraging their ability to effectively model relational data. Unlike conventional convolutional neural networks (CNNs), our method represents images as graphs, where pixels or regions function as vertices. These graphs are then refined using corresponding Delaunay triangulations, optimizing their representation. The proposed model achieves significant improvements in both preprocessing efficiency and classification accuracy across various benchmark datasets, surpassing state-of-the-art approaches, particularly in challenging scenarios involving intricate scenes and fine-grained categories. Experimental results, validated through cross-validation, underscore the effectiveness of combining GCNs with Voronoi diagrams for advancing image classification. This research not only presents a novel perspective on image classification but also expands the potential applications of graph-based learning paradigms in computer vision and unstructured data analysis.

Method: Single-stage query-based framework using transformer detector with lightweight pose head, bounding-box-conditioned translation module, and 6D-aware Hungarian matching cost. Trained end-to-end with only RGB images and category-level pose labels.

Result: Sets new state-of-the-art on three benchmarks. On REAL275: 79.6% IoU50 and 54.1% under 10°10cm metric, surpassing prior RGB-only methods and closing gap to RGB-D systems.

Conclusion: YOPO demonstrates that object detection and 9-DoF pose estimation can be unified with high performance using only RGB images, providing a simpler yet effective alternative to complex multi-stage approaches.

Abstract: Accurately recovering the full 9-DoF pose of unseen instances within specific categories from a single RGB image remains a core challenge for robotics and automation. Most existing solutions still rely on pseudo-depth, CAD models, or multi-stage cascades that separate 2D detection from pose estimation. Motivated by the need for a simpler, RGB-only alternative that learns directly at the category level, we revisit a longstanding question: Can object detection and 9-DoF pose estimation be unified with high performance, without any additional data? We show that they can with our method, YOPO, a single-stage, query-based framework that treats category-level 9-DoF estimation as a natural extension of 2D detection. YOPO augments a transformer detector with a lightweight pose head, a bounding-box-conditioned translation module, and a 6D-aware Hungarian matching cost. The model is trained end-to-end only with RGB images and category-level pose labels. Despite its minimalist design, YOPO sets a new state of the art on three benchmarks. On the REAL275 dataset, it achieves 79.6% $\rm{IoU}_{50}$ and 54.1% under the $10^\circ$$10{\rm{cm}}$ metric, surpassing prior RGB-only methods and closing much of the gap to RGB-D systems. The code, models, and additional qualitative results can be found on our project.

Method: Proposes Paired-Sampling Contrastive Framework that uses automatically matched pairs of genuine and attack selfies to learn modality-agnostic liveness cues through contrastive learning.

Result: Achieves 2.10% average classification error rate (ACER) on the 6th Face Anti-Spoofing Challenge benchmark, outperforming prior solutions. The framework is lightweight (4.46 GFLOPs) and trains in under one hour.

Conclusion: The proposed unified framework effectively detects both physical and digital face spoofing attacks with superior performance and practical efficiency, making it suitable for real-world deployment.

Abstract: Modern face recognition systems remain vulnerable to spoofing attempts, including both physical presentation attacks and digital forgeries. Traditionally, these two attack vectors have been handled by separate models, each targeting its own artifacts and modalities. However, maintaining distinct detectors increases system complexity and inference latency and leaves systems exposed to combined attack vectors. We propose the Paired-Sampling Contrastive Framework, a unified training approach that leverages automatically matched pairs of genuine and attack selfies to learn modality-agnostic liveness cues. Evaluated on the 6th Face Anti-Spoofing Challenge Unified Physical-Digital Attack Detection benchmark, our method achieves an average classification error rate (ACER) of 2.10 percent, outperforming prior solutions. The framework is lightweight (4.46 GFLOPs) and trains in under one hour, making it practical for real-world deployment. Code and pretrained models are available at https://github.com/xPONYx/iccv2025_deepfake_challenge.

Method: TAIGen injects perturbations during the mixing step interval without processing all timesteps. It uses a selective RGB channel strategy: attention maps on red channel and GradCAM-guided perturbations on green/blue channels to preserve structure while maximizing misclassification.

Result: TAIGen achieves 70.6% success against ResNet, 80.8% against MNASNet, and 97.8% against ShuffleNet on ImageNet with VGGNet source. Maintains PSNR above 30 dB and generates adversarial examples 10x faster than existing diffusion-based attacks.

Conclusion: TAIGen is the most impactful attack method as it achieves the lowest robust accuracy, indicating defense mechanisms are least successful in purifying images generated by this approach.

Abstract: Adversarial attacks from generative models often produce low-quality images and require substantial computational resources. Diffusion models, though capable of high-quality generation, typically need hundreds of sampling steps for adversarial generation. This paper introduces TAIGen, a training-free black-box method for efficient adversarial image generation. TAIGen produces adversarial examples using only 3-20 sampling steps from unconditional diffusion models. Our key finding is that perturbations injected during the mixing step interval achieve comparable attack effectiveness without processing all timesteps. We develop a selective RGB channel strategy that applies attention maps to the red channel while using GradCAM-guided perturbations on green and blue channels. This design preserves image structure while maximizing misclassification in target models. TAIGen maintains visual quality with PSNR above 30 dB across all tested datasets. On ImageNet with VGGNet as source, TAIGen achieves 70.6% success against ResNet, 80.8% against MNASNet, and 97.8% against ShuffleNet. The method generates adversarial examples 10x faster than existing diffusion-based attacks. Our method achieves the lowest robust accuracy, indicating it is the most impactful attack as the defense mechanism is least successful in purifying the images generated by TAIGen.

Method: Formulates CVP as optimization problem unfolded into multi-stage network with three modules: CORE (mask domain reversible modeling), CARE (RGB domain reversible modeling), and FINE (targeted Bernoulli diffusion for uncertain regions).

Result: Provides principled reversible modeling across both domains, efficient uncertainty resolution through targeted diffusion, and reduced false positives/negatives by focusing on ambiguous areas.

Conclusion: RUN++ introduces a novel bi-level optimization framework that synergizes unfolding networks with diffusion models for robust CVP systems effective under real-world degradations.

Abstract: Existing methods for concealed visual perception (CVP) often leverage reversible strategies to decrease uncertainty, yet these are typically confined to the mask domain, leaving the potential of the RGB domain underexplored. To address this, we propose a reversible unfolding network with generative refinement, termed RUN++. Specifically, RUN++ first formulates the CVP task as a mathematical optimization problem and unfolds the iterative solution into a multi-stage deep network. This approach provides a principled way to apply reversible modeling across both mask and RGB domains while leveraging a diffusion model to resolve the resulting uncertainty. Each stage of the network integrates three purpose-driven modules: a Concealed Object Region Extraction (CORE) module applies reversible modeling to the mask domain to identify core object regions; a Context-Aware Region Enhancement (CARE) module extends this principle to the RGB domain to foster better foreground-background separation; and a Finetuning Iteration via Noise-based Enhancement (FINE) module provides a final refinement. The FINE module introduces a targeted Bernoulli diffusion model that refines only the uncertain regions of the segmentation mask, harnessing the generative power of diffusion for fine-detail restoration without the prohibitive computational cost of a full-image process. This unique synergy, where the unfolding network provides a strong uncertainty prior for the diffusion model, allows RUN++ to efficiently direct its focus toward ambiguous areas, significantly mitigating false positives and negatives. Furthermore, we introduce a new paradigm for building robust CVP systems that remain effective under real-world degradations and extend this concept into a broader bi-level optimization framework.

Method: Uses a novel Mix Twin encoder alternating between spatially-reduced global attention and locally-grouped attention mechanisms, with a lightweight LR-ASPP decoder for multi-scale feature aggregation in a unified framework.

Result: Achieves 74.47% mIoU and 83.63% mF1 for segmentation, 100% dietary classification accuracy, with only 3.348M parameters, 3.428G FLOPs, and 114.9 FPS inference speed. Also introduces first comprehensive beef cattle methane emission dataset with 11,694 annotated frames.

Conclusion: GasTwinFormer establishes a practical solution for real-time livestock emission monitoring, validated by extensive ablation studies demonstrating the effectiveness of leveraging diet-emission correlations.

Abstract: Livestock methane emissions represent 32% of human-caused methane production, making automated monitoring critical for climate mitigation strategies. We introduce GasTwinFormer, a hybrid vision transformer for real-time methane emission segmentation and dietary classification in optical gas imaging through a novel Mix Twin encoder alternating between spatially-reduced global attention and locally-grouped attention mechanisms. Our architecture incorporates a lightweight LR-ASPP decoder for multi-scale feature aggregation and enables simultaneous methane segmentation and dietary classification in a unified framework. We contribute the first comprehensive beef cattle methane emission dataset using OGI, containing 11,694 annotated frames across three dietary treatments. GasTwinFormer achieves 74.47% mIoU and 83.63% mF1 for segmentation while maintaining exceptional efficiency with only 3.348M parameters, 3.428G FLOPs, and 114.9 FPS inference speed. Additionally, our method achieves perfect dietary classification accuracy (100%), demonstrating the effectiveness of leveraging diet-emission correlations. Extensive ablation studies validate each architectural component, establishing GasTwinFormer as a practical solution for real-time livestock emission monitoring. Please see our project page at gastwinformer.github.io.

Method: VAREdit formulates image editing as sequential next-scale prediction over discrete visual tokens. It uses a Scale-Aligned Reference (SAR) module to inject scale-matched conditioning information from source image features into the first self-attention layer.

Result: Outperforms leading diffusion-based methods by 30%+ higher GPT-Balance score, completes 512×512 editing in 1.2 seconds (2.2× faster than similarly sized UltraEdit).

Conclusion: Visual autoregressive framework with scale-aligned conditioning provides superior editing adherence and efficiency compared to diffusion-based approaches, demonstrating the advantages of sequential compositional generation for image editing tasks.

Abstract: Recent advances in diffusion models have brought remarkable visual fidelity to instruction-guided image editing. However, their global denoising process inherently entangles the edited region with the entire image context, leading to unintended spurious modifications and compromised adherence to editing instructions. In contrast, autoregressive models offer a distinct paradigm by formulating image synthesis as a sequential process over discrete visual tokens. Their causal and compositional mechanism naturally circumvents the adherence challenges of diffusion-based methods. In this paper, we present VAREdit, a visual autoregressive (VAR) framework that reframes image editing as a next-scale prediction problem. Conditioned on source image features and text instructions, VAREdit generates multi-scale target features to achieve precise edits. A core challenge in this paradigm is how to effectively condition the source image tokens. We observe that finest-scale source features cannot effectively guide the prediction of coarser target features. To bridge this gap, we introduce a Scale-Aligned Reference (SAR) module, which injects scale-matched conditioning information into the first self-attention layer. VAREdit demonstrates significant advancements in both editing adherence and efficiency. On standard benchmarks, it outperforms leading diffusion-based methods by 30%+ higher GPT-Balance score. Moreover, it completes a $512\times512$ editing in 1.2 seconds, making it 2.2$\times$ faster than the similarly sized UltraEdit. The models are available at https://github.com/HiDream-ai/VAREdit.

Method: A novel flow matching framework that learns smooth, non-linear trajectories by incorporating curvature guidance into the flow path with robust curvature regularization that penalizes abrupt changes in trajectory dynamics.

Result: State-of-the-art performance on MS COCO 2014 and 2017 datasets, significantly outperforming standard rectified flow variants and non-linear baselines like Rectified Diffusion, with especially notable improvements in semantic consistency metrics (BLEU, METEOR, ROUGE, CLAIR).

Conclusion: Curvature-aware modeling substantially enhances the model’s ability to faithfully follow complex instructions while maintaining high image quality, demonstrating the importance of non-linear trajectories for semantic alignment in text-to-image generation.

Abstract: Existing rectified flow models are based on linear trajectories between data and noise distributions. This linearity enforces zero curvature, which can inadvertently force the image generation process through low-probability regions of the data manifold. A key question remains underexplored: how does the curvature of these trajectories correlate with the semantic alignment between generated images and their corresponding captions, i.e., instructional compliance? To address this, we introduce CurveFlow, a novel flow matching framework designed to learn smooth, non-linear trajectories by directly incorporating curvature guidance into the flow path. Our method features a robust curvature regularization technique that penalizes abrupt changes in the trajectory’s intrinsic dynamics.Extensive experiments on MS COCO 2014 and 2017 demonstrate that CurveFlow achieves state-of-the-art performance in text-to-image generation, significantly outperforming both standard rectified flow variants and other non-linear baselines like Rectified Diffusion. The improvements are especially evident in semantic consistency metrics such as BLEU, METEOR, ROUGE, and CLAIR. This confirms that our curvature-aware modeling substantially enhances the model’s ability to faithfully follow complex instructions while simultaneously maintaining high image quality. The code is made publicly available at https://github.com/Harvard-AI-and-Robotics-Lab/CurveFlow.

Method: Combines ranking and contrastive learning with novel higher-order ranking loss, embedding distance loss, and training-time contrastive alignment loss guided by structured textual prompts. Uses only synthetic distortions for training.

Result: Achieves state-of-the-art performance, strong generalization to authentic degradations (lens flare, haze, motion blur, low-light), and offers a lightweight variant (HiRQA-S) with 3.5ms inference time per image.

Conclusion: HiRQA provides an effective self-supervised framework for NR-IQA that generalizes well from synthetic to real-world distortions while maintaining real-time performance capabilities.

Abstract: Despite significant progress in no-reference image quality assessment (NR-IQA), dataset biases and reliance on subjective labels continue to hinder their generalization performance. We propose HiRQA, Hierarchical Ranking and Quality Alignment), a self-supervised, opinion-unaware framework that offers a hierarchical, quality-aware embedding through a combination of ranking and contrastive learning. Unlike prior approaches that depend on pristine references or auxiliary modalities at inference time, HiRQA predicts quality scores using only the input image. We introduce a novel higher-order ranking loss that supervises quality predictions through relational ordering across distortion pairs, along with an embedding distance loss that enforces consistency between feature distances and perceptual differences. A training-time contrastive alignment loss, guided by structured textual prompts, further enhances the learned representation. Trained only on synthetic distortions, HiRQA generalizes effectively to authentic degradations, as demonstrated through evaluation on various distortions such as lens flare, haze, motion blur, and low-light conditions. For real-time deployment, we introduce \textbf{HiRQA-S}, a lightweight variant with an inference time of only 3.5 ms per image. Extensive experiments across synthetic and authentic benchmarks validate HiRQA’s state-of-the-art (SOTA) performance, strong generalization ability, and scalability.

Method: Proposes a portfolio theory-inspired optimization approach using genetic algorithm to manage edge resources and select cameras that guarantee reconstruction quality despite spatiotemporal disruptions.

Result: Demonstrated benefits over traditional baseline strategies using public and customized 3D datasets, showing reliable 3D reconstruction under spatiotemporal disruptions.

Conclusion: The portfolio-based camera selection strategy effectively ensures reliable 3D reconstruction quality in dynamic edge environments with correlated disruptions.

Abstract: Multi-view 3D reconstruction applications are revolutionizing critical use cases that require rapid situational-awareness, such as emergency response, tactical scenarios, and public safety. In many cases, their near-real-time latency requirements and ad-hoc needs for compute resources necessitate adoption of `Just-in-time’ edge environments where the system is set up on the fly to support the applications during the mission lifetime. However, reliability issues can arise from the inherent dynamism and operational adversities of such edge environments, resulting in spatiotemporally correlated disruptions that impact the camera operations, which can lead to sustained degradation of reconstruction quality. In this paper, we propose a novel portfolio theory inspired edge resource management strategy for reliable multi-view 3D reconstruction against possible system disruptions. Our proposed methodology can guarantee reconstruction quality satisfaction even when the cameras are prone to spatiotemporally correlated disruptions. The portfolio theoretic optimization problem is solved using a genetic algorithm that converges quickly for realistic system settings. Using publicly available and customized 3D datasets, we demonstrate the proposed camera selection strategy’s benefits in guaranteeing reliable 3D reconstruction against traditional baseline strategies, under spatiotemporal disruptions.

Method: Proposes XDR-LVLM framework with specialized Medical Vision Encoder, LVLM Core, Multi-task Prompt Engineering, and Multi-stage Fine-tuning to understand pathological features in fundus images and generate comprehensive diagnostic reports with explanations.

Result: Achieved state-of-the-art performance: 84.55% Balanced Accuracy and 79.92% F1 Score for disease diagnosis, 77.95% BACC and 66.88% F1 for concept detection. Human evaluations confirmed high fluency, accuracy, and clinical utility of explanations.

Conclusion: XDR-LVLM successfully bridges the gap between automated diagnosis and clinical needs by providing robust, interpretable insights through natural language explanations, making it suitable for clinical adoption.

Abstract: Diabetic Retinopathy (DR) is a major cause of global blindness, necessitating early and accurate diagnosis. While deep learning models have shown promise in DR detection, their black-box nature often hinders clinical adoption due to a lack of transparency and interpretability. To address this, we propose XDR-LVLM (eXplainable Diabetic Retinopathy Diagnosis with LVLM), a novel framework that leverages Vision-Language Large Models (LVLMs) for high-precision DR diagnosis coupled with natural language-based explanations. XDR-LVLM integrates a specialized Medical Vision Encoder, an LVLM Core, and employs Multi-task Prompt Engineering and Multi-stage Fine-tuning to deeply understand pathological features within fundus images and generate comprehensive diagnostic reports. These reports explicitly include DR severity grading, identification of key pathological concepts (e.g., hemorrhages, exudates, microaneurysms), and detailed explanations linking observed features to the diagnosis. Extensive experiments on the Diabetic Retinopathy (DDR) dataset demonstrate that XDR-LVLM achieves state-of-the-art performance, with a Balanced Accuracy of 84.55% and an F1 Score of 79.92% for disease diagnosis, and superior results for concept detection (77.95% BACC, 66.88% F1). Furthermore, human evaluations confirm the high fluency, accuracy, and clinical utility of the generated explanations, showcasing XDR-LVLM’s ability to bridge the gap between automated diagnosis and clinical needs by providing robust and interpretable insights.

Method: Three-stage pipeline: 1) Generate geometrically consistent sparse views using Cascaded Outpainting ControlNets, 2) Propagate denser intermediate views via AGInpaint component, 3) Eliminate visual inconsistencies globally using GCAlign module. Concurrently reconstruct 3D Gaussian Splatting scene initialized on mesh surface.

Result: Outperforms existing approaches in both geometric alignment and generation quality. Enables diverse style rendering through relighting and style transfer techniques.

Conclusion: MeSS successfully addresses texture generation for city mesh models, providing high-quality, consistent outdoor scenes suitable for virtual navigation and autonomous driving applications.

Abstract: Mesh models have become increasingly accessible for numerous cities; however, the lack of realistic textures restricts their application in virtual urban navigation and autonomous driving. To address this, this paper proposes MeSS (Meshbased Scene Synthesis) for generating high-quality, styleconsistent outdoor scenes with city mesh models serving as the geometric prior. While image and video diffusion models can leverage spatial layouts (such as depth maps or HD maps) as control conditions to generate street-level perspective views, they are not directly applicable to 3D scene generation. Video diffusion models excel at synthesizing consistent view sequences that depict scenes but often struggle to adhere to predefined camera paths or align accurately with rendered control videos. In contrast, image diffusion models, though unable to guarantee cross-view visual consistency, can produce more geometry-aligned results when combined with ControlNet. Building on this insight, our approach enhances image diffusion models by improving cross-view consistency. The pipeline comprises three key stages: first, we generate geometrically consistent sparse views using Cascaded Outpainting ControlNets; second, we propagate denser intermediate views via a component dubbed AGInpaint; and third, we globally eliminate visual inconsistencies (e.g., varying exposure) using the GCAlign module. Concurrently with generation, a 3D Gaussian Splatting (3DGS) scene is reconstructed by initializing Gaussian balls on the mesh surface. Our method outperforms existing approaches in both geometric alignment and generation quality. Once synthesized, the scene can be rendered in diverse styles through relighting and style transfer techniques.

Method: Created SurgWound dataset with 697 surgical wound images annotated by 3 professional surgeons. Developed a three-stage learning framework (WoundQwen) using multiple MLLMs: first stage predicts wound characteristics, second stage diagnoses outcomes, third stage generates comprehensive reports.

Result: First open-source surgical wound dataset and benchmark established. The proposed WoundQwen framework can analyze wound characteristics, assess infection risk, and generate comprehensive diagnostic reports from surgical images.

Conclusion: This work enables personalized wound care through automated surgical wound screening, providing timely intervention guidance and improving patient outcomes by addressing critical gaps in surgical wound diagnosis infrastructure.

Abstract: Surgical site infection (SSI) is one of the most common and costly healthcare-associated infections and and surgical wound care remains a significant clinical challenge in preventing SSIs and improving patient outcomes. While recent studies have explored the use of deep learning for preliminary surgical wound screening, progress has been hindered by concerns over data privacy and the high costs associated with expert annotation. Currently, no publicly available dataset or benchmark encompasses various types of surgical wounds, resulting in the absence of an open-source Surgical-Wound screening tool. To address this gap: (1) we present SurgWound, the first open-source dataset featuring a diverse array of surgical wound types. It contains 697 surgical wound images annotated by 3 professional surgeons with eight fine-grained clinical attributes. (2) Based on SurgWound, we introduce the first benchmark for surgical wound diagnosis, which includes visual question answering (VQA) and report generation tasks to comprehensively evaluate model performance. (3) Furthermore, we propose a three-stage learning framework, WoundQwen, for surgical wound diagnosis. In the first stage, we employ five independent MLLMs to accurately predict specific surgical wound characteristics. In the second stage, these predictions serve as additional knowledge inputs to two MLLMs responsible for diagnosing outcomes, which assess infection risk and guide subsequent interventions. In the third stage, we train a MLLM that integrates the diagnostic results from the previous two stages to produce a comprehensive report. This three-stage framework can analyze detailed surgical wound characteristics and provide subsequent instructions to patients based on surgical images, paving the way for personalized wound care, timely intervention, and improved patient outcomes.

Method: A learning-based method is presented to derive adversarial behavior for rule-based agents that can cause failure scenarios. The approach evaluates the adversarial agent against all rule-based agents.

Result: The evaluation shows a decrease in cumulative reward when the adversarial agent is deployed against rule-based agents, demonstrating the effectiveness of the adversarial approach in creating failure scenarios.

Conclusion: The proposed learning-based adversarial method successfully identifies vulnerabilities in rule-based agent systems, highlighting the importance of robust modeling for safety-critical applications like autonomous driving where failure scenarios must be anticipated and mitigated.

Abstract: Existing approaches in reinforcement learning train an agent to learn desired optimal behavior in an environment with rule based surrounding agents. In safety critical applications such as autonomous driving it is crucial that the rule based agents are modelled properly. Several behavior modelling strategies and IDM models are used currently to model the surrounding agents. We present a learning based method to derive the adversarial behavior for the rule based agents to cause failure scenarios. We evaluate our adversarial agent against all the rule based agents and show the decrease in cumulative reward.

Method: Integrated watershed, skeletonization, and morphological operations within a unified framework with adaptive geometric refinement and class-specific hyperparameter tuning. Systematic parameter optimization separates adherent and heterogeneous structures.

Result: Outperforms individual classical approaches and naive combinations, enabling superior instance separation and more accurate morphometric analysis in renal pathology workflows.

Conclusion: DyMorph-B2I provides a robust solution for converting binary masks to instance-level segmentations in renal pathology, facilitating precise morphological quantification of functional units with publicly available implementation.

Abstract: Accurate morphological quantification of renal pathology functional units relies on instance-level segmentation, yet most existing datasets and automated methods provide only binary (semantic) masks, limiting the precision of downstream analyses. Although classical post-processing techniques such as watershed, morphological operations, and skeletonization, are often used to separate semantic masks into instances, their individual effectiveness is constrained by the diverse morphologies and complex connectivity found in renal tissue. In this study, we present DyMorph-B2I, a dynamic, morphology-guided binary-to-instance segmentation pipeline tailored for renal pathology. Our approach integrates watershed, skeletonization, and morphological operations within a unified framework, complemented by adaptive geometric refinement and customizable hyperparameter tuning for each class of functional unit. Through systematic parameter optimization, DyMorph-B2I robustly separates adherent and heterogeneous structures present in binary masks. Experimental results demonstrate that our method outperforms individual classical approaches and na"ive combinations, enabling superior instance separation and facilitating more accurate morphometric analysis in renal pathology workflows. The pipeline is publicly available at: https://github.com/ddrrnn123/DyMorph-B2I.

Method: Incorporates improved spatio-temporal features and aggregates them through a recurrent network to enhance graph neural networks for accident prediction from dash-cam videos.

Result: Achieves higher average precision and mean time-to-collision values than previous methods across three publicly available datasets, with strong performance in both cross-validation and cross-dataset testing scenarios.

Conclusion: The proposed STAGNet model provides an effective and practical solution for accident prediction using only dash-cam video, offering better performance than state-of-the-art methods while being more cost-effective and deployable.

Abstract: Accident prediction and timely warnings play a key role in improving road safety by reducing the risk of injury to road users and minimizing property damage. Advanced Driver Assistance Systems (ADAS) are designed to support human drivers and are especially useful when they can anticipate potential accidents before they happen. While many existing systems depend on a range of sensors such as LiDAR, radar, and GPS, relying solely on dash-cam video input presents a more challenging but a more cost-effective and easily deployable solution. In this work, we incorporate better spatio-temporal features and aggregate them through a recurrent network to improve upon state-of-the-art graph neural networks for predicting accidents from dash-cam videos. Experiments using three publicly available datasets show that our proposed STAGNet model achieves higher average precision and mean time-to-collision values than previous methods, both when cross-validated on a given dataset and when trained and tested on different datasets.

Method: 1) Collaborative multi-modal coding integrating modality-specific features while preserving unique strengths; 2) Auxiliary 2D and 3D supervision for robustness; 3) Triplane latent diffusion model for high-quality 3D asset generation.

Result: Achieves competitive performance with models trained on large-scale datasets despite using small training data. Successfully incorporates RGB-D datasets, demonstrating feasibility of multi-modal 3D generation.

Conclusion: TriMM effectively harnesses multi-modal data for 3D modeling, producing superior quality assets with enhanced textures and geometric details while demonstrating scalability to incorporate various multi-modal datasets.

Abstract: 3D content inherently encompasses multi-modal characteristics and can be projected into different modalities (e.g., RGB images, RGBD, and point clouds). Each modality exhibits distinct advantages in 3D asset modeling: RGB images contain vivid 3D textures, whereas point clouds define fine-grained 3D geometries. However, most existing 3D-native generative architectures either operate predominantly within single-modality paradigms-thus overlooking the complementary benefits of multi-modality data-or restrict themselves to 3D structures, thereby limiting the scope of available training datasets. To holistically harness multi-modalities for 3D modeling, we present TriMM, the first feed-forward 3D-native generative model that learns from basic multi-modalities (e.g., RGB, RGBD, and point cloud). Specifically, 1) TriMM first introduces collaborative multi-modal coding, which integrates modality-specific features while preserving their unique representational strengths. 2) Furthermore, auxiliary 2D and 3D supervision are introduced to raise the robustness and performance of multi-modal coding. 3) Based on the embedded multi-modal code, TriMM employs a triplane latent diffusion model to generate 3D assets of superior quality, enhancing both the texture and the geometric detail. Extensive experiments on multiple well-known datasets demonstrate that TriMM, by effectively leveraging multi-modality, achieves competitive performance with models trained on large-scale datasets, despite utilizing a small amount of training data. Furthermore, we conduct additional experiments on recent RGB-D datasets, verifying the feasibility of incorporating other multi-modal datasets into 3D generation.

Method: Proposes a framework with: 1) Soft prototype contrastive module using sample-to-center probability as weights to calculate category prototypes, 2) Dual consistency learning module that aligns different transformations of same samples and neighborhoods of different samples.

Result: Extensive experiments on five datasets show the method is effective compared to state-of-the-art approaches.

Conclusion: The proposed framework successfully reduces prototype drift, avoids inter-class conflicts, and provides reliable prototype calculation while ensuring transformation-invariant semantic information and compact intra-cluster distribution.

Abstract: Contrastive learning is widely used in clustering tasks due to its discriminative representation. However, the conflict problem between classes is difficult to solve effectively. Existing methods try to solve this problem through prototype contrast, but there is a deviation between the calculation of hard prototypes and the true cluster center. To address this problem, we propose a center-oriented prototype contrastive clustering framework, which consists of a soft prototype contrastive module and a dual consistency learning module. In short, the soft prototype contrastive module uses the probability that the sample belongs to the cluster center as a weight to calculate the prototype of each category, while avoiding inter-class conflicts and reducing prototype drift. The dual consistency learning module aligns different transformations of the same sample and the neighborhoods of different samples respectively, ensuring that the features have transformation-invariant semantic information and compact intra-cluster distribution, while providing reliable guarantees for the calculation of prototypes. Extensive experiments on five datasets show that the proposed method is effective compared to the SOTA. Our code is published on https://github.com/LouisDong95/CPCC.

Method: Uses Conflict Maps (CMs) from ray-to-model-prior analysis, semantically segments real-world CMs with synthetically generated CMs from Semantic Conflict Map Generator (SCMG), and fuses additional segmentation of textured models with CMs using confidence scores.

Result: Achieves 61% performance with uncertainty-aware fusion of segmented building textures, effectively segmenting and reconstructing building openings.

Conclusion: The research advances automated LoD3 model reconstruction, enabling scalable and efficient 3D city modeling for urban planning, digital twins, and disaster management applications.

Abstract: Detailed 3D building models are crucial for urban planning, digital twins, and disaster management applications. While Level of Detail 1 (LoD)1 and LoD2 building models are widely available, they lack detailed facade elements essential for advanced urban analysis. In contrast, LoD3 models address this limitation by incorporating facade elements such as windows, doors, and underpasses. However, their generation has traditionally required manual modeling, making large-scale adoption challenging. In this contribution, CM2LoD3, we present a novel method for reconstructing LoD3 building models leveraging Conflict Maps (CMs) obtained from ray-to-model-prior analysis. Unlike previous works, we concentrate on semantically segmenting real-world CMs with synthetically generated CMs from our developed Semantic Conflict Map Generator (SCMG). We also observe that additional segmentation of textured models can be fused with CMs using confidence scores to further increase segmentation performance and thus increase 3D reconstruction accuracy. Experimental results demonstrate the effectiveness of our CM2LoD3 method in segmenting and reconstructing building openings, with the 61% performance with uncertainty-aware fusion of segmented building textures. This research contributes to the advancement of automated LoD3 model reconstruction, paving the way for scalable and efficient 3D city modeling. Our project is available: https://github.com/InFraHank/CM2LoD3

Method: Propose DuAl-VLN task with two UAVs: high-altitude UAV uses multimodal LLM (Pilot-LLM) for target reasoning, low-altitude UAV uses lightweight multi-stage policy for navigation. Only exchange coordinate information for efficiency.

Result: Created HaL-13k dataset with 13,838 collaborative trajectories and target-oriented instructions. Includes unseen maps and object validation sets for systematic generalization evaluation.

Conclusion: Dual-UAV collaboration enables effective aerial navigation by combining high-level environmental reasoning with precise low-altitude execution, providing a scalable solution for complex UAV-VLN tasks.

Abstract: Aerial Vision-and-Language Navigation (VLN) is an emerging task that enables Unmanned Aerial Vehicles (UAVs) to navigate outdoor environments using natural language instructions and visual cues. However, due to the extended trajectories and complex maneuverability of UAVs, achieving reliable UAV-VLN performance is challenging and often requires human intervention or overly detailed instructions. To harness the advantages of UAVs’ high mobility, which could provide multi-grained perspectives, while maintaining a manageable motion space for learning, we introduce a novel task called Dual-Altitude UAV Collaborative VLN (DuAl-VLN). In this task, two UAVs operate at distinct altitudes: a high-altitude UAV responsible for broad environmental reasoning, and a low-altitude UAV tasked with precise navigation. To support the training and evaluation of the DuAl-VLN, we construct the HaL-13k, a dataset comprising 13,838 collaborative high-low UAV demonstration trajectories, each paired with target-oriented language instructions. This dataset includes both unseen maps and an unseen object validation set to systematically evaluate the model’s generalization capabilities across novel environments and unfamiliar targets. To consolidate their complementary strengths, we propose a dual-UAV collaborative VLN framework, AeroDuo, where the high-altitude UAV integrates a multimodal large language model (Pilot-LLM) for target reasoning, while the low-altitude UAV employs a lightweight multi-stage policy for navigation and target grounding. The two UAVs work collaboratively and only exchange minimal coordinate information to ensure efficiency.

Method: Proposed skipped-step sampling mechanism that bypasses multiple intermediate denoising steps in the iterative generation process. This approach is derived from the same training objective as standard diffusion models and can be integrated with DDIM for enhanced performance.

Result: Extensive experiments on OpenAI ADM, Stable Diffusion, and Open Sora models show the method achieves high-quality generation with significantly reduced sampling steps compared to traditional approaches.

Conclusion: Accelerated sampling via skipped-step sampling is an intrinsic property of pretrained diffusion models, enabling efficient Markovian generation without compromising quality, and can be effectively combined with existing methods like DDIM.

Abstract: Diffusion models have been achieving state-of-the-art results across various generation tasks. However, a notable drawback is their sequential generation process, requiring long-sequence step-by-step generation. Existing methods, such as DDIM, attempt to reduce sampling steps by constructing a class of non-Markovian diffusion processes that maintain the same training objective. However, there remains a gap in understanding whether the original diffusion process can achieve the same efficiency without resorting to non-Markovian processes. In this paper, we provide a confirmative answer and introduce skipped-step sampling, a mechanism that bypasses multiple intermediate denoising steps in the iterative generation process, in contrast with the traditional step-by-step refinement of standard diffusion inference. Crucially, we demonstrate that this skipped-step sampling mechanism is derived from the same training objective as the standard diffusion model, indicating that accelerated sampling via skipped-step sampling via a Markovian way is an intrinsic property of pretrained diffusion models. Additionally, we propose an enhanced generation method by integrating our accelerated sampling technique with DDIM. Extensive experiments on popular pretrained diffusion models, including the OpenAI ADM, Stable Diffusion, and Open Sora models, show that our method achieves high-quality generation with significantly reduced sampling steps.

Method: Three key innovations: 1) Multi-functional coding framework unifying various coding modes, 2) Interactive coding agent using augmented in-context learning with expert feedback, 3) IIC-bench benchmark for evaluation.

Result: Comp-X efficiently understands coding requests with impressive textual interaction capability while maintaining comparable compression performance with a single framework.

Conclusion: The system provides a promising avenue for AGI in image compression by enabling intelligent, user-friendly compression through LLM reasoning capabilities.

Abstract: We present Comp-X, the first intelligently interactive image compression paradigm empowered by the impressive reasoning capability of large language model (LLM) agent. Notably, commonly used image codecs usually suffer from limited coding modes and rely on manual mode selection by engineers, making them unfriendly for unprofessional users. To overcome this, we advance the evolution of image coding paradigm by introducing three key innovations: (i) multi-functional coding framework, which unifies different coding modes of various objective/requirements, including human-machine perception, variable coding, and spatial bit allocation, into one framework. (ii) interactive coding agent, where we propose an augmented in-context learning method with coding expert feedback to teach the LLM agent how to understand the coding request, mode selection, and the use of the coding tools. (iii) IIC-bench, the first dedicated benchmark comprising diverse user requests and the corresponding annotations from coding experts, which is systematically designed for intelligently interactive image compression evaluation. Extensive experimental results demonstrate that our proposed Comp-X can understand the coding requests efficiently and achieve impressive textual interaction capability. Meanwhile, it can maintain comparable compression performance even with a single coding framework, providing a promising avenue for artificial general intelligence (AGI) in image compression.

Method: Built on a pre-trained vision-language model with a lightweight trainable MLP, Ano-NAViLa incorporates both normal and abnormal pathology knowledge to enhance accuracy and robustness while providing interpretability through image-text associations.

Result: Evaluated on two lymph node datasets from different organs, Ano-NAViLa achieves state-of-the-art performance in both anomaly detection and localization, outperforming competing models.

Conclusion: Ano-NAViLa successfully addresses the challenges of pathology anomaly detection by leveraging vision-language modeling with pathology-specific knowledge, demonstrating superior performance and interpretability compared to existing methods.

Abstract: Anomaly detection in computational pathology aims to identify rare and scarce anomalies where disease-related data are often limited or missing. Existing anomaly detection methods, primarily designed for industrial settings, face limitations in pathology due to computational constraints, diverse tissue structures, and lack of interpretability. To address these challenges, we propose Ano-NAViLa, a Normal and Abnormal pathology knowledge-augmented Vision-Language model for Anomaly detection in pathology images. Ano-NAViLa is built on a pre-trained vision-language model with a lightweight trainable MLP. By incorporating both normal and abnormal pathology knowledge, Ano-NAViLa enhances accuracy and robustness to variability in pathology images and provides interpretability through image-text associations. Evaluated on two lymph node datasets from different organs, Ano-NAViLa achieves the state-of-the-art performance in anomaly detection and localization, outperforming competing models.

Method: Restructured Transformer decoder by swapping cross-attention and self-attention layers to retain redundant predictions before suppression, and instantiated multiple parallel cross-attention blocks with independent parameters to enhance lane diversity.

Result: Extensive experiments on OpenLane-V2 show RATopo is model-agnostic, seamlessly integrates into existing frameworks, and consistently improves both lane-lane and lane-traffic topology performance.

Conclusion: The proposed redundancy assignment strategy enables quantity-rich and geometry-diverse topology supervision, effectively addressing the limitations of traditional one-to-one assignment methods in lane topology reasoning.

Abstract: Lane topology reasoning plays a critical role in autonomous driving by modeling the connections among lanes and the topological relationships between lanes and traffic elements. Most existing methods adopt a first-detect-then-reason paradigm, where topological relationships are supervised based on the one-to-one assignment results obtained during the detection stage. This supervision strategy results in suboptimal topology reasoning performance due to the limited range of valid supervision. In this paper, we propose RATopo, a Redundancy Assignment strategy for lane Topology reasoning that enables quantity-rich and geometry-diverse topology supervision. Specifically, we restructure the Transformer decoder by swapping the cross-attention and self-attention layers. This allows redundant lane predictions to be retained before suppression, enabling effective one-to-many assignment. We also instantiate multiple parallel cross-attention blocks with independent parameters, which further enhances the diversity of detected lanes. Extensive experiments on OpenLane-V2 demonstrate that our RATopo strategy is model-agnostic and can be seamlessly integrated into existing topology reasoning frameworks, consistently improving both lane-lane and lane-traffic topology performance.

Method: Developed DesignCLIP framework using CLIP models with class-aware classification and contrastive learning. Utilized generated detailed captions for patent images and multi-view image learning. Built on a large-scale dataset of U.S. design patents.

Result: DesignCLIP consistently outperforms baseline and state-of-the-art models across various downstream tasks including patent classification and patent retrieval. Also enables multimodal patent retrieval for enhanced creativity and innovation.

Conclusion: Multimodal approaches show significant promise in advancing patent analysis. DesignCLIP demonstrates the effectiveness of combining vision-language models with patent-specific adaptations for improved performance in design patent applications.

Abstract: In the field of design patent analysis, traditional tasks such as patent classification and patent image retrieval heavily depend on the image data. However, patent images – typically consisting of sketches with abstract and structural elements of an invention – often fall short in conveying comprehensive visual context and semantic information. This inadequacy can lead to ambiguities in evaluation during prior art searches. Recent advancements in vision-language models, such as CLIP, offer promising opportunities for more reliable and accurate AI-driven patent analysis. In this work, we leverage CLIP models to develop a unified framework DesignCLIP for design patent applications with a large-scale dataset of U.S. design patents. To address the unique characteristics of patent data, DesignCLIP incorporates class-aware classification and contrastive learning, utilizing generated detailed captions for patent images and multi-views image learning. We validate the effectiveness of DesignCLIP across various downstream tasks, including patent classification and patent retrieval. Additionally, we explore multimodal patent retrieval, which provides the potential to enhance creativity and innovation in design by offering more diverse sources of inspiration. Our experiments show that DesignCLIP consistently outperforms baseline and SOTA models in the patent domain on all tasks. Our findings underscore the promise of multimodal approaches in advancing patent analysis. The codebase is available here: https://anonymous.4open.science/r/PATENTCLIP-4661/README.md.

Method: Temporal Prompt Alignment (TPA) extracts frame features using image encoder, aggregates with temporal extractor, aligns with class-specific text prompts via contrastive loss, and uses CVAESM module for uncertainty quantification and style modulation.

Result: TPA achieves 85.40% macro F1 for CHD diagnosis, reduces calibration error by 5.38-6.8%, and boosts EchoNet-Dynamic three-class task performance by 4.73% (to 58.62% macro F1).

Conclusion: TPA effectively addresses limitations of current methods by integrating temporal modeling, prompt learning, and uncertainty quantification, demonstrating superior performance and clinical reliability for ultrasound video analysis.

Abstract: Congenital heart defect (CHD) detection in ultrasound videos is hindered by image noise and probe positioning variability. While automated methods can reduce operator dependence, current machine learning approaches often neglect temporal information, limit themselves to binary classification, and do not account for prediction calibration. We propose Temporal Prompt Alignment (TPA), a method leveraging foundation image-text model and prompt-aware contrastive learning to classify fetal CHD on cardiac ultrasound videos. TPA extracts features from each frame of video subclips using an image encoder, aggregates them with a trainable temporal extractor to capture heart motion, and aligns the video representation with class-specific text prompts via a margin-hinge contrastive loss. To enhance calibration for clinical reliability, we introduce a Conditional Variational Autoencoder Style Modulation (CVAESM) module, which learns a latent style vector to modulate embeddings and quantifies classification uncertainty. Evaluated on a private dataset for CHD detection and on a large public dataset, EchoNet-Dynamic, for systolic dysfunction, TPA achieves state-of-the-art macro F1 scores of 85.40% for CHD diagnosis, while also reducing expected calibration error by 5.38% and adaptive ECE by 6.8%. On EchoNet-Dynamic’s three-class task, it boosts macro F1 by 4.73% (from 53.89% to 58.62%). Temporal Prompt Alignment (TPA) is a framework for fetal congenital heart defect (CHD) classification in ultrasound videos that integrates temporal modeling, prompt-aware contrastive learning, and uncertainty quantification.

Method: Created BasketLiDAR dataset with 4,445 frames and 3,105 player IDs from professional games, featuring synchronized LiDAR point clouds and multi-view camera footage. Developed novel MOT algorithm with two pipelines: LiDAR-only real-time tracking and multimodal fusion of LiDAR+camera data.

Result: Achieved real-time operation (previously difficult with camera-only methods) and superior tracking performance even under occlusion conditions. The method provides complete 3D positional information with high precision.

Conclusion: LiDAR-based approach overcomes limitations of traditional camera systems, enabling robust real-time 3D player tracking in challenging basketball scenarios with frequent occlusions and complex movements.

Abstract: Real-time 3D trajectory player tracking in sports plays a crucial role in tactical analysis, performance evaluation, and enhancing spectator experience. Traditional systems rely on multi-camera setups, but are constrained by the inherently two-dimensional nature of video data and the need for complex 3D reconstruction processing, making real-time analysis challenging. Basketball, in particular, represents one of the most difficult scenarios in the MOT field, as ten players move rapidly and complexly within a confined court space, with frequent occlusions caused by intense physical contact. To address these challenges, this paper constructs BasketLiDAR, the first multimodal dataset in the sports MOT field that combines LiDAR point clouds with synchronized multi-view camera footage in a professional basketball environment, and proposes a novel MOT framework that simultaneously achieves improved tracking accuracy and reduced computational cost. The BasketLiDAR dataset contains a total of 4,445 frames and 3,105 player IDs, with fully synchronized IDs between three LiDAR sensors and three multi-view cameras. We recorded 5-on-5 and 3-on-3 game data from actual professional basketball players, providing complete 3D positional information and ID annotations for each player. Based on this dataset, we developed a novel MOT algorithm that leverages LiDAR’s high-precision 3D spatial information. The proposed method consists of a real-time tracking pipeline using LiDAR alone and a multimodal tracking pipeline that fuses LiDAR and camera data. Experimental results demonstrate that our approach achieves real-time operation, which was difficult with conventional camera-only methods, while achieving superior tracking performance even under occlusion conditions. The dataset is available upon request at: https://sites.google.com/keio.jp/keio-csg/projects/basket-lidar

Method: Two-stage approach: 1) RAG stage uses unsupervised clustering to build retrieval database and generate coarse masks as prompts, 2) SEG stage uses SAM2 for precise mask refinement. Entirely training-free.

Result: Competitive performance on benchmark COD datasets, performing on par with or surpassing state-of-the-art methods. All experiments conducted on a personal laptop, demonstrating computational efficiency.

Conclusion: RAG-SEG provides an efficient, training-free solution for COD that maintains high performance while being computationally practical, eliminating the need for conventional training procedures.

Abstract: Camouflaged object detection (COD) poses a significant challenge in computer vision due to the high similarity between objects and their backgrounds. Existing approaches often rely on heavy training and large computational resources. While foundation models such as the Segment Anything Model (SAM) offer strong generalization, they still struggle to handle COD tasks without fine-tuning and require high-quality prompts to yield good performance. However, generating such prompts manually is costly and inefficient. To address these challenges, we propose \textbf{First RAG, Second SEG (RAG-SEG)}, a training-free paradigm that decouples COD into two stages: Retrieval-Augmented Generation (RAG) for generating coarse masks as prompts, followed by SAM-based segmentation (SEG) for refinement. RAG-SEG constructs a compact retrieval database via unsupervised clustering, enabling fast and effective feature retrieval. During inference, the retrieved features produce pseudo-labels that guide precise mask generation using SAM2. Our method eliminates the need for conventional training while maintaining competitive performance. Extensive experiments on benchmark COD datasets demonstrate that RAG-SEG performs on par with or surpasses state-of-the-art methods. Notably, all experiments are conducted on a \textbf{personal laptop}, highlighting the computational efficiency and practicality of our approach. We present further analysis in the Appendix, covering limitations, salient object detection extension, and possible improvements.

Method: Two-stage plug-and-play process: Selective Prompt Embedding Adjustment (SPEA) and Adversarial-Resilient Noise Guidance (ARNG) that integrates with existing T2V diffusion models without retraining.

Result: Achieves 46% average reduction in undesirable content across four tasks (object, style, celebrity, and explicit content erasure), outperforming prior methods in efficacy, integrity, fidelity, robustness, and generalizability.

Conclusion: VideoEraser provides an effective training-free solution for content safety in T2V generation, achieving state-of-the-art performance in suppressing undesirable concepts while maintaining video quality.

Abstract: The rapid growth of text-to-video (T2V) diffusion models has raised concerns about privacy, copyright, and safety due to their potential misuse in generating harmful or misleading content. These models are often trained on numerous datasets, including unauthorized personal identities, artistic creations, and harmful materials, which can lead to uncontrolled production and distribution of such content. To address this, we propose VideoEraser, a training-free framework that prevents T2V diffusion models from generating videos with undesirable concepts, even when explicitly prompted with those concepts. Designed as a plug-and-play module, VideoEraser can seamlessly integrate with representative T2V diffusion models via a two-stage process: Selective Prompt Embedding Adjustment (SPEA) and Adversarial-Resilient Noise Guidance (ARNG). We conduct extensive evaluations across four tasks, including object erasure, artistic style erasure, celebrity erasure, and explicit content erasure. Experimental results show that VideoEraser consistently outperforms prior methods regarding efficacy, integrity, fidelity, robustness, and generalizability. Notably, VideoEraser achieves state-of-the-art performance in suppressing undesirable content during T2V generation, reducing it by 46% on average across four tasks compared to baselines.

Method: Used deep learning for pose detection combined with sequence modeling techniques (GRU, LSTM, 1D CNN) on video data to create an end-to-end framework for temporal prediction of crossing intent.

Result: GRU performed better than LSTM for intent prediction accuracy, while 1D CNN was the fastest model in terms of processing speed.

Conclusion: The study successfully developed a deep learning framework for pedestrian intent prediction, demonstrating that different sequence models have trade-offs between accuracy and speed for autonomous vehicle applications.

Abstract: The world is constantly moving towards AI based systems and autonomous vehicles are now reality in different parts of the world. These vehicles require sensors and cameras to detect objects and maneuver according to that. It becomes important to for such vehicles to also predict from a distant if a person is about to cross a road or not. The current study focused on predicting the intent of crossing the road by pedestrians in an experimental setup. The study involved working with deep learning models to predict poses and sequence modelling for temporal predictions. The study analysed three different sequence modelling to understand the prediction behaviour and it was found out that GRU was better in predicting the intent compared to LSTM model but 1D CNN was the best model in terms of speed. The study involved video analysis, and the output of pose detection model was integrated later on to sequence modelling techniques for an end-to-end deep learning framework for predicting road crossing intents.

Method: Proposed dual-stream HDR video generation paradigm that decouples temporal luminance anchoring from exposure-variant detail reconstruction. Designed asynchronous dual-camera system (DCS) for independent exposure control without synchronization. Developed exposure-adaptive fusion network (EAFNet) with pre-alignment, asymmetric cross-feature fusion, and reconstruction subnetworks.

Result: Achieves state-of-the-art performance across various datasets, demonstrating remarkable potential in HDR video reconstruction with eliminated flicker artifacts.

Conclusion: The proposed dual-camera system with EAFNet effectively addresses temporal flicker in HDR video while maintaining cost-effectiveness, representing a significant advancement over traditional alternating exposure methods.

Abstract: High Dynamic Range (HDR) video acquisition using the alternating exposure (AE) paradigm has garnered significant attention due to its cost-effectiveness with a single consumer camera. However, despite progress driven by deep neural networks, these methods remain prone to temporal flicker in real-world applications due to inter-frame exposure inconsistencies. To address this challenge while maintaining the cost-effectiveness of the AE paradigm, we propose a novel learning-based HDR video generation solution. Specifically, we propose a dual-stream HDR video generation paradigm that decouples temporal luminance anchoring from exposure-variant detail reconstruction, overcoming the inherent limitations of the AE paradigm. To support this, we design an asynchronous dual-camera system (DCS), which enables independent exposure control across two cameras, eliminating the need for synchronization typically required in traditional multi-camera setups. Furthermore, an exposure-adaptive fusion network (EAFNet) is formulated for the DCS system. EAFNet integrates a pre-alignment subnetwork that aligns features across varying exposures, ensuring robust feature extraction for subsequent fusion, an asymmetric cross-feature fusion subnetwork that emphasizes reference-based attention to effectively merge these features across exposures, and a reconstruction subnetwork to mitigate ghosting artifacts and preserve fine details. Extensive experimental evaluations demonstrate that the proposed method achieves state-of-the-art performance across various datasets, showing the remarkable potential of our solution in HDR video reconstruction. The codes and data captured by DCS will be available at https://zqqqyu.github.io/DCS-HDR/.

Method: Proposes RCDINO, a transformer-based model that fuses camera and radar data with DINOv2 features to enrich visual representations while maintaining baseline architecture compatibility.

Result: Achieves state-of-the-art performance on nuScenes dataset with 56.4 NDS and 48.1 mAP among radar-camera models.

Conclusion: The fusion of DINOv2 features with multimodal data significantly improves 3D object detection performance while preserving architectural compatibility.

Abstract: Three-dimensional object detection is essential for autonomous driving and robotics, relying on effective fusion of multimodal data from cameras and radar. This work proposes RCDINO, a multimodal transformer-based model that enhances visual backbone features by fusing them with semantically rich representations from the pretrained DINOv2 foundation model. This approach enriches visual representations and improves the model’s detection performance while preserving compatibility with the baseline architecture. Experiments on the nuScenes dataset demonstrate that RCDINO achieves state-of-the-art performance among radar-camera models, with 56.4 NDS and 48.1 mAP. Our implementation is available at https://github.com/OlgaMatykina/RCDINO.

Method: The authors use attention knockouts as the primary analytical tool, designing three variants: Video Temporal Knockout, Video Spatial Knockout, and Language-to-Video Knockout. They apply these knockouts across different layers in both global and fine-grained settings to study model behavior.

Result: Key findings include: (1) Video information extraction occurs primarily in early layers with a two-stage process (perceptual encoding in lower layers, abstract reasoning in higher layers); (2) Certain intermediate layers have outsized impact while most contribute minimally; (3) Spatial-temporal modeling relies more on language-guided retrieval than on expensive intra-/inter-frame self-attention.

Conclusion: The study provides the first systematic understanding of Video-LLM internal processing mechanisms, offering both interpretability insights and practical efficiency improvements by demonstrating how attention computation can be reduced while maintaining performance.

Abstract: Taking advantage of large-scale data and pretrained language models, Video Large Language Models (Video-LLMs) have shown strong capabilities in answering video questions. However, most existing efforts focus on improving performance, with limited attention to understanding their internal mechanisms. This paper aims to bridge this gap through a systematic empirical study. To interpret existing VideoLLMs, we adopt attention knockouts as our primary analytical tool and design three variants: Video Temporal Knockout, Video Spatial Knockout, and Language-to-Video Knockout. Then, we apply these three knockouts on different numbers of layers (window of layers). By carefully controlling the window of layers and types of knockouts, we provide two settings: a global setting and a fine-grained setting. Our study reveals three key findings: (1) Global setting indicates Video information extraction primarily occurs in early layers, forming a clear two-stage process – lower layers focus on perceptual encoding, while higher layers handle abstract reasoning; (2) In the fine-grained setting, certain intermediate layers exert an outsized impact on video question answering, acting as critical outliers, whereas most other layers contribute minimally; (3) In both settings, we observe that spatial-temporal modeling relies more on language-guided retrieval than on intra- and inter-frame self-attention among video tokens, despite the latter’s high computational cost. Finally, we demonstrate that these insights can be leveraged to reduce attention computation in Video-LLMs. To our knowledge, this is the first work to systematically uncover how Video-LLMs internally process and understand video content, offering interpretability and efficiency perspectives for future research.

Method: BioTune uses an evolutionary algorithm to identify and selectively fine-tune only the most relevant layers in pre-trained models, optimizing for target task performance while minimizing trainable parameters.

Result: Evaluation across nine image classification datasets shows BioTune achieves competitive or improved accuracy and efficiency compared to existing methods like AutoRGN and LoRA, with reduced computational costs.

Conclusion: BioTune provides an effective evolutionary approach for selective layer fine-tuning that enhances transfer learning efficiency across diverse domains while maintaining performance and reducing computational demands.

Abstract: Deep learning has shown substantial progress in image analysis. However, the computational demands of large, fully trained models remain a consideration. Transfer learning offers a strategy for adapting pre-trained models to new tasks. Traditional fine-tuning often involves updating all model parameters, which can potentially lead to overfitting and higher computational costs. This paper introduces BioTune, an evolutionary adaptive fine-tuning technique that selectively fine-tunes layers to enhance transfer learning efficiency. BioTune employs an evolutionary algorithm to identify a focused set of layers for fine-tuning, aiming to optimize model performance on a given target task. Evaluation across nine image classification datasets from various domains indicates that BioTune achieves competitive or improved accuracy and efficiency compared to existing fine-tuning methods such as AutoRGN and LoRA. By concentrating the fine-tuning process on a subset of relevant layers, BioTune reduces the number of trainable parameters, potentially leading to decreased computational cost and facilitating more efficient transfer learning across diverse data characteristics and distributions.

Method: Proposes an Image-Conditioned Gaussian Splat Quantizer that jointly exploits inter-Gaussian and inter-attribute correlations, uses shared codebooks across scenes, and enables conditional decoding based on images captured at decoding time.

Result: ICGS-Quantizer reduces 3DGS storage requirements to kilobyte range while preserving visual fidelity and outperforms state-of-the-art methods in both compression efficiency and adaptability to scene changes.

Conclusion: The proposed method successfully addresses the limitations of current 3DGS compression by achieving significantly better compression ratios and providing adaptability to post-archival scene changes through image-conditioned decoding.

Abstract: 3D Gaussian Splatting (3DGS) has attracted considerable attention for enabling high-quality real-time rendering. Although 3DGS compression methods have been proposed for deployment on storage-constrained devices, two limitations hinder archival use: (1) they compress medium-scale scenes only to the megabyte range, which remains impractical for large-scale scenes or extensive scene collections; and (2) they lack mechanisms to accommodate scene changes after long-term archival. To address these limitations, we propose an Image-Conditioned Gaussian Splat Quantizer (ICGS-Quantizer) that substantially enhances compression efficiency and provides adaptability to scene changes after archiving. ICGS-Quantizer improves quantization efficiency by jointly exploiting inter-Gaussian and inter-attribute correlations and by using shared codebooks across all training scenes, which are then fixed and applied to previously unseen test scenes, eliminating the overhead of per-scene codebooks. This approach effectively reduces the storage requirements for 3DGS to the kilobyte range while preserving visual fidelity. To enable adaptability to post-archival scene changes, ICGS-Quantizer conditions scene decoding on images captured at decoding time. The encoding, quantization, and decoding processes are trained jointly, ensuring that the codes, which are quantized representations of the scene, are effective for conditional decoding. We evaluate ICGS-Quantizer on 3D scene compression and 3D scene updating. Experimental results show that ICGS-Quantizer consistently outperforms state-of-the-art methods in compression efficiency and adaptability to scene changes. Our code, model, and data will be publicly available on GitHub.

Method: Uses neural Gaussian representations with dynamic-static decoupling, region-wise voxel initialization (near/middle/far regions), deformable neural Gaussians for non-rigid dynamic actors with learnable deformation network, and supervision from depth and normal priors from pre-trained models.

Result: Achieves state-of-the-art performance on Waymo and KITTI datasets for novel-view synthesis in driving scenarios.

Conclusion: DriveSplat provides high-quality reconstruction for driving scenarios with improved geometric accuracy and robustness in novel view rendering through its comprehensive approach to handling both static and dynamic components.

Abstract: In the realm of driving scenarios, the presence of rapidly moving vehicles, pedestrians in motion, and large-scale static backgrounds poses significant challenges for 3D scene reconstruction. Recent methods based on 3D Gaussian Splatting address the motion blur problem by decoupling dynamic and static components within the scene. However, these decoupling strategies overlook background optimization with adequate geometry relationships and rely solely on fitting each training view by adding Gaussians. Therefore, these models exhibit limited robustness in rendering novel views and lack an accurate geometric representation. To address the above issues, we introduce DriveSplat, a high-quality reconstruction method for driving scenarios based on neural Gaussian representations with dynamic-static decoupling. To better accommodate the predominantly linear motion patterns of driving viewpoints, a region-wise voxel initialization scheme is employed, which partitions the scene into near, middle, and far regions to enhance close-range detail representation. Deformable neural Gaussians are introduced to model non-rigid dynamic actors, whose parameters are temporally adjusted by a learnable deformation network. The entire framework is further supervised by depth and normal priors from pre-trained models, improving the accuracy of geometric structures. Our method has been rigorously evaluated on the Waymo and KITTI datasets, demonstrating state-of-the-art performance in novel-view synthesis for driving scenarios.

Method: Used WIDER FACE dataset to evaluate three models (YOLOv11, YOLOv12, MTCNN) across multiple resolutions (160x160, 320x320, 640x640) using precision, recall, mAP50, mAP50-95, and inference time metrics.

Result: YOLOv11 outperformed others in detection accuracy especially at higher resolutions, YOLOv12 had slightly better recall, MTCNN was competitive in landmark localization but slower for real-time inference.

Conclusion: Provides actionable insights for selecting resolution-aware face detection models based on operational constraints, with YOLOv11 recommended for high-accuracy applications and considerations for resolution requirements.

Abstract: Face detection is a crucial component in many AI-driven applications such as surveillance, biometric authentication, and human-computer interaction. However, real-world conditions like low-resolution imagery present significant challenges that degrade detection performance. In this study, we systematically investigate the impact of input resolution on the accuracy and robustness of three prominent deep learning-based face detectors: YOLOv11, YOLOv12, and MTCNN. Using the WIDER FACE dataset, we conduct extensive evaluations across multiple image resolutions (160x160, 320x320, and 640x640) and assess each model’s performance using metrics such as precision, recall, mAP50, mAP50-95, and inference time. Results indicate that YOLOv11 outperforms YOLOv12 and MTCNN in terms of detection accuracy, especially at higher resolutions, while YOLOv12 exhibits slightly better recall. MTCNN, although competitive in landmark localization, lags in real-time inference speed. Our findings provide actionable insights for selecting resolution-aware face detection models suitable for varying operational constraints.

Method: Adopts a causal chain modeling perspective for RPM tasks, designs DIO baseline network architecture, but finds mutual information maximization inadequate. Proposes three progressive improvement methods to address the limitations.

Result: Experiments reveal that the initial optimization objective (maximizing variational lower bound of mutual information) fails to enable genuine acquisition of human reasoning logic due to lower bound tightness issues and statistical nature of mutual information.

Conclusion: The paper identifies fundamental limitations in using mutual information for abstract reasoning tasks and progressively develops three improvement methods to better capture causal relationships and human reasoning logic in RPM problems.

Abstract: Despite the outstanding performance of current deep learning models across various domains, their fundamental bottleneck in abstract reasoning remains unresolved. To address this challenge, the academic community has introduced Raven’s Progressive Matrices (RPM) problems as an authoritative benchmark for evaluating the abstract reasoning capabilities of deep learning algorithms, with a focus on core intelligence dimensions such as abstract reasoning, pattern recognition, and complex problem-solving. Therefore, this paper centers on solving RPM problems, aiming to contribute to enhancing the abstract reasoning abilities of machine intelligence. Firstly, this paper adopts a ``causal chain modeling’’ perspective to systematically analyze the complete causal chain in RPM tasks: image $\rightarrow$ abstract attributes $\rightarrow$ progressive attribute patterns $\rightarrow$ pattern consistency $\rightarrow$ correct answer. Based on this analysis, the network architecture of the baseline model DIO is designed. However, experiments reveal that the optimization objective formulated for DIO, namely maximizing the variational lower bound of mutual information between the context and the correct option, fails to enable the model to genuinely acquire the predefined human reasoning logic. This is attributed to two main reasons: the tightness of the lower bound significantly impacts the effectiveness of mutual information maximization, and mutual information, as a statistical measure, does not capture the causal relationship between subjects and objects. To overcome these limitations, this paper progressively proposes three improvement methods:

Method: Proposes Spiking Variational Graph Network with: 1) SNN-based keyframe extractor for autonomous feature learning, 2) Dynamic Aggregation Graph Reasoner for fine-grained reasoning, and 3) Variational Inference Reconstruction Module with ELBO optimization to handle uncertainty and noise.

Result: SpiVG surpasses existing methods across multiple datasets including SumMe, TVSum, VideoXum, and QFVS, demonstrating superior performance in video summarization.

Conclusion: The proposed SpiVG Network effectively addresses challenges in video summarization by combining spiking neural networks with variational inference, achieving state-of-the-art results while enhancing information density and reducing computational complexity.

Abstract: With the rise of short video content, efficient video summarization techniques for extracting key information have become crucial. However, existing methods struggle to capture the global temporal dependencies and maintain the semantic coherence of video content. Additionally, these methods are also influenced by noise during multi-channel feature fusion. We propose a Spiking Variational Graph (SpiVG) Network, which enhances information density and reduces computational complexity. First, we design a keyframe extractor based on Spiking Neural Networks (SNN), leveraging the event-driven computation mechanism of SNNs to learn keyframe features autonomously. To enable fine-grained and adaptable reasoning across video frames, we introduce a Dynamic Aggregation Graph Reasoner, which decouples contextual object consistency from semantic perspective coherence. We present a Variational Inference Reconstruction Module to address uncertainty and noise arising during multi-channel feature fusion. In this module, we employ Evidence Lower Bound Optimization (ELBO) to capture the latent structure of multi-channel feature distributions, using posterior distribution regularization to reduce overfitting. Experimental results show that SpiVG surpasses existing methods across multiple datasets such as SumMe, TVSum, VideoXum, and QFVS. Our codes and pre-trained models are available at https://github.com/liwrui/SpiVG.

Method: The authors analytically derive features learned by linear MAEs and extend this analysis to non-linear MAEs, examining how masking ratio and patch size influence the learning of spatial correlations at different ranges.

Result: The study shows that MAE hyperparameters (masking ratio and patch size) can selectively capture features representing short- and long-range spatial correlations, with MAE representations adapting to dataset-specific correlation structures beyond second-order statistics.

Conclusion: The analysis provides theoretical insights into MAE behavior and practical guidance for hyperparameter selection based on the spatial correlation properties of target datasets, helping optimize MAE performance for different vision tasks.

Abstract: Masked Autoencoders (MAEs) have emerged as a powerful pretraining technique for vision foundation models. Despite their effectiveness, they require extensive hyperparameter tuning (masking ratio, patch size, encoder/decoder layers) when applied to novel datasets. While prior theoretical works have analyzed MAEs in terms of their attention patterns and hierarchical latent variable models, the connection between MAE hyperparameters and performance on downstream tasks is relatively unexplored. This work investigates how MAEs learn spatial correlations in the input image. We analytically derive the features learned by a linear MAE and show that masking ratio and patch size can be used to select for features that capture short- and long-range spatial correlations. We extend this analysis to non-linear MAEs to show that MAE representations adapt to spatial correlations in the dataset, beyond second-order statistics. Finally, we discuss some insights on how to select MAE hyper-parameters in practice.

Method: Proposes bidirectional propagation strategy with Local Temporal Motion Fusion (LTMF) module for local spatio-temporal dependency modeling and Global Temporal Motion Fusion (GTMF) module for global feature aggregation. Uses both detection loss and Spatio-Temporal Fusion (STF) loss for joint optimization.

Result: Extensive experiments demonstrate state-of-the-art performance with fast inference speed.

Conclusion: BIRD effectively addresses limitations of sliding-window methods by leveraging both local and global temporal information through bidirectional propagation, achieving superior detection performance while maintaining computational efficiency.

Abstract: Moving infrared small target detection is broadly adopted in infrared search and track systems, and has attracted considerable research focus in recent years. The existing learning-based multi-frame methods mainly aggregate the information of adjacent frames in a sliding window fashion to assist the detection of the current frame. However, the sliding-window-based methods do not consider joint optimization of the entire video clip and ignore the global temporal information outside the sliding window, resulting in redundant computation and sub-optimal performance. In this paper, we propose a Bidirectional temporal information propagation method for moving InfraRed small target Detection, dubbed BIRD. The bidirectional propagation strategy simultaneously utilizes local temporal information of adjacent frames and global temporal information of past and future frames in a recursive fashion. Specifically, in the forward and backward propagation branches, we first design a Local Temporal Motion Fusion (LTMF) module to model local spatio-temporal dependency between a target frame and its two adjacent frames. Then, a Global Temporal Motion Fusion (GTMF) module is developed to further aggregate the global propagation feature with the local fusion feature. Finally, the bidirectional aggregated features are fused and input into the detection head for detection. In addition, the entire video clip is jointly optimized by the traditional detection loss and the additional Spatio-Temporal Fusion (STF) loss. Extensive experiments demonstrate that the proposed BIRD method not only achieves the state-of-the-art performance but also shows a fast inference speed.

Method: Used YOLOv8 object detection framework with custom variants (YOLOv8m-t4 and YOLOv8m-t42), trained on a bespoke dataset of annotated car surface images under various conditions with real-time data augmentation for robustness.

Result: YOLOv8m-t42 model achieved precision 0.86, recall 0.84, F1-score 0.85, mAP@0.5 of 0.60, and PR curve area of 0.88, outperforming YOLOv8m-t4 model (precision: 0.81, recall: 0.79, F1-score: 0.80).

Conclusion: The proposed deep learning approach provides an effective solution for real-time microscopic dent detection with high accuracy, making it suitable for practical applications like automated insurance assessments and vehicle inspections.

Abstract: Conventional car damage inspection techniques are labor-intensive, manual, and frequently overlook tiny surface imperfections like microscopic dents. Machine learning provides an innovative solution to the increasing demand for quicker and more precise inspection methods. The paper uses the YOLOv8 object recognition framework to provide a deep learning-based solution for automatically detecting microscopic surface flaws, notably tiny dents, on car exteriors. Traditional automotive damage inspection procedures are manual, time-consuming, and frequently unreliable at detecting tiny flaws. To solve this, a bespoke dataset containing annotated photos of car surfaces under various lighting circumstances, angles, and textures was created. To improve robustness, the YOLOv8m model and its customized variants, YOLOv8m-t4 and YOLOv8m-t42, were trained employing real-time data augmentation approaches. Experimental results show that the technique has excellent detection accuracy and low inference latency, making it suited for real-time applications such as automated insurance evaluations and automobile inspections. Evaluation parameters such as mean Average Precision (mAP), precision, recall, and F1-score verified the model’s efficacy. With a precision of 0.86, recall of 0.84, and F1-score of 0.85, the YOLOv8m-t42 model outperformed the YOLOv8m-t4 model (precision: 0.81, recall: 0.79, F1-score: 0.80) in identifying microscopic surface defects. With a little reduced mAP@0.5:0.95 of 0.20, the mAP@0.5 for YOLOv8m-t42 stabilized at 0.60. Furthermore, YOLOv8m-t42’s PR curve area was 0.88, suggesting more consistent performance than YOLOv8m-t4 (0.82). YOLOv8m-t42 has greater accuracy and is more appropriate for practical dent detection applications, even though its convergence is slower.

Method: MATR uses transformer architecture with dual-stage sequence alignment to condition target video representations on query features, plus self-supervised pre-training by localizing random clips within videos.

Result: Achieved 13.1% R@1 and 8.1% mIoU improvement on ActivityNet-VRL, and 14.7% R@1 and 14.4% mIoU gain on new SportsMoments dataset over state-of-the-art methods.

Conclusion: MATR effectively addresses the challenges of video-to-video moment retrieval through transformer-based alignment and self-supervised pre-training, demonstrating substantial performance gains across multiple datasets.

Abstract: Video-to-video moment retrieval (Vid2VidMR) is the task of localizing unseen events or moments in a target video using a query video. This task poses several challenges, such as the need for semantic frame-level alignment and modeling complex dependencies between query and target videos. To tackle this challenging problem, we introduce MATR (Moment Alignment TRansformer), a transformer-based model designed to capture semantic context as well as the temporal details necessary for precise moment localization. MATR conditions target video representations on query video features using dual-stage sequence alignment that encodes the required correlations and dependencies. These representations are then used to guide foreground/background classification and boundary prediction heads, enabling the model to accurately identify moments in the target video that semantically match with the query video. Additionally, to provide a strong task-specific initialization for MATR, we propose a self-supervised pre-training technique that involves training the model to localize random clips within videos. Extensive experiments demonstrate that MATR achieves notable performance improvements of 13.1% in R@1 and 8.1% in mIoU on an absolute scale compared to state-of-the-art methods on the popular ActivityNet-VRL dataset. Additionally, on our newly proposed dataset, SportsMoments, MATR shows a 14.7% gain in R@1 and a 14.4% gain in mIoU on an absolute scale over strong baselines.

Method: Proposes dense stereo module for camera pose estimation and global dense point cloud initialization, coherent view interpolation for additional supervision, and multi-scale regularization techniques for enhanced geometry and rendering quality.

Result: Achieves 2.75dB PSNR improvement under extremely sparse-view conditions (2 training views), with minimal distortion and preserved high-frequency details, outperforming other state-of-the-art 3DGS approaches.

Conclusion: The proposed SfM-free approach successfully addresses the limitations of traditional 3DGS in sparse-view scenarios, enabling high-quality novel view synthesis from minimal input views without requiring precise camera pose information.

Abstract: 3D Gaussian Splatting (3DGS) has demonstrated remarkable real-time performance in novel view synthesis, yet its effectiveness relies heavily on dense multi-view inputs with precisely known camera poses, which are rarely available in real-world scenarios. When input views become extremely sparse, the Structure-from-Motion (SfM) method that 3DGS depends on for initialization fails to accurately reconstruct the 3D geometric structures of scenes, resulting in degraded rendering quality. In this paper, we propose a novel SfM-free 3DGS-based method that jointly estimates camera poses and reconstructs 3D scenes from extremely sparse-view inputs. Specifically, instead of SfM, we propose a dense stereo module to progressively estimates camera pose information and reconstructs a global dense point cloud for initialization. To address the inherent problem of information scarcity in extremely sparse-view settings, we propose a coherent view interpolation module that interpolates camera poses based on training view pairs and generates viewpoint-consistent content as additional supervision signals for training. Furthermore, we introduce multi-scale Laplacian consistent regularization and adaptive spatial-aware multi-scale geometry regularization to enhance the quality of geometrical structures and rendered content. Experiments show that our method significantly outperforms other state-of-the-art 3DGS-based approaches, achieving a remarkable 2.75dB improvement in PSNR under extremely sparse-view conditions (using only 2 training views). The images synthesized by our method exhibit minimal distortion while preserving rich high-frequency details, resulting in superior visual quality compared to existing techniques.

Method: Proposes LGMSNet framework with: 1) heterogeneous intra-layer kernels to extract local high-frequency information while reducing channel redundancy, and 2) sparse transformer-convolutional hybrid branches to capture low-frequency global information.

Result: Extensive experiments on six public datasets show LGMSNet outperforms state-of-the-art methods. It maintains exceptional performance in zero-shot generalization tests on four unseen datasets, demonstrating strong real-world applicability.

Conclusion: LGMSNet provides an effective lightweight solution for medical image segmentation that balances performance and efficiency, with excellent generalization capabilities suitable for deployment in resource-limited medical scenarios.

Abstract: Medical image segmentation plays a pivotal role in disease diagnosis and treatment planning, particularly in resource-constrained clinical settings where lightweight and generalizable models are urgently needed. However, existing lightweight models often compromise performance for efficiency and rarely adopt computationally expensive attention mechanisms, severely restricting their global contextual perception capabilities. Additionally, current architectures neglect the channel redundancy issue under the same convolutional kernels in medical imaging, which hinders effective feature extraction. To address these challenges, we propose LGMSNet, a novel lightweight framework based on local and global dual multiscale that achieves state-of-the-art performance with minimal computational overhead. LGMSNet employs heterogeneous intra-layer kernels to extract local high-frequency information while mitigating channel redundancy. In addition, the model integrates sparse transformer-convolutional hybrid branches to capture low-frequency global information. Extensive experiments across six public datasets demonstrate LGMSNet’s superiority over existing state-of-the-art methods. In particular, LGMSNet maintains exceptional performance in zero-shot generalization tests on four unseen datasets, underscoring its potential for real-world deployment in resource-limited medical scenarios. The whole project code is in https://github.com/cq-dong/LGMSNet.

Method: Uses Joint Multi-view Body Optimization (JMBO) to fit SMPL-X body model across all views with multi-view consistency, then adds geometric details via normal map integration from front and back views.

Result: Consistently improves fidelity over single-view baseline and achieves competitive performance compared to modern few-shot 3D reconstruction methods.

Conclusion: MExECON successfully extends single-view reconstruction to multi-view scenarios, achieving better geometry and detail capture for clothed human avatars without requiring network retraining.

Abstract: This work presents MExECON, a novel pipeline for 3D reconstruction of clothed human avatars from sparse multi-view RGB images. Building on the single-view method ECON, MExECON extends its capabilities to leverage multiple viewpoints, improving geometry and body pose estimation. At the core of the pipeline is the proposed Joint Multi-view Body Optimization (JMBO) algorithm, which fits a single SMPL-X body model jointly across all input views, enforcing multi-view consistency. The optimized body model serves as a low-frequency prior that guides the subsequent surface reconstruction, where geometric details are added via normal map integration. MExECON integrates normal maps from both front and back views to accurately capture fine-grained surface details such as clothing folds and hairstyles. All multi-view gains are achieved without requiring any network re-training. Experimental results show that MExECON consistently improves fidelity over the single-view baseline and achieves competitive performance compared to modern few-shot 3D reconstruction methods.

Method: Proposes AdaSFFuse with two key innovations: Adaptive Approximate Wavelet Transform (AdaWAT) for frequency decoupling, and Spatial-Frequency Mamba Blocks for efficient cross-domain fusion in both spatial and frequency domains.

Result: Superior fusion performance demonstrated on four MMIF tasks (IVF, MFF, MEF, MIF) with low computational cost and compact network architecture, achieving good balance between performance and efficiency.

Conclusion: AdaSFFuse effectively addresses MMIF challenges through adaptive cross-domain co-fusion learning, improving feature alignment, reducing frequency loss, and preserving critical details across diverse modalities.

Abstract: Multimodal Image Fusion (MMIF) aims to integrate complementary information from different imaging modalities to overcome the limitations of individual sensors. It enhances image quality and facilitates downstream applications such as remote sensing, medical diagnostics, and robotics. Despite significant advancements, current MMIF methods still face challenges such as modality misalignment, high-frequency detail destruction, and task-specific limitations. To address these challenges, we propose AdaSFFuse, a novel framework for task-generalized MMIF through adaptive cross-domain co-fusion learning. AdaSFFuse introduces two key innovations: the Adaptive Approximate Wavelet Transform (AdaWAT) for frequency decoupling, and the Spatial-Frequency Mamba Blocks for efficient multimodal fusion. AdaWAT adaptively separates the high- and low-frequency components of multimodal images from different scenes, enabling fine-grained extraction and alignment of distinct frequency characteristics for each modality. The Spatial-Frequency Mamba Blocks facilitate cross-domain fusion in both spatial and frequency domains, enhancing this process. These blocks dynamically adjust through learnable mappings to ensure robust fusion across diverse modalities. By combining these components, AdaSFFuse improves the alignment and integration of multimodal features, reduces frequency loss, and preserves critical details. Extensive experiments on four MMIF tasks – Infrared-Visible Image Fusion (IVF), Multi-Focus Image Fusion (MFF), Multi-Exposure Image Fusion (MEF), and Medical Image Fusion (MIF) – demonstrate AdaSFFuse’s superior fusion performance, ensuring both low computational cost and a compact network, offering a strong balance between performance and efficiency. The code will be publicly available at https://github.com/Zhen-yu-Liu/AdaSFFuse.

Method: Proposes ExtraGS framework with Road Surface Gaussian representation (hybrid Gaussian-SDF design), Far Field Gaussians with learnable scaling factors, and self-supervised uncertainty estimation using spherical harmonics for selective generative prior integration.

Result: Extensive experiments show ExtraGS significantly enhances realism and geometric consistency of extrapolated views while maintaining high fidelity along original trajectories across multiple datasets and camera setups.

Conclusion: ExtraGS provides an effective holistic solution for trajectory extrapolation that successfully integrates both geometric and generative priors to overcome limitations of previous methods.

Abstract: Synthesizing extrapolated views from recorded driving logs is critical for simulating driving scenes for autonomous driving vehicles, yet it remains a challenging task. Recent methods leverage generative priors as pseudo ground truth, but often lead to poor geometric consistency and over-smoothed renderings. To address these limitations, we propose ExtraGS, a holistic framework for trajectory extrapolation that integrates both geometric and generative priors. At the core of ExtraGS is a novel Road Surface Gaussian(RSG) representation based on a hybrid Gaussian-Signed Distance Function (SDF) design, and Far Field Gaussians (FFG) that use learnable scaling factors to efficiently handle distant objects. Furthermore, we develop a self-supervised uncertainty estimation framework based on spherical harmonics that enables selective integration of generative priors only where extrapolation artifacts occur. Extensive experiments on multiple datasets, diverse multi-camera setups, and various generative priors demonstrate that ExtraGS significantly enhances the realism and geometric consistency of extrapolated views, while preserving high fidelity along the original trajectory.

Method: Two-stage pipeline: 1) Motion Initialization - interactively divide sketch into semantic groups and define key frames for coarse animation via interpolation; 2) Motion Refinement - use Group-based Displacement Network (GDN) with Context-conditioned Feature Enhancement to predict group-specific displacement fields using text-to-video model priors.

Result: Extensive experiments show GroupSketch significantly outperforms existing methods in generating high-quality, temporally consistent animations for complex multi-object sketches.

Conclusion: The method expands practical applications of sketch animation by effectively handling multi-object interactions and complex motions with improved temporal consistency.

Abstract: We introduce GroupSketch, a novel method for vector sketch animation that effectively handles multi-object interactions and complex motions. Existing approaches struggle with these scenarios, either being limited to single-object cases or suffering from temporal inconsistency and poor generalization. To address these limitations, our method adopts a two-stage pipeline comprising Motion Initialization and Motion Refinement. In the first stage, the input sketch is interactively divided into semantic groups and key frames are defined, enabling the generation of a coarse animation via interpolation. In the second stage, we propose a Group-based Displacement Network (GDN), which refines the coarse animation by predicting group-specific displacement fields, leveraging priors from a text-to-video model. GDN further incorporates specialized modules, such as Context-conditioned Feature Enhancement (CCFE), to improve temporal consistency. Extensive experiments demonstrate that our approach significantly outperforms existing methods in generating high-quality, temporally consistent animations for complex, multi-object sketches, thus expanding the practical applications of sketch animation.

Method: D3FNet builds on D-LinkNet with three innovations: 1) Differential Attention Dilation Extraction module for enhanced road features, 2) Dual-stream Decoding Fusion Mechanism for spatial-semantic balance, and 3) multi-scale dilation strategy (rates 1,3,5,9) to reduce artifacts.

Result: Extensive experiments on DeepGlobe and CHN6-CUG benchmarks show superior IoU and recall on challenging road regions, outperforming state-of-the-art baselines.

Conclusion: D3FNet provides a robust solution for fine-grained narrow road extraction in complex remote and cooperative perception scenarios, with ablation studies confirming the synergy of attention-guided encoding and dual-path decoding.

Abstract: Extracting narrow roads from high-resolution remote sensing imagery remains a significant challenge due to their limited width, fragmented topology, and frequent occlusions. To address these issues, we propose D3FNet, a Dilated Dual-Stream Differential Attention Fusion Network designed for fine-grained road structure segmentation in remote perception systems. Built upon the encoder-decoder backbone of D-LinkNet, D3FNet introduces three key innovations:(1) a Differential Attention Dilation Extraction (DADE) module that enhances subtle road features while suppressing background noise at the bottleneck; (2) a Dual-stream Decoding Fusion Mechanism (DDFM) that integrates original and attention-modulated features to balance spatial precision with semantic context; and (3) a multi-scale dilation strategy (rates 1, 3, 5, 9) that mitigates gridding artifacts and improves continuity in narrow road prediction. Unlike conventional models that overfit to generic road widths, D3FNet specifically targets fine-grained, occluded, and low-contrast road segments. Extensive experiments on the DeepGlobe and CHN6-CUG benchmarks show that D3FNet achieves superior IoU and recall on challenging road regions, outperforming state-of-the-art baselines. Ablation studies further verify the complementary synergy of attention-guided encoding and dual-path decoding. These results confirm D3FNet as a robust solution for fine-grained narrow road extraction in complex remote and cooperative perception scenarios.

Method: Reframes TTA as Gaussian probabilistic inference task using gradually updated class means and shared covariance matrix. Uses lightweight regularization with CLIP priors and historical knowledge bank to correct likelihood bias. No source data, gradient updates, or full target data access required.

Result: Achieves state-of-the-art performance across diverse benchmarks under various distribution shifts with superior scalability and robustness.

Conclusion: ADAPT provides an effective backpropagation-free solution for test-time adaptation that enables closed-form, training-free inference while maintaining high performance and scalability.

Abstract: Test-time adaptation (TTA) enhances the zero-shot robustness under distribution shifts by leveraging unlabeled test data during inference. Despite notable advances, several challenges still limit its broader applicability. First, most methods rely on backpropagation or iterative optimization, which limits scalability and hinders real-time deployment. Second, they lack explicit modeling of class-conditional feature distributions. This modeling is crucial for producing reliable decision boundaries and calibrated predictions, but it remains underexplored due to the lack of both source data and supervision at test time. In this paper, we propose ADAPT, an Advanced Distribution-Aware and backPropagation-free Test-time adaptation method. We reframe TTA as a Gaussian probabilistic inference task by modeling class-conditional likelihoods using gradually updated class means and a shared covariance matrix. This enables closed-form, training-free inference. To correct potential likelihood bias, we introduce lightweight regularization guided by CLIP priors and a historical knowledge bank. ADAPT requires no source data, no gradient updates, and no full access to target data, supporting both online and transductive settings. Extensive experiments across diverse benchmarks demonstrate that our method achieves state-of-the-art performance under a wide range of distribution shifts with superior scalability and robustness.

Method: Two-stage training: 1) Use neighbor-aware soft mask to assign higher weights to pixels with strong local variations, encouraging early focus on fine details 2) Transition to full-image training

Result: Consistently improves reconstruction quality and complements existing INR methods

Conclusion: Pioneering approach that offers a new avenue for mitigating spectral bias by assigning frequency-aware importance to pixels in image INR training

Abstract: Implicit Neural Representations (INRs) have emerged as a powerful alternative to traditional pixel-based formats by modeling images as continuous functions over spatial coordinates. A key challenge, however, lies in the spectral bias of neural networks, which tend to favor low-frequency components while struggling to capture high-frequency (HF) details such as sharp edges and fine textures. While prior approaches have addressed this limitation through architectural modifications or specialized activation functions, we propose an orthogonal direction by directly guiding the training process. Specifically, we introduce a two-stage training strategy where a neighbor-aware soft mask adaptively assigns higher weights to pixels with strong local variations, encouraging early focus on fine details. The model then transitions to full-image training. Experimental results show that our approach consistently improves reconstruction quality and complements existing INR methods. As a pioneering attempt to assign frequency-aware importance to pixels in image INR, our work offers a new avenue for mitigating the spectral bias problem.

Method: Proposed a linear time convex relaxation approach combined with a contractor method to accelerate Branch and Bound (BnB) optimization for the truncated least squares formulation.

Result: The solver achieves provable global optimality in less than 0.5 seconds for 100-point 3D registration when rotation axis is provided - 100x faster than STRIDE SDP solver for rotation-only TLS problems.

Conclusion: The method provides dramatic speed improvements while maintaining global optimality guarantees, though currently limited to problems with known rotation axis rather than full 6DoF registration.

Abstract: Recent results showed that point cloud registration with given correspondences can be made robust to outlier rates of up to 95% using the truncated least squares (TLS) formulation. However, solving this combinatorial optimization problem to global optimality is challenging. Provably globally optimal approaches using semidefinite programming (SDP) relaxations take hundreds of seconds for 100 points. In this paper, we propose a novel linear time convex relaxation as well as a contractor method to speed up Branch and Bound (BnB). Our solver can register two 3D point clouds with 100 points to provable global optimality in less than half a second when the axis of rotation is provided. Although it currently cannot solve the full 6DoF problem, it is two orders of magnitude faster than the state-of-the-art SDP solver STRIDE when solving the rotation-only TLS problem. In addition to providing a formal proof for global optimality, we present empirical evidence of global optimality using adversarial instances with local minimas close to the global minimum.

Method: Combined visible/NIR camera traps and thermal infrared drones for data collection. Built deep learning models (YOLOv11s and enhanced Faster RCNN) for automated detection. Performed spatial pattern analysis to identify activity hotspots and conflict zones.

Result: YOLOv11s achieved 96.2% precision, 92.3% recall, 96.7% mAP50 for camera trap detection. Drone thermal imagery provided complementary aerial perspective. Spatial analysis identified clear wildlife and human activity hotspots with overlapping patterns indicating potential conflict zones.

Conclusion: Integrating multiperspective monitoring with automated object detection enhances wildlife surveillance and landscape management by revealing human-wildlife conflicts within conserved landscapes.

Abstract: Wildlife and human activities are key components of landscape systems. Understanding their spatial distribution is essential for evaluating human wildlife interactions and informing effective conservation planning. Multiperspective monitoring of wildlife and human activities by combining camera traps and drone imagery. Capturing the spatial patterns of their distributions, which allows the identification of the overlap of their activity zones and the assessment of the degree of human wildlife conflict. The study was conducted in Chitwan National Park (CNP), Nepal, and adjacent regions. Images collected by visible and nearinfrared camera traps and thermal infrared drones from February to July 2022 were processed to create training and testing datasets, which were used to build deep learning models to automatic identify wildlife and human activities. Drone collected thermal imagery was used for detecting targets to provide a multiple monitoring perspective. Spatial pattern analysis was performed to identify animal and resident activity hotspots and delineation potential human wildlife conflict zones. Among the deep learning models tested, YOLOv11s achieved the highest performance with a precision of 96.2%, recall of 92.3%, mAP50 of 96.7%, and mAP50 of 81.3%, making it the most effective for detecting objects in camera trap imagery. Drone based thermal imagery, analyzed with an enhanced Faster RCNN model, added a complementary aerial viewpoint for camera trap detections. Spatial pattern analysis identified clear hotspots for both wildlife and human activities and their overlapping patterns within certain areas in the CNP and buffer zones indicating potential conflict. This study reveals human wildlife conflicts within the conserved landscape. Integrating multiperspective monitoring with automated object detection enhances wildlife surveillance and landscape management.

Method: Three key innovations: 1) Diffusion Temporal Latent encoder for boundary sensitivity and temporal consistency, 2) Object grounded representations that bind query entities to visual evidence, 3) Mixed token scheme with discrete temporal tokens for explicit timestamp modeling.

Result: Achieves state-of-the-art results on Charades STA, NExT GQA, and multiple VideoQA benchmarks, demonstrating robust grounding capabilities.

Conclusion: Grounded VideoDiT successfully overcomes limitations in temporal perception for Video LLMs through explicit temporal modeling and entity grounding, enabling fine-grained video reasoning.

Abstract: Understanding videos requires more than answering open ended questions, it demands the ability to pinpoint when events occur and how entities interact across time. While recent Video LLMs have achieved remarkable progress in holistic reasoning, they remain coarse in temporal perception: timestamps are encoded only implicitly, frame level features are weak in capturing continuity, and language vision alignment often drifts from the entities of interest. In this paper, we present Grounded VideoDiT, a Video LLM designed to overcome these limitations by introducing three key innovations. First, a Diffusion Temporal Latent (DTL) encoder enhances boundary sensitivity and maintains temporal consistency. Second, object grounded representations explicitly bind query entities to localized visual evidence, strengthening alignment. Third, a mixed token scheme with discrete temporal tokens provides explicit timestamp modeling, enabling fine grained temporal reasoning. Together, these designs equip Grounded VideoDiT with robust grounding capabilities, as validated by state of the art results on Charades STA, NExT GQA, and multiple VideoQA benchmarks.

Method: Human operators provide quality ratings on initial segmentation results, which train a rating model to classify segmentation outputs. The segmentation model is then finetuned using feedback from this rating model.

Result: Improved the original segmentation model by 34% in correctly identified tree instances while reducing non-tree instance predictions. Performance reduced in sparsely forested regions with small trees (<2m) or complex surroundings.

Conclusion: The weakly supervised approach effectively improves tree instance segmentation with reduced labeling costs, though challenges remain in complex terrain with small trees and shrubs.

Abstract: Tree instance segmentation of airborne laser scanning (ALS) data is of utmost importance for forest monitoring, but remains challenging due to variations in the data caused by factors such as sensor resolution, vegetation state at acquisition time, terrain characteristics, etc. Moreover, obtaining a sufficient amount of precisely labeled data to train fully supervised instance segmentation methods is expensive. To address these challenges, we propose a weakly supervised approach where labels of an initial segmentation result obtained either by a non-finetuned model or a closed form algorithm are provided as a quality rating by a human operator. The labels produced during the quality assessment are then used to train a rating model, whose task is to classify a segmentation output into the same classes as specified by the human operator. Finally, the segmentation model is finetuned using feedback from the rating model. This in turn improves the original segmentation model by 34% in terms of correctly identified tree instances while considerably reducing the number of non-tree instances predicted. Challenges still remain in data over sparsely forested regions characterized by small trees (less than two meters in height) or within complex surroundings containing shrubs, boulders, etc. which can be confused as trees where the performance of the proposed method is reduced.

Method: Proposes CFG method that computes feature importance to prioritize corruption of critical features likely adopted by diverse architectures. Uses loss function with explicit constraints on maximum deviation to ensure imperceptibility.

Result: Extensive experiments on ModelNet40 and ScanObjectNN datasets show CFG outperforms state-of-the-art attack methods by a large margin.

Conclusion: CFG successfully improves transferability of adversarial point clouds by leveraging critical feature guidance and maintaining imperceptibility, demonstrating effectiveness in black-box attack scenarios.

Abstract: Deep neural networks for 3D point clouds have been demonstrated to be vulnerable to adversarial examples. Previous 3D adversarial attack methods often exploit certain information about the target models, such as model parameters or outputs, to generate adversarial point clouds. However, in realistic scenarios, it is challenging to obtain any information about the target models under conditions of absolute security. Therefore, we focus on transfer-based attacks, where generating adversarial point clouds does not require any information about the target models. Based on our observation that the critical features used for point cloud classification are consistent across different DNN architectures, we propose CFG, a novel transfer-based black-box attack method that improves the transferability of adversarial point clouds via the proposed Critical Feature Guidance. Specifically, our method regularizes the search of adversarial point clouds by computing the importance of the extracted features, prioritizing the corruption of critical features that are likely to be adopted by diverse architectures. Further, we explicitly constrain the maximum deviation extent of the generated adversarial point clouds in the loss function to ensure their imperceptibility. Extensive experiments conducted on the ModelNet40 and ScanObjectNN benchmark datasets demonstrate that the proposed CFG outperforms the state-of-the-art attack methods by a large margin.

Method: Proposes MapKD with Teacher-Coach-Student paradigm: 1) multimodal teacher with map priors, 2) vision-centric coach with simulated LiDAR to bridge modality gap, 3) lightweight student. Uses Token-Guided 2D Patch Distillation and Masked Semantic Response Distillation for knowledge transfer.

Result: On nuScenes dataset, improves student model by +6.68 mIoU and +10.94 mAP while accelerating inference speed compared to baseline vision-only approaches.

Conclusion: MapKD effectively transfers knowledge from multimodal models to vision-only student models, enabling efficient and high-performance online HD map construction without dependency on expensive sensor suites or stale map data.

Abstract: Online HD map construction is a fundamental task in autonomous driving systems, aiming to acquire semantic information of map elements around the ego vehicle based on real-time sensor inputs. Recently, several approaches have achieved promising results by incorporating offline priors such as SD maps and HD maps or by fusing multi-modal data. However, these methods depend on stale offline maps and multi-modal sensor suites, resulting in avoidable computational overhead at inference. To address these limitations, we employ a knowledge distillation strategy to transfer knowledge from multimodal models with prior knowledge to an efficient, low-cost, and vision-centric student model. Specifically, we propose MapKD, a novel multi-level cross-modal knowledge distillation framework with an innovative Teacher-Coach-Student (TCS) paradigm. This framework consists of: (1) a camera-LiDAR fusion model with SD/HD map priors serving as the teacher; (2) a vision-centric coach model with prior knowledge and simulated LiDAR to bridge the cross-modal knowledge transfer gap; and (3) a lightweight vision-based student model. Additionally, we introduce two targeted knowledge distillation strategies: Token-Guided 2D Patch Distillation (TGPD) for bird’s eye view feature alignment and Masked Semantic Response Distillation (MSRD) for semantic learning guidance. Extensive experiments on the challenging nuScenes dataset demonstrate that MapKD improves the student model by +6.68 mIoU and +10.94 mAP while simultaneously accelerating inference speed. The code is available at:https://github.com/2004yan/MapKD2026.

Method: Multi-dimensional Dynamic Prompt Routing (MDPR) constructs comprehensive knowledge base across five visual-semantic dimensions, uses dynamic routing to align global classes, retrieve optimal prompts, and balance fine-grained semantics with logits fusion.

Result: Achieves comparable results with SOTA methods on long-tailed benchmarks (CIFAR-LT, ImageNet-LT, Places-LT), shows effectiveness for tail classes, and incurs minimal computational overhead.

Conclusion: MDPR provides flexible and efficient enhancement for VLM fine-tuning under data imbalance through semantic library and dynamic routing mechanism.

Abstract: Pre-trained vision-language models (VLMs), such as CLIP, have demonstrated impressive capability in visual tasks, but their fine-tuning often suffers from bias in class-imbalanced scene. Recent works have introduced large language models (LLMs) to enhance VLM fine-tuning with supplementing semantic information. However, they often overlook inherent class imbalance in VLMs’ pre-training, which may lead to bias accumulation in downstream tasks. To address this problem, this paper proposes a Multi-dimensional Dynamic Prompt Routing (MDPR) framework. MDPR constructs a comprehensive knowledge base for classes, spanning five visual-semantic dimensions. During fine-tuning, the dynamic routing mechanism aligns global visual classes, retrieves optimal prompts, and balances fine-grained semantics, yielding stable predictions through logits fusion. Extensive experiments on long-tailed benchmarks, including CIFAR-LT, ImageNet-LT, and Places-LT, demonstrate that MDPR achieves comparable results with current SOTA methods. Ablation studies further confirm the effectiveness of our semantic library for tail classes, and show that our dynamic routing incurs minimal computational overhead, making MDPR a flexible and efficient enhancement for VLM fine-tuning under data imbalance.

Method: StreamMem encodes new video frames in a streaming manner and compresses KV cache using attention scores between visual tokens and generic query tokens, while maintaining a fixed-size KV memory for efficient question answering in memory-constrained scenarios.

Result: Evaluation on three long video understanding and two streaming video QA benchmarks shows StreamMem achieves state-of-the-art performance in query-agnostic KV cache compression and is competitive with query-aware approaches.

Conclusion: StreamMem provides an efficient solution for streaming video understanding in MLLMs by enabling query-agnostic KV cache compression with fixed memory usage, making long-video processing practical without requiring prior question knowledge.

Abstract: Multimodal large language models (MLLMs) have made significant progress in visual-language reasoning, but their ability to efficiently handle long videos remains limited. Despite recent advances in long-context MLLMs, storing and attending to the key-value (KV) cache for long visual contexts incurs substantial memory and computational overhead. Existing visual compression methods require either encoding the entire visual context before compression or having access to the questions in advance, which is impractical for long video understanding and multi-turn conversational settings. In this work, we propose StreamMem, a query-agnostic KV cache memory mechanism for streaming video understanding. Specifically, StreamMem encodes new video frames in a streaming manner, compressing the KV cache using attention scores between visual tokens and generic query tokens, while maintaining a fixed-size KV memory to enable efficient question answering (QA) in memory-constrained, long-video scenarios. Evaluation on three long video understanding and two streaming video question answering benchmarks shows that StreamMem achieves state-of-the-art performance in query-agnostic KV cache compression and is competitive with query-aware compression approaches.

Method: The framework jointly predicts perceptual conditions and color information from unified representations, leverages depth cues to create a memory bank for clearer context, and uses segmented noise scheduling for training prediction groups to reduce drift and computational costs.

Result: Extensive experiments on diffusion- and rectified flow-based models demonstrate that WorldWeaver effectively reduces temporal drift and improves the fidelity of generated videos in long-horizon scenarios.

Conclusion: WorldWeaver provides a robust solution for long video generation by combining joint modeling of RGB and perceptual conditions, depth-based memory preservation, and segmented noise scheduling, significantly enhancing temporal consistency and video quality.

Abstract: Generative video modeling has made significant strides, yet ensuring structural and temporal consistency over long sequences remains a challenge. Current methods predominantly rely on RGB signals, leading to accumulated errors in object structure and motion over extended durations. To address these issues, we introduce WorldWeaver, a robust framework for long video generation that jointly models RGB frames and perceptual conditions within a unified long-horizon modeling scheme. Our training framework offers three key advantages. First, by jointly predicting perceptual conditions and color information from a unified representation, it significantly enhances temporal consistency and motion dynamics. Second, by leveraging depth cues, which we observe to be more resistant to drift than RGB, we construct a memory bank that preserves clearer contextual information, improving quality in long-horizon video generation. Third, we employ segmented noise scheduling for training prediction groups, which further mitigates drift and reduces computational cost. Extensive experiments on both diffusion- and rectified flow-based models demonstrate the effectiveness of WorldWeaver in reducing temporal drift and improving the fidelity of generated videos.

Method: Full-stack approach with: (1) molecular-empowered learning for lay annotators, (2) SAM adapter for semantic adaptation, and (3) Molecular-Oriented Corrective Learning (MOCL) for refinement. Combines vision foundation models with molecular data.

Result: Significant improvement in cell classification performance across both in-house and public datasets, even with varying annotation quality.

Conclusion: Reduces annotator workload and extends precise biomedical image analysis to resource-limited settings, advancing medical diagnostics and pathology automation.

Abstract: Purpose: Recent developments in computational pathology have been driven by advances in Vision Foundation Models, particularly the Segment Anything Model (SAM). This model facilitates nuclei segmentation through two primary methods: prompt-based zero-shot segmentation and the use of cell-specific SAM models for direct segmentation. These approaches enable effective segmentation across a range of nuclei and cells. However, general vision foundation models often face challenges with fine-grained semantic segmentation, such as identifying specific nuclei subtypes or particular cells. Approach: In this paper, we propose the molecular-empowered All-in-SAM Model to advance computational pathology by leveraging the capabilities of vision foundation models. This model incorporates a full-stack approach, focusing on: (1) annotation-engaging lay annotators through molecular-empowered learning to reduce the need for detailed pixel-level annotations, (2) learning-adapting the SAM model to emphasize specific semantics, which utilizes its strong generalizability with SAM adapter, and (3) refinement-enhancing segmentation accuracy by integrating Molecular-Oriented Corrective Learning (MOCL). Results: Experimental results from both in-house and public datasets show that the All-in-SAM model significantly improves cell classification performance, even when faced with varying annotation quality. Conclusions: Our approach not only reduces the workload for annotators but also extends the accessibility of precise biomedical image analysis to resource-limited settings, thereby advancing medical diagnostics and automating pathology image analysis.

Method: Uses Hybrid Stream DiT architecture for modality alignment and faster training convergence, implements comprehensive data curation pipeline with MLLM-based video quality filtering, and provides detailed training/inference recipes.

Result: Achieves Top 3 ranking on both T2V and I2V leaderboards, outperforms open-source models and matches/surpasses commercial solutions in motion amplitude and temporal consistency.

Conclusion: Waver represents a significant advancement in video generation technology, providing an efficient framework for training high-quality models and accelerating progress in the field.

Abstract: We present Waver, a high-performance foundation model for unified image and video generation. Waver can directly generate videos with durations ranging from 5 to 10 seconds at a native resolution of 720p, which are subsequently upscaled to 1080p. The model simultaneously supports text-to-video (T2V), image-to-video (I2V), and text-to-image (T2I) generation within a single, integrated framework. We introduce a Hybrid Stream DiT architecture to enhance modality alignment and accelerate training convergence. To ensure training data quality, we establish a comprehensive data curation pipeline and manually annotate and train an MLLM-based video quality model to filter for the highest-quality samples. Furthermore, we provide detailed training and inference recipes to facilitate the generation of high-quality videos. Building on these contributions, Waver excels at capturing complex motion, achieving superior motion amplitude and temporal consistency in video synthesis. Notably, it ranks among the Top 3 on both the T2V and I2V leaderboards at Artificial Analysis (data as of 2025-07-30 10:00 GMT+8), consistently outperforming existing open-source models and matching or surpassing state-of-the-art commercial solutions. We hope this technical report will help the community more efficiently train high-quality video generation models and accelerate progress in video generation technologies. Official page: https://github.com/FoundationVision/Waver.

Method: Learned from 600k high-resolution scans using 240 synchronized cameras, ATLAS explicitly decouples shape and skeleton bases by grounding mesh representation in the human skeleton, using non-linear pose correctives.

Result: ATLAS outperforms existing methods by fitting unseen subjects in diverse poses more accurately, with quantitative evaluations showing non-linear pose correctives capture complex poses better than linear models.

Conclusion: The skeleton-shape decoupling approach enables enhanced shape expressivity, fine-grained customization, and keypoint fitting independent of soft-tissue characteristics, representing a significant advancement in parametric body modeling.

Abstract: Parametric body models offer expressive 3D representation of humans across a wide range of poses, shapes, and facial expressions, typically derived by learning a basis over registered 3D meshes. However, existing human mesh modeling approaches struggle to capture detailed variations across diverse body poses and shapes, largely due to limited training data diversity and restrictive modeling assumptions. Moreover, the common paradigm first optimizes the external body surface using a linear basis, then regresses internal skeletal joints from surface vertices. This approach introduces problematic dependencies between internal skeleton and outer soft tissue, limiting direct control over body height and bone lengths. To address these issues, we present ATLAS, a high-fidelity body model learned from 600k high-resolution scans captured using 240 synchronized cameras. Unlike previous methods, we explicitly decouple the shape and skeleton bases by grounding our mesh representation in the human skeleton. This decoupling enables enhanced shape expressivity, fine-grained customization of body attributes, and keypoint fitting independent of external soft-tissue characteristics. ATLAS outperforms existing methods by fitting unseen subjects in diverse poses more accurately, and quantitative evaluations show that our non-linear pose correctives more effectively capture complex poses compared to linear models.

Method: Uses a feature aggregation module integrating local/global scene information from visual and geometric encoders, coupled with a position head for simultaneous 3D asset generation and spatial positioning.

Result: Extensive evaluations confirm efficiency and robust generation abilities, with direct extensibility to multi-image inputs despite single-image training.

Conclusion: SceneGen offers a novel solution for high-quality 3D content generation that could advance practical applications in downstream tasks.

Abstract: 3D content generation has recently attracted significant research interest due to its applications in VR/AR and embodied AI. In this work, we address the challenging task of synthesizing multiple 3D assets within a single scene image. Concretely, our contributions are fourfold: (i) we present SceneGen, a novel framework that takes a scene image and corresponding object masks as input, simultaneously producing multiple 3D assets with geometry and texture. Notably, SceneGen operates with no need for optimization or asset retrieval; (ii) we introduce a novel feature aggregation module that integrates local and global scene information from visual and geometric encoders within the feature extraction module. Coupled with a position head, this enables the generation of 3D assets and their relative spatial positions in a single feedforward pass; (iii) we demonstrate SceneGen’s direct extensibility to multi-image input scenarios. Despite being trained solely on single-image inputs, our architectural design enables improved generation performance with multi-image inputs; and (iv) extensive quantitative and qualitative evaluations confirm the efficiency and robust generation abilities of our approach. We believe this paradigm offers a novel solution for high-quality 3D content generation, potentially advancing its practical applications in downstream tasks. The code and model will be publicly available at: https://mengmouxu.github.io/SceneGen.

Method: Formulates group inference as a quadratic integer assignment problem where candidate outputs are graph nodes, selecting subsets to optimize quality (unary term) and maximize diversity (binary term) with progressive candidate pruning for efficiency.

Result: Extensive experiments show significant improvements in group diversity and quality compared to independent sampling baselines and recent inference algorithms across various tasks.

Conclusion: The framework enables generative models to treat multiple outputs as cohesive groups rather than independent samples, generalizing across text-to-image, image-to-image, image prompting, and video generation tasks.

Abstract: Generative models typically sample outputs independently, and recent inference-time guidance and scaling algorithms focus on improving the quality of individual samples. However, in real-world applications, users are often presented with a set of multiple images (e.g., 4-8) for each prompt, where independent sampling tends to lead to redundant results, limiting user choices and hindering idea exploration. In this work, we introduce a scalable group inference method that improves both the diversity and quality of a group of samples. We formulate group inference as a quadratic integer assignment problem: candidate outputs are modeled as graph nodes, and a subset is selected to optimize sample quality (unary term) while maximizing group diversity (binary term). To substantially improve runtime efficiency, we progressively prune the candidate set using intermediate predictions, allowing our method to scale up to large candidate sets. Extensive experiments show that our method significantly improves group diversity and quality compared to independent sampling baselines and recent inference algorithms. Our framework generalizes across a wide range of tasks, including text-to-image, image-to-image, image prompting, and video generation, enabling generative models to treat multiple outputs as cohesive groups rather than independent samples.

Method: Proposes CineScale, a tuning-free inference paradigm with dedicated variants for different video generation architectures (T2I, T2V, I2V, V2V), built on state-of-the-art open-source frameworks to handle high-frequency information issues.

Result: Extensive experiments show superior performance in higher-resolution visual generation, enabling 8k image generation without fine-tuning and 4k video generation with minimal LoRA fine-tuning.

Conclusion: CineScale successfully extends the capabilities of pre-trained models for high-resolution visual generation across multiple modalities, overcoming the limitations of previous methods that were prone to repetitive patterns and quality degradation.

Abstract: Visual diffusion models achieve remarkable progress, yet they are typically trained at limited resolutions due to the lack of high-resolution data and constrained computation resources, hampering their ability to generate high-fidelity images or videos at higher resolutions. Recent efforts have explored tuning-free strategies to exhibit the untapped potential higher-resolution visual generation of pre-trained models. However, these methods are still prone to producing low-quality visual content with repetitive patterns. The key obstacle lies in the inevitable increase in high-frequency information when the model generates visual content exceeding its training resolution, leading to undesirable repetitive patterns deriving from the accumulated errors. In this work, we propose CineScale, a novel inference paradigm to enable higher-resolution visual generation. To tackle the various issues introduced by the two types of video generation architectures, we propose dedicated variants tailored to each. Unlike existing baseline methods that are confined to high-resolution T2I and T2V generation, CineScale broadens the scope by enabling high-resolution I2V and V2V synthesis, built atop state-of-the-art open-source video generation frameworks. Extensive experiments validate the superiority of our paradigm in extending the capabilities of higher-resolution visual generation for both image and video models. Remarkably, our approach enables 8k image generation without any fine-tuning, and achieves 4k video generation with only minimal LoRA fine-tuning. Generated video samples are available at our website: https://eyeline-labs.github.io/CineScale/.

Method: Projects both data types into RoadNet Sequence integer series, uses non-autoregressive sequence-to-sequence Transformer, introduces Topology-Inherited Training, and leverages SD-Maps prior information from open-source datasets.

Result: Extensive experiments on nuScenes dataset show superiority over state-of-the-art alternatives in both efficiency and accuracy for road network extraction.

Conclusion: The RoadNet Sequence representation and non-autoregressive approach successfully bridge the gap between Euclidean and non-Euclidean data domains, achieving better performance in landmark detection and topology reasoning for HD map generation.

Abstract: The extraction of road network is essential for the generation of high-definition maps since it enables the precise localization of road landmarks and their interconnections. However, generating road network poses a significant challenge due to the conflicting underlying combination of Euclidean (e.g., road landmarks location) and non-Euclidean (e.g., road topological connectivity) structures. Existing methods struggle to merge the two types of data domains effectively, but few of them address it properly. Instead, our work establishes a unified representation of both types of data domain by projecting both Euclidean and non-Euclidean data into an integer series called RoadNet Sequence. Further than modeling an auto-regressive sequence-to-sequence Transformer model to understand RoadNet Sequence, we decouple the dependency of RoadNet Sequence into a mixture of auto-regressive and non-autoregressive dependency. Building on this, our proposed non-autoregressive sequence-to-sequence approach leverages non-autoregressive dependencies while fixing the gap towards auto-regressive dependencies, resulting in success in both efficiency and accuracy. We further identify two main bottlenecks in the current RoadNetTransformer on a non-overfitting split of the dataset: poor landmark detection limited by the BEV Encoder and error propagation to topology reasoning. Therefore, we propose Topology-Inherited Training to inherit better topology knowledge into RoadNetTransformer. Additionally, we collect SD-Maps from open-source map datasets and use this prior information to significantly improve landmark detection and reachability. Extensive experiments on the nuScenes dataset demonstrate the superiority of RoadNet Sequence representation and the non-autoregressive approach compared to existing state-of-the-art alternatives.

Method: Proposes an architecture with a multiview inlier/outlier classification module that respects model equivariance, combined with a robust bundle adjustment step to handle outlier-contaminated point tracks.

Result: The method successfully handles realistic settings with large image collections and point tracks containing many outliers, achieving state-of-the-art accuracies superior to existing deep methods and on-par with leading classical sequential and global methods.

Conclusion: The proposed robust deep learning approach effectively addresses the outlier problem in multiview structure from motion, making deep methods practical for real-world applications with noisy input data.

Abstract: Multiview Structure from Motion is a fundamental and challenging computer vision problem. A recent deep-based approach utilized matrix equivariant architectures for simultaneous recovery of camera pose and 3D scene structure from large image collections. That work, however, made the unrealistic assumption that the point tracks given as input are almost clean of outliers. Here, we propose an architecture suited to dealing with outliers by adding a multiview inlier/outlier classification module that respects the model equivariance and by utilizing a robust bundle adjustment step. Experiments demonstrate that our method can be applied successfully in realistic settings that include large image collections and point tracks extracted with common heuristics that include many outliers, achieving state-of-the-art accuracies in almost all runs, superior to existing deep-based methods and on-par with leading classical (non-deep) sequential and global methods.

Method: Novel transformer-based architecture with attention mechanisms, versatile training procedure that enables operation across different N-way K-shot configurations and prompt types with a single model.

Result: Achieves state-of-the-art performance on COCO-20^i benchmark among multi-way few-shot segmentation methods, and significantly outperforms leading single-class models in multi-class settings.

Conclusion: Label Anything provides a highly flexible and generalizable framework that reduces annotation burden while maintaining high accuracy through diverse visual prompts and transformer architecture.

Abstract: Few-shot semantic segmentation aims to segment objects from previously unseen classes using only a limited number of labeled examples. In this paper, we introduce Label Anything, a novel transformer-based architecture designed for multi-prompt, multi-way few-shot semantic segmentation. Our approach leverages diverse visual prompts – points, bounding boxes, and masks – to create a highly flexible and generalizable framework that significantly reduces annotation burden while maintaining high accuracy. Label Anything makes three key contributions: ($\textit{i}$) we introduce a new task formulation that relaxes conventional few-shot segmentation constraints by supporting various types of prompts, multi-class classification, and enabling multiple prompts within a single image; ($\textit{ii}$) we propose a novel architecture based on transformers and attention mechanisms; and ($\textit{iii}$) we design a versatile training procedure allowing our model to operate seamlessly across different $N$-way $K$-shot and prompt-type configurations with a single trained model. Our extensive experimental evaluation on the widely used COCO-$20^i$ benchmark demonstrates that Label Anything achieves state-of-the-art performance among existing multi-way few-shot segmentation methods, while significantly outperforming leading single-class models when evaluated in multi-class settings. Code and trained models are available at https://github.com/pasqualedem/LabelAnything.

Method: Proposes adaptive computation framework that dynamically exits Transformer blocks based on task complexity, and enforces feature invariance to simulated motion blur for robust representations.

Result: Extensive experiments on four tracking benchmarks validate effectiveness and versatility, demonstrating practical real-time UAV tracking performance.

Conclusion: BDTrack provides an efficient and effective solution for real-time UAV tracking with adaptive computation and motion blur robustness, showing strong potential for practical applications.

Abstract: Unmanned aerial vehicle (UAV) tracking is critical for applications like surveillance, search-and-rescue, and autonomous navigation. However, the high-speed movement of UAVs and targets introduces unique challenges, including real-time processing demands and severe motion blur, which degrade the performance of existing generic trackers. While single-stream vision transformer (ViT) architectures have shown promise in visual tracking, their computational inefficiency and lack of UAV-specific optimizations limit their practicality in this domain. In this paper, we boost the efficiency of this framework by tailoring it into an adaptive computation framework that dynamically exits Transformer blocks for real-time UAV tracking. The motivation behind this is that tracking tasks with fewer challenges can be adequately addressed using low-level feature representations. Simpler tasks can often be handled with less demanding, lower-level features. This approach allows the model use computational resources more efficiently by focusing on complex tasks and conserving resources for easier ones. Another significant enhancement introduced in this paper is the improved effectiveness of ViTs in handling motion blur, a common issue in UAV tracking caused by the fast movements of either the UAV, the tracked objects, or both. This is achieved by acquiring motion blur robust representations through enforcing invariance in the feature representation of the target with respect to simulated motion blur. We refer to our proposed approach as BDTrack. Extensive experiments conducted on four tracking benchmarks validate the effectiveness and versatility of our approach, demonstrating its potential as a practical and effective approach for real-time UAV tracking. Code is released at: https://github.com/wuyou3474/BDTrack.

Method: FPBA (Frequency-based Post-train Bayesian Attack) adds perturbations in the frequency domain to push images away from their original frequency distribution, and uses a novel post-train Bayesian strategy to simulate diverse victim models from a single surrogate without re-training.

Result: FPBA demonstrates successful black-box attacks across different models (CNNs and ViTs), generators, defense methods, and can even evade cross-generator detection, showing that adversarial attacks pose a real threat to AIGI detectors.

Conclusion: Adversarial attacks are a significant threat to AIGI detectors, and the proposed FPBA method effectively compromises these detectors across various scenarios, highlighting the need for more robust detection systems.

Abstract: Recent advancements in image synthesis, particularly with the advent of GAN and Diffusion models, have amplified public concerns regarding the dissemination of disinformation. To address such concerns, numerous AI-generated Image (AIGI) Detectors have been proposed and achieved promising performance in identifying fake images. However, there still lacks a systematic understanding of the adversarial robustness of AIGI detectors. In this paper, we examine the vulnerability of state-of-the-art AIGI detectors against adversarial attack under white-box and black-box settings, which has been rarely investigated so far. To this end, we propose a new method to attack AIGI detectors. First, inspired by the obvious difference between real images and fake images in the frequency domain, we add perturbations under the frequency domain to push the image away from its original frequency distribution. Second, we explore the full posterior distribution of the surrogate model to further narrow this gap between heterogeneous AIGI detectors, e.g. transferring adversarial examples across CNNs and ViTs. This is achieved by introducing a novel post-train Bayesian strategy that turns a single surrogate into a Bayesian one, capable of simulating diverse victim models using one pre-trained surrogate, without the need for re-training. We name our method as Frequency-based Post-train Bayesian Attack, or FPBA. Through FPBA, we show that adversarial attack is truly a real threat to AIGI detectors, because FPBA can deliver successful black-box attacks across models, generators, defense methods, and even evade cross-generator detection, which is a crucial real-world detection scenario. The code will be shared upon acceptance.

Method: Combines shape, Mahalanobis distance, and novel soft BIoU similarity for richer similarity measure. Introduces soft detection confidence boost technique and varying similarity thresholds to handle infrequently updated tracklets.

Result: Achieves near state-of-the-art results on MOT17 dataset and new state-of-the-art HOTA and IDF1 scores on MOT20 dataset.

Conclusion: The proposed improvements to BoostTrack’s confidence boosting method effectively enhance MOT performance through better similarity measures and confidence scoring, with components that can be integrated into any MOT algorithm.

Abstract: Multiple object tracking (MOT) depends heavily on selection of true positive detected bounding boxes. However, this aspect of the problem is mostly overlooked or mitigated by employing two-stage association and utilizing low confidence detections in the second stage. Recently proposed BoostTrack attempts to avoid the drawbacks of multiple stage association approach and use low-confidence detections by applying detection confidence boosting. In this paper, we identify the limitations of the confidence boost used in BoostTrack and propose a method to improve its performance. To construct a richer similarity measure and enable a better selection of true positive detections, we propose to use a combination of shape, Mahalanobis distance and novel soft BIoU similarity. We propose a soft detection confidence boost technique which calculates new confidence scores based on the similarity measure and the previous confidence scores, and we introduce varying similarity threshold to account for lower similarity measure between detections and tracklets which are not regularly updated. The proposed additions are mutually independent and can be used in any MOT algorithm. Combined with the BoostTrack+ baseline, our method achieves near state of the art results on the MOT17 dataset and new state of the art HOTA and IDF1 scores on the MOT20 dataset. The source code is available at: https://github.com/vukasin-stanojevic/BoostTrack .

Method: TripleMixer integrates three specialized mixer modules for spatial, frequency, and channel-wise processing to suppress high-frequency noise while preserving geometric structures. The method is plug-and-play compatible with existing pipelines.

Result: TripleMixer achieves state-of-the-art denoising performance and yields substantial improvements across semantic segmentation, place recognition, and object detection tasks without requiring retraining of downstream models.

Conclusion: Denoising serves as an effective task-agnostic preprocessing strategy that enhances LiDAR robustness in real-world autonomous driving applications under adverse weather conditions.

Abstract: Adverse weather conditions such as snow, fog, and rain pose significant challenges to LiDAR-based perception models by introducing noise and corrupting point cloud measurements. To address this issue, we propose TripleMixer, a robust and efficient point cloud denoising network that integrates spatial, frequency, and channel-wise processing through three specialized mixer modules. TripleMixer effectively suppresses high-frequency noise while preserving essential geometric structures and can be seamlessly deployed as a plug-and-play module within existing LiDAR perception pipelines. To support the development and evaluation of denoising methods, we construct two large-scale simulated datasets, Weather-KITTI and Weather-NuScenes, covering diverse weather scenarios with dense point-wise semantic and noise annotations. Based on these datasets, we establish four benchmarks: Denoising, Semantic Segmentation (SS), Place Recognition (PR), and Object Detection (OD). These benchmarks enable systematic evaluation of denoising generalization, transferability, and downstream impact under both simulated and real-world adverse weather conditions. Extensive experiments demonstrate that TripleMixer achieves state-of-the-art denoising performance and yields substantial improvements across all downstream tasks without requiring retraining. Our results highlight the potential of denoising as a task-agnostic preprocessing strategy to enhance LiDAR robustness in real-world autonomous driving applications.

Method: Replaces ADAM with Levenberg-Marquardt optimizer, proposes caching data structure for intermediate gradients, uses custom CUDA kernels for Jacobian-vector products, and combines update directions from multiple image subsets in weighted mean.

Result: 20% faster optimization than original 3DGS while achieving the same reconstruction quality.

Conclusion: The method provides significant speed improvement without quality loss and is compatible with other 3DGS acceleration techniques, enabling even faster performance compared to vanilla 3DGS.

Abstract: We present 3DGS-LM, a new method that accelerates the reconstruction of 3D Gaussian Splatting (3DGS) by replacing its ADAM optimizer with a tailored Levenberg-Marquardt (LM). Existing methods reduce the optimization time by decreasing the number of Gaussians or by improving the implementation of the differentiable rasterizer. However, they still rely on the ADAM optimizer to fit Gaussian parameters of a scene in thousands of iterations, which can take up to an hour. To this end, we change the optimizer to LM that runs in conjunction with the 3DGS differentiable rasterizer. For efficient GPU parallization, we propose a caching data structure for intermediate gradients that allows us to efficiently calculate Jacobian-vector products in custom CUDA kernels. In every LM iteration, we calculate update directions from multiple image subsets using these kernels and combine them in a weighted mean. Overall, our method is 20% faster than the original 3DGS while obtaining the same reconstruction quality. Our optimization is also agnostic to other methods that acclerate 3DGS, thus enabling even faster speedups compared to vanilla 3DGS.

Method: Incorporates (1) an additional temporal stream to encode frame relationships, (2) discrete temporal tokens with time knowledge for timestamp representation, and uses multi-stage training from simple video-captioning to complex temporal grounding tasks. Also creates a grounded VideoQA dataset via automatic annotation.

Result: Extensive experiments show Grounded-VideoLLM excels in fine-grained grounding tasks including temporal sentence grounding, dense video captioning, and grounded VideoQA, while also demonstrating strong general video understanding capabilities.

Conclusion: Grounded-VideoLLM successfully addresses the fine-grained temporal grounding limitations of current Video-LLMs and shows great potential as a versatile video assistant for both specific temporal reasoning and general video understanding tasks.

Abstract: Video Large Language Models (Video-LLMs) have demonstrated remarkable capabilities in coarse-grained video understanding, however, they struggle with fine-grained temporal grounding. In this paper, we introduce Grounded-VideoLLM, a novel Video-LLM adept at perceiving and reasoning over specific video moments in a fine-grained manner. We identify that current Video-LLMs have limitations for fine-grained video understanding since they lack effective temporal modeling and timestamp representation. In light of this, we sharpen our model by incorporating (1) an additional temporal stream to encode the relationships between frames and (2) discrete temporal tokens enriched with specific time knowledge to represent timestamps. To optimize the training of Grounded-VideoLLM, we employ a multi-stage training scheme, beginning with simple video-captioning tasks and progressively introducing video temporal grounding tasks of increasing complexity. To further enhance Grounded-VideoLLM’s temporal reasoning capability, we also curate a grounded VideoQA dataset by an automatic annotation pipeline. Extensive experiments demonstrate that Grounded-VideoLLM not only excels in fine-grained grounding tasks such as temporal sentence grounding, dense video captioning, and grounded VideoQA, but also shows great potential as a versatile video assistant for general video understanding.

Method: RouteT2I establishes multi-dimensional quality metrics by evaluating image similarity to positive/negative texts describing each metric. It identifies key prompt tokens and uses Pareto relative superiority to compare multi-metric quality, then allocates prompts to edge or cloud based on cost constraints.

Result: RouteT2I significantly reduces requests to large cloud models while maintaining high-quality image generation.

Conclusion: The proposed routing framework effectively balances cost and performance by intelligently distributing prompts between edge and cloud models based on quality predictions.

Abstract: Large text-to-image models demonstrate impressive generation capabilities; however, their substantial size necessitates expensive cloud servers for deployment. Conversely, light-weight models can be deployed on edge devices at lower cost but often with inferior generation quality for complex user prompts. To strike a balance between performance and cost, we propose a routing framework, called RouteT2I, which dynamically selects either the large cloud model or the light-weight edge model for each user prompt. Since generated image quality is challenging to measure and compare directly, RouteT2I establishes multi-dimensional quality metrics, particularly, by evaluating the similarity between the generated images and both positive and negative texts that describe each specific quality metric. RouteT2I then predicts the expected quality of the generated images by identifying key tokens in the prompt and comparing their impact on the quality. RouteT2I further introduces the Pareto relative superiority to compare the multi-metric quality of the generated images. Based on this comparison and predefined cost constraints, RouteT2I allocates prompts to either the edge or the cloud. Evaluation reveals that RouteT2I significantly reduces the number of requesting large cloud model while maintaining high-quality image generation.

Method: Proposes Evaluation Agent framework that employs human-like strategies for efficient, dynamic, multi-round evaluations using only a few samples per round, offering promptable evaluation tailored to user needs with detailed explanations.

Result: Experiments show Evaluation Agent reduces evaluation time to 10% of traditional methods while delivering comparable results, with advantages in efficiency, promptable evaluation, explainability, and scalability.

Conclusion: The Evaluation Agent framework provides an efficient and explainable alternative to traditional evaluation methods for visual generative models, and is fully open-sourced to advance research in this field.

Abstract: Recent advancements in visual generative models have enabled high-quality image and video generation, opening diverse applications. However, evaluating these models often demands sampling hundreds or thousands of images or videos, making the process computationally expensive, especially for diffusion-based models with inherently slow sampling. Moreover, existing evaluation methods rely on rigid pipelines that overlook specific user needs and provide numerical results without clear explanations. In contrast, humans can quickly form impressions of a model’s capabilities by observing only a few samples. To mimic this, we propose the Evaluation Agent framework, which employs human-like strategies for efficient, dynamic, multi-round evaluations using only a few samples per round, while offering detailed, user-tailored analyses. It offers four key advantages: 1) efficiency, 2) promptable evaluation tailored to diverse user needs, 3) explainability beyond single numerical scores, and 4) scalability across various models and tools. Experiments show that Evaluation Agent reduces evaluation time to 10% of traditional methods while delivering comparable results. The Evaluation Agent framework is fully open-sourced to advance research in visual generative models and their efficient evaluation.

Method: Proposes Hadamard Attention Recurrent Stereo Transformer (HART) with two key components: 1) Dense Attention Kernel (DAK) that maps attention weights to high-dimensional space without upper bound constraints, and 2) Multi Kernel & Order Interaction (MKOI) module that unifies semantic and spatial knowledge learning.

Result: HART ranked 1st on the KITTI 2012 benchmark among all published methods at submission time, particularly excelling in reflective areas.

Conclusion: The proposed HART framework effectively addresses the low-rank bottleneck problem in stereo matching transformers and demonstrates superior performance in challenging reflective conditions.

Abstract: Constrained by the low-rank bottleneck inherent in attention mechanisms, current stereo matching transformers suffer from limited nonlinear expressivity, which renders their feature representations sensitive to challenging conditions such as reflections. To overcome this difficulty, we present the Hadamard Attention Recurrent Stereo Transformer (HART). HART includes a novel attention mechanism that incorporates the following components: 1) The Dense Attention Kernel (DAK) maps the attention weight distribution into a high-dimensional space over (0, +$\infty$). By removing the upper bound constraint on attention weights, DAK enables more flexible modeling of complex feature interactions. This reduces feature collinearity. 2) The Multi Kernel & Order Interaction (MKOI) module extends the attention mechanism by unifying semantic and spatial knowledge learning. This integration improves the ability of HART to learn features in binocular images. Experimental results demonstrate the effectiveness of our HART. In reflective area, HART ranked 1st on the KITTI 2012 benchmark among all published methods at the time of submission. Code is available at https://github.com/ZYangChen/HART.

Method: Explored various deep learning models including traditional techniques and newer architectures, trained multiple models on SP Cup 2025 deep fake detection dataset.

Result: Models were rigorously assessed using performance metrics such as accuracy, though specific accuracy values are not provided in the abstract.

Conclusion: The study demonstrates the application of deep learning approaches for deep fake detection, contributing to the development of more reliable authentication systems for digital media.

Abstract: The proliferation of deep fake technology poses significant challenges to digital media authenticity, necessitating robust detection mechanisms. This project evaluates deep fake detection using the SP Cup’s 2025 deep fake detection challenge dataset. We focused on exploring various deep learning models for detecting deep fake content, utilizing traditional deep learning techniques alongside newer architectures. Our approach involved training a series of models and rigorously assessing their performance using metrics such as accuracy.

Method: Uses a vision-language model backbone to deeply integrate image features with natural language instructions, creating strongly unified vision-language representations.

Result: Achieves best-for-size performance on MSCOCO image-to-text retrieval, top performance on classification and VQA tasks in MMEB, and excels on the CtrlBench benchmark for instruction-based retrieval.

Conclusion: ABC advances visual embeddings by providing high-quality representations with natural language control, solving subtle and ambiguous visual retrieval problems through deep modality integration.

Abstract: Visual embedding models excel at zero-shot tasks like visual retrieval and classification. However, these models cannot be used for tasks that contain ambiguity or require user instruction. These tasks necessitate an embedding model which outputs can use a natural language instruction to control the representation of a visual embedding. Existing CLIP-based approaches embed images and text independently, and fuse the result. We find that this results in weak interactions between modalities, and poor user control over the representation. We introduce ABC, an open-source multimodal embedding model that uses a vision-language model backbone to deeply integrate image features with natural language instructions. ABC achieves best-for-size performance on MSCOCO image-to-text retrieval and is the top performing model on classification and VQA tasks in the Massive Multimodal Embedding Benchmark. With a strongly unified vision-language representation, ABC can use natural language to solve subtle and potentially ambiguous visual retrieval problems. To evaluate this capability, we design CtrlBench, a benchmark that requires interleaving textual instructions with image content for correct retrieval. ABC advances the state of visual embeddings, outputting high-quality visual representations with natural language control. Our model and datasets are available at our project page: https://tiger-ai-lab.github.io/ABC/

Method: The paper addresses two key dimensions: 1) Revealing fundamental flaws in evaluation benchmarks where up to 13% of objects in OoD test sets actually belong to in-distribution classes, and 2) Introducing a training-time mitigation paradigm that fine-tunes detectors using carefully synthesized OoD datasets that semantically resemble in-distribution objects, shaping defensive decision boundaries by suppressing objectness on OoD objects.

Result: The approach achieves a 91% reduction in hallucination error of a YOLO model on BDD-100K. The methodology generalizes across detection paradigms including YOLO, Faster R-CNN, and RT-DETR, and supports few-shot adaptation.

Conclusion: The contributions offer a principled and effective way to reduce OoD-induced hallucination in object detectors by addressing both evaluation benchmark quality issues and providing a novel training-time mitigation approach that operates independently of external OoD detectors.

Abstract: Out-of-distribution (OoD) inputs pose a persistent challenge to deep learning models, often triggering overconfident predictions on non-target objects. While prior work has primarily focused on refining scoring functions and adjusting test-time thresholds, such algorithmic improvements offer only incremental gains. We argue that a rethinking of the entire development lifecycle is needed to mitigate these risks effectively. This work addresses two overlooked dimensions of OoD detection in object detection. First, we reveal fundamental flaws in widely used evaluation benchmarks: contrary to their design intent, up to 13% of objects in the OoD test sets actually belong to in-distribution classes, and vice versa. These quality issues severely distort the reported performance of existing methods and contribute to their high false positive rates. Second, we introduce a novel training-time mitigation paradigm that operates independently of external OoD detectors. Instead of relying solely on post-hoc scoring, we fine-tune the detector using a carefully synthesized OoD dataset that semantically resembles in-distribution objects. This process shapes a defensive decision boundary by suppressing objectness on OoD objects, leading to a 91% reduction in hallucination error of a YOLO model on BDD-100K. Our methodology generalizes across detection paradigms such as YOLO, Faster R-CNN, and RT-DETR, and supports few-shot adaptation. Together, these contributions offer a principled and effective way to reduce OoD-induced hallucination in object detectors. Code and data are available at: https://gricad-gitlab.univ-grenoble-alpes.fr/dnn-safety/m-hood.

Method: BannerAgency - MLLM agent system that collaborates with advertisers, generates background images, creates design blueprints, and renders final creatives as editable components

Result: Introduced BannerRequest400 benchmark with 400 diverse requests; quantitative and qualitative evaluations show effective banner generation, adaptability, and strong editability

Conclusion: Component-based approach enables fully automated banner design with high editability, streamlining production across diverse marketing contexts with minimal manual effort

Abstract: Advertising banners are critical for capturing user attention and enhancing advertising campaign effectiveness. Creating aesthetically pleasing banner designs while conveying the campaign messages is challenging due to the large search space involving multiple design elements. Additionally, advertisers need multiple sizes for different displays and various versions to target different sectors of audiences. Since design is intrinsically an iterative and subjective process, flexible editability is also in high demand for practical usage. While current models have served as assistants to human designers in various design tasks, they typically handle only segments of the creative design process or produce pixel-based outputs that limit editability. This paper introduces a training-free framework for fully automated banner ad design creation, enabling frontier multimodal large language models (MLLMs) to streamline the production of effective banners with minimal manual effort across diverse marketing contexts. We present BannerAgency, an MLLM agent system that collaborates with advertisers to understand their brand identity and banner objectives, generates matching background images, creates blueprints for foreground design elements, and renders the final creatives as editable components in Figma or SVG formats rather than static pixels. To facilitate evaluation and future research, we introduce BannerRequest400, a benchmark featuring 100 unique logos paired with 400 diverse banner requests. Through quantitative and qualitative evaluations, we demonstrate the framework’s effectiveness, emphasizing the quality of the generated banner designs, their adaptability to various banner requests, and their strong editability enabled by this component-based approach.

Method: Semi-automatic construction of procedural knowledge graph (PKG) from COIN dataset, ConceptNet, and LLM outputs; graph traversal templates for QA generation; neurosymbolic Knowledge Module Learning (KML) approach combining neural modules with LLM-based structured reasoning.

Result: The approach improves reasoning performance on the dataset and enables step-by-step reasoning traces for interpretability, with theoretical analysis showing trained models satisfy near optimal conditions for learning KG relations.

Conclusion: PKR-QA provides a valuable benchmark for procedural reasoning, and the KML approach effectively combines neural learning with symbolic reasoning for interpretable procedural knowledge question answering.

Abstract: We introduce \dataset (Procedural Knowledge Reasoning Question Answering), a new benchmark for question answering over procedural tasks that require structured reasoning. PKR-QA is constructed semi-automatically using a procedural knowledge graph (PKG), which encodes task-specific knowledge across diverse domains. The PKG is built by curating and linking information from the COIN instructional video dataset and the ontology, enriched with commonsense knowledge from ConceptNet and structured outputs from Large Language Models (LLMs), followed by manual verification. To generate question-answer pairs, we design graph traversal templates where each template is applied systematically over PKG. To enable interpretable reasoning, we propose a neurosymbolic approach called Knowledge Module Learning (KML), which learns procedural relations via neural modules and composes them for structured reasoning with LLMs. Experiments demonstrate that this paradigm improves reasoning performance on our dataset and enables step-by-step reasoning traces that facilitate interpretability. Our theoretical analysis on KML learning shows that our trained models satisfy near optimal conditions for learning KG relations as neural network mapping models. Code and dataset will be released soon.

Method: Created TrackID3x3 dataset with three subsets (Indoor fixed-camera, Outdoor fixed-camera, Drone camera) and introduced Track-ID task for fixed-camera scenarios. Proposed baseline Track-ID algorithm and benchmarked with MOT algorithms (BoT-SORT-ReID) and pose estimation methods (HRNet, RTMPose, SwinPose).

Result: Demonstrated robust performance results and identified remaining challenges in 3x3 basketball tracking and pose estimation through comprehensive benchmark experiments.

Conclusion: The dataset and evaluation benchmarks provide a solid foundation for advancing automated analytics in 3x3 basketball, addressing gaps in current sports analytics research for non-mainstream sports and fixed-camera scenarios.

Abstract: Multi-object tracking, player identification, and pose estimation are fundamental components of sports analytics, essential for analyzing player movements, performance, and tactical strategies. However, existing datasets and methodologies primarily target mainstream team sports such as soccer and conventional 5-on-5 basketball, often overlooking scenarios involving fixed-camera setups commonly used at amateur levels, less mainstream sports, or datasets that explicitly incorporate pose annotations. In this paper, we propose the TrackID3x3 dataset, the first publicly available comprehensive dataset specifically designed for multi-player tracking, player identification, and pose estimation in 3x3 basketball scenarios. The dataset comprises three distinct subsets (Indoor fixed-camera, Outdoor fixed-camera, and Drone camera footage), capturing diverse full-court camera perspectives and environments. We also introduce the Track-ID task, a simplified variant of the game state reconstruction task that excludes field detection and focuses exclusively on fixed-camera scenarios. To evaluate performance, we propose a baseline algorithm called Track-ID algorithm, tailored to assess tracking and identification quality. Furthermore, our benchmark experiments, utilizing recent multi-object tracking algorithms (e.g., BoT-SORT-ReID) and top-down pose estimation methods (HRNet, RTMPose, and SwinPose), demonstrate robust results and highlight remaining challenges. Our dataset and evaluation benchmarks provide a solid foundation for advancing automated analytics in 3x3 basketball. Dataset and code will be available at https://github.com/open-starlab/TrackID3x3.

Method: Proposes tri-modal video-gesture-speech-text representation learning using global phrase contrastive loss and local gesture-word coupling loss. Weakly supervised learning from in-the-wild videos.

Result: Learned representations outperform previous methods including large vision-language models. Speech and text modalities capture distinct gesture-related signals.

Conclusion: Shared tri-modal embedding space provides advantages for gesture understanding. Framework enables comprehensive gesture-speech-text association analysis with state-of-the-art performance.

Abstract: Co-speech gestures play a vital role in non-verbal communication. In this paper, we introduce a new framework for co-speech gesture understanding in the wild. Specifically, we propose three new tasks and benchmarks to evaluate a model’s capability to comprehend gesture-speech-text associations: (i) gesture based retrieval, (ii) gesture word spotting, and (iii) active speaker detection using gestures. We present a new approach that learns a tri-modal video-gesture-speech-text representation to solve these tasks. By leveraging a combination of global phrase contrastive loss and local gesture-word coupling loss, we demonstrate that a strong gesture representation can be learned in a weakly supervised manner from videos in the wild. Our learned representations outperform previous methods, including large vision-language models (VLMs). Further analysis reveals that speech and text modalities capture distinct gesture related signals, underscoring the advantages of learning a shared tri-modal embedding space. The dataset, model, and code are available at: https://www.robots.ox.ac.uk/~vgg/research/jegal.

Method: Uses three parallel modules: Diffusion Prior Depth Estimator for depth, Semantic Aware Segmentation Network for semantics, and Multi-View Interaction Network for cross-view features. These are integrated through a Text-Guided Semantic Fusion Module with attention-based feature aggregation.

Result: Experimental results on benchmark datasets show improved performance across multiple evaluation metrics compared to existing methods.

Conclusion: The framework successfully integrates text guidance with 3D Gaussian Splatting, producing high-fidelity reconstructions with enhanced semantic and geometric alignment.

Abstract: Recent advancements in Generalizable Gaussian Splatting have enabled robust 3D reconstruction from sparse input views by utilizing feed-forward Gaussian Splatting models, achieving superior cross-scene generalization. However, while many methods focus on geometric consistency, they often neglect the potential of text-driven guidance to enhance semantic understanding, which is crucial for accurately reconstructing fine-grained details in complex scenes. To address this limitation, we propose TextSplat–the first text-driven Generalizable Gaussian Splatting framework. By employing a text-guided fusion of diverse semantic cues, our framework learns robust cross-modal feature representations that improve the alignment of geometric and semantic information, producing high-fidelity 3D reconstructions. Specifically, our framework employs three parallel modules to obtain complementary representations: the Diffusion Prior Depth Estimator for accurate depth information, the Semantic Aware Segmentation Network for detailed semantic information, and the Multi-View Interaction Network for refined cross-view features. Then, in the Text-Guided Semantic Fusion Module, these representations are integrated via the text-guided and attention-based feature aggregation mechanism, resulting in enhanced 3D Gaussian parameters enriched with detailed semantic cues. Experimental results on various benchmark datasets demonstrate improved performance compared to existing methods across multiple evaluation metrics, validating the effectiveness of our framework. The code will be publicly available.

Method: Constructed Any2Omni dataset with 60,000+ training data covering diverse input conditions and 9 ODI tasks. Proposed Omni² model that handles various ODI generation and editing tasks using one unified model.

Result: Extensive experiments demonstrate superiority and effectiveness of Omni² model for both ODI generation and editing tasks.

Conclusion: The proposed solution bridges the gap in ODI synthesis, providing comprehensive dataset and effective model for 360° image generation and editing in VR/AR applications.

Abstract: $360^{\circ}$ omnidirectional images (ODIs) have gained considerable attention recently, and are widely used in various virtual reality (VR) and augmented reality (AR) applications. However, capturing such images is expensive and requires specialized equipment, making ODI synthesis increasingly important. While common 2D image generation and editing methods are rapidly advancing, these models struggle to deliver satisfactory results when generating or editing ODIs due to the unique format and broad 360$^{\circ}$ Field-of-View (FoV) of ODIs. To bridge this gap, we construct \textbf{\textit{Any2Omni}}, the first comprehensive ODI generation-editing dataset comprises 60,000+ training data covering diverse input conditions and up to 9 ODI generation and editing tasks. Built upon Any2Omni, we propose an \textbf{\underline{Omni}} model for \textbf{\underline{Omni}}-directional image generation and editing (\textbf{\textit{Omni$^2$}}), with the capability of handling various ODI generation and editing tasks under diverse input conditions using one model. Extensive experiments demonstrate the superiority and effectiveness of the proposed Omni$^2$ model for both the ODI generation and editing tasks. Both the Any2Omni dataset and the Omni$^2$ model are publicly available at: https://github.com/IntMeGroup/Omni2.

Method: Designs method solely based on first-order optimizers, identifies and eliminates two key performance bottlenecks: computational complexity and kernel implementation of each optimization step.

Result: FastMap is up to 10 times faster than COLMAP and GLOMAP with GPU acceleration while achieving comparable pose accuracy.

Conclusion: First-order optimization approach provides significant speed improvements for structure from motion while maintaining accuracy, making it suitable for large-scale applications.

Abstract: We propose FastMap, a new global structure from motion method focused on speed and simplicity. Previous methods like COLMAP and GLOMAP are able to estimate high-precision camera poses, but suffer from poor scalability when the number of matched keypoint pairs becomes large, mainly due to the time-consuming process of second-order Gauss-Newton optimization. Instead, we design our method solely based on first-order optimizers. To obtain maximal speedup, we identify and eliminate two key performance bottlenecks: computational complexity and the kernel implementation of each optimization step. Through extensive experiments, we show that FastMap is up to 10 times faster than COLMAP and GLOMAP with GPU acceleration and achieves comparable pose accuracy.

Method: Used Apple Vision Pro to collect egocentric video with paired 3D hand and finger tracking data at recording time, leveraging multiple calibrated cameras and on-device SLAM for precise joint tracking across 194 different tabletop manipulation tasks.

Result: Created EgoDex with 829 hours of video covering diverse manipulation behaviors with household objects, trained imitation learning policies for hand trajectory prediction, and established metrics and benchmarks for evaluation.

Conclusion: EgoDex pushes the frontiers of robotics, computer vision, and foundation models by providing a publicly available large-scale dataset to advance research in dexterous manipulation and imitation learning.

Abstract: Imitation learning for manipulation has a well-known data scarcity problem. Unlike natural language and 2D computer vision, there is no Internet-scale corpus of data for dexterous manipulation. One appealing option is egocentric human video, a passively scalable data source. However, existing large-scale datasets such as Ego4D do not have native hand pose annotations and do not focus on object manipulation. To this end, we use Apple Vision Pro to collect EgoDex: the largest and most diverse dataset of dexterous human manipulation to date. EgoDex has 829 hours of egocentric video with paired 3D hand and finger tracking data collected at the time of recording, where multiple calibrated cameras and on-device SLAM can be used to precisely track the pose of every joint of each hand. The dataset covers a wide range of diverse manipulation behaviors with everyday household objects in 194 different tabletop tasks ranging from tying shoelaces to folding laundry. Furthermore, we train and systematically evaluate imitation learning policies for hand trajectory prediction on the dataset, introducing metrics and benchmarks for measuring progress in this increasingly important area. By releasing this large-scale dataset, we hope to push the frontier of robotics, computer vision, and foundation models. EgoDex is publicly available for download at https://github.com/apple/ml-egodex.

Method: Uses visual encoders (ResNet/ViT) to extract attributes from tool images and fine-tuned language models (GPT-2/LLaMA/DeepSeek) to derive required attributes from task descriptions, with a comprehensive dataset (ToolNet) of 115 tools and 13 attributes.

Result: Achieves 74% accuracy in tool selection tasks, significantly outperforming direct tool matching (20%) and smaller multimodal models (21%-58%), while approaching GPT-4o performance (73%) with fewer parameters.

Conclusion: Provides a parameter-efficient, interpretable solution that mimics human-like tool cognition, with manipulation-related attributes proving most critical across modalities.

Abstract: Flexible tool selection reflects a complex cognitive ability that distinguishes humans from other species, yet computational models that capture this ability remain underdeveloped. We developed a framework using low-dimensional attribute representations to bridge visual tool perception and linguistic task understanding. We constructed a comprehensive dataset (ToolNet) containing 115 common tools labeled with 13 carefully designed attributes spanning physical, functional, and psychological properties, paired with natural language scenarios describing tool usage. Visual encoders (ResNet or ViT) extract attributes from tool images while fine-tuned language models (GPT-2, LLaMA, DeepSeek) derive required attributes from task descriptions. Our approach achieves 74% accuracy in tool selection tasks-significantly outperforming direct tool matching (20%) and smaller multimodal models (21%-58%), while approaching performance of much larger models like GPT-4o (73%) with substantially fewer parameters. Human evaluation studies validate our framework’s alignment with human decision-making patterns, and generalization experiments demonstrate effective performance on novel tool categories. Ablation studies revealed that manipulation-related attributes (graspability, elongation, hand-relatedness) consistently prove most critical across modalities. This work provides a parameter-efficient, interpretable solution that mimics human-like tool cognition, advancing both cognitive science understanding and practical applications in tool selection tasks.

Method: Automated deep learning pipeline including layout analysis, table detection, cell classification, and handwriting recognition applied to ~200,000 images of handwritten migration records from Evangelical-Lutheran parishes.

Result: Successfully extracted over six million migration entries, creating a structured dataset suitable for research, with a case study demonstrating reconstruction of local migration histories in Elimäki parish.

Conclusion: Demonstrates that large-scale handwritten archival materials can be effectively transformed into structured datasets using automated deep learning methods, supporting historical and demographic research on migration patterns.

Abstract: This article presents a large-scale effort to create a structured dataset of internal migration in Finland between 1800 and 1920 using digitized church moving records. These records, maintained by Evangelical-Lutheran parishes, document the migration of individuals and families and offer a valuable source for studying historical demographic patterns. The dataset includes over six million entries extracted from approximately 200,000 images of handwritten migration records. The data extraction process was automated using a deep learning pipeline that included layout analysis, table detection, cell classification, and handwriting recognition. The complete pipeline was applied to all images, resulting in a structured dataset suitable for research. The dataset can be used to study internal migration, urbanization, and family migration, and the spread of disease in preindustrial Finland. A case study from the Elim"aki parish shows how local migration histories can be reconstructed. The work demonstrates how large volumes of handwritten archival material can be transformed into structured data to support historical and demographic research.

Method: Proposed CLARE method with dual-stream architecture: textual branch encodes referring expressions, visual branch detects objects and extracts features, uses contrastive language-instance alignment for training.

Result: Created REIRCOCO benchmark from MSCOCO and RefCOCO datasets using VLMs to generate referring expressions, and developed CLARE as an end-to-end baseline solution.

Conclusion: REIR addresses the gap between TIR and REC tasks, providing a comprehensive solution for instance-level retrieval and localization with the proposed benchmark and baseline method.

Abstract: Using natural language to query visual information is a fundamental need in real-world applications. Text-Image Retrieval (TIR) retrieves a target image from a gallery based on an image-level description, while Referring Expression Comprehension (REC) localizes a target object within a given image using an instance-level description. However, real-world applications often present more complex demands. Users typically query an instance-level description across a large gallery and expect to receive both relevant image and the corresponding instance location. In such scenarios, TIR struggles with fine-grained descriptions and object-level localization, while REC is limited in its ability to efficiently search large galleries and lacks an effective ranking mechanism. In this paper, we introduce a new task called \textbf{Referring Expression Instance Retrieval (REIR)}, which supports both instance-level retrieval and localization based on fine-grained referring expressions. First, we propose a large-scale benchmark for REIR, named REIRCOCO, constructed by prompting advanced vision-language models to generate high-quality referring expressions for instances in the MSCOCO and RefCOCO datasets. Second, we present a baseline method, Contrastive Language-Instance Alignment with Relation Experts (CLARE), which employs a dual-stream architecture to address REIR in an end-to-end manner. Given a referring expression, the textual branch encodes it into a query embedding. The visual branch detects candidate objects and extracts their instance-level visual features. The most similar candidate to the query is selected for bounding box prediction. CLARE is first trained on object detection and REC datasets to establish initial grounding capabilities, then optimized via Contrastive Language-Instance Alignment (CLIA) for improved retrieval across images. We will release our code and benchmark publicly.

Method: Uses MLLMs to produce visual clues as input to diffusion decoders, with lightweight architectural additions (vision head and adapter) and efficient multi-stage training for limited data/resources.

Result: The model demonstrates satisfactory generalization across video generation, editing, and understanding tasks.

Conclusion: Omni-Video successfully creates a unified framework for multiple video tasks by effectively connecting MLLMs with diffusion models through visual clue generation.

Abstract: Notable breakthroughs in unified understanding and generation modeling have led to remarkable advancements in image understanding, reasoning, production and editing, yet current foundational models predominantly focus on processing images, creating a gap in the development of unified models for video understanding and generation. This report presents Omni-Video, an efficient and effective unified framework for video understanding, generation, as well as instruction-based editing. Our key insight is to teach existing multimodal large language models (MLLMs) to produce continuous visual clues that are used as the input of diffusion decoders, which produce high-quality videos conditioned on these visual clues. To fully unlock the potential of our system for unified video modeling, we integrate several technical improvements: 1) a lightweight architectural design that respectively attaches a vision head on the top of MLLMs and a adapter before the input of diffusion decoders, the former produce visual tokens for the latter, which adapts these visual tokens to the conditional space of diffusion decoders; and 2) an efficient multi-stage training scheme that facilitates a fast connection between MLLMs and diffusion decoders with limited data and computational resources. We empirically demonstrate that our model exhibits satisfactory generalization abilities across video generation, editing and understanding tasks.

Method: Uses two concept banks (pathology and anatomy), aligns visual features with medical concepts, employs anatomy-based contrastive learning, matching loss for pathological features, and feature gating mechanism to filter low-quality concepts.

Result: Superior performance on MIMIC-CXR and CheXpert Plus benchmarks, demonstrating effectiveness in radiology report generation.

Conclusion: MCA-RG successfully addresses feature-text alignment challenges in radiology report generation through explicit medical concept alignment and knowledge-driven framework.

Abstract: Despite significant advancements in adapting Large Language Models (LLMs) for radiology report generation (RRG), clinical adoption remains challenging due to difficulties in accurately mapping pathological and anatomical features to their corresponding text descriptions. Additionally, semantic agnostic feature extraction further hampers the generation of accurate diagnostic reports. To address these challenges, we introduce Medical Concept Aligned Radiology Report Generation (MCA-RG), a knowledge-driven framework that explicitly aligns visual features with distinct medical concepts to enhance the report generation process. MCA-RG utilizes two curated concept banks: a pathology bank containing lesion-related knowledge, and an anatomy bank with anatomical descriptions. The visual features are aligned with these medical concepts and undergo tailored enhancement. We further propose an anatomy-based contrastive learning procedure to improve the generalization of anatomical features, coupled with a matching loss for pathological features to prioritize clinically relevant regions. Additionally, a feature gating mechanism is employed to filter out low-quality concept features. Finally, the visual features are corresponding to individual medical concepts, and are leveraged to guide the report generation process. Experiments on two public benchmarks (MIMIC-CXR and CheXpert Plus) demonstrate that MCA-RG achieves superior performance, highlighting its effectiveness in radiology report generation.

Method: Propose an unsupervised debiasing approach that reduces the model’s tendency to predict EOS tokens prematurely, encouraging longer caption generation.

Result: Experiments with three VLMs on three benchmarks show substantial increase in caption length and relevant details, though with increased hallucination rates.

Conclusion: Simple EOS token debiasing effectively improves detailed captioning without complex supervision, making it easily applicable to any pretrained VLM.

Abstract: Despite significant advances in vision-language models (VLMs), image captioning often suffers from a lack of detail, with base models producing short, generic captions. This limitation persists even though VLMs are equipped with strong vision and language backbones. While supervised data and complex reward functions have been proposed to improve detailed image captioning, we identify a simpler underlying issue: a bias towards the end-of-sequence (EOS) token, which is introduced during cross-entropy training. We propose an unsupervised method to debias the model’s tendency to predict the EOS token prematurely. By reducing this bias, we encourage the generation of longer, more detailed captions without the need for intricate reward functions or supervision. Our approach is straightforward, effective, and easily applicable to any pretrained model. We demonstrate its effectiveness through experiments with three VLMs and on three detailed captioning benchmarks. Our results show a substantial increase in caption length and relevant details, albeit with an expected increase in the rate of hallucinations.

Method: Cross-modality masked learning approach with two branches: Slice-Depth Transformer for 3D CT features and graph-based Transformer for clinical tabular data, using masked modality strategy to reconstruct missing components for better integration.

Result: Superior performance in multi-modal integration for NSCLC survival prediction, surpassing existing methods and setting new benchmark for prognostic models.

Conclusion: The proposed framework effectively addresses multi-modal feature fusion challenges and demonstrates significant improvements in survival prediction accuracy for NSCLC immunotherapy patients.

Abstract: Accurate prognosis of non-small cell lung cancer (NSCLC) patients undergoing immunotherapy is essential for personalized treatment planning, enabling informed patient decisions, and improving both treatment outcomes and quality of life. However, the lack of large, relevant datasets and effective multi-modal feature fusion strategies pose significant challenges in this domain. To address these challenges, we present a large-scale dataset and introduce a novel framework for multi-modal feature fusion aimed at enhancing the accuracy of survival prediction. The dataset comprises 3D CT images and corresponding clinical records from NSCLC patients treated with immune checkpoint inhibitors (ICI), along with progression-free survival (PFS) and overall survival (OS) data. We further propose a cross-modality masked learning approach for medical feature fusion, consisting of two distinct branches, each tailored to its respective modality: a Slice-Depth Transformer for extracting 3D features from CT images and a graph-based Transformer for learning node features and relationships among clinical variables in tabular data. The fusion process is guided by a masked modality learning strategy, wherein the model utilizes the intact modality to reconstruct missing components. This mechanism improves the integration of modality-specific features, fostering more effective inter-modality relationships and feature interactions. Our approach demonstrates superior performance in multi-modal integration for NSCLC survival prediction, surpassing existing methods and setting a new benchmark for prognostic models in this context.

Method: Hybrid approach combining autoregressive and diffusion models, with Multi-Scale Pose Representation module for detailed feature extraction and Confidence-Aware Causal Attention for dynamic pose generation guidance.

Result: Extensive experiments on PHOENIX14T and How2Sign datasets demonstrate effectiveness in both generation quality and real-time streaming efficiency.

Conclusion: The proposed hybrid model successfully leverages strengths of both autoregressive and diffusion approaches while maintaining real-time applicability for sign language production.

Abstract: Earlier Sign Language Production (SLP) models typically relied on autoregressive methods that generate output tokens one by one, which inherently provide temporal alignment. Although techniques like Teacher Forcing can prevent model collapse during training, they still cannot solve the problem of error accumulation during inference, since ground truth is unavailable at that stage. In contrast, more recent approaches based on diffusion models leverage step-by-step denoising to enable high-quality generation. However, the iterative nature of these models and the requirement to denoise entire sequences limit their applicability in real-time tasks like SLP. To address it, we apply a hybrid approach combining autoregressive and diffusion models to SLP for the first time, leveraging the strengths of both models in sequential dependency modeling and output refinement. To capture fine-grained body movements, we design a Multi-Scale Pose Representation module that separately extracts detailed features from distinct articulators and integrates them via a Multi-Scale Fusion module. Furthermore, we introduce a Confidence-Aware Causal Attention mechanism that utilizes joint-level confidence scores to dynamically guide the pose generation process, improving accuracy and robustness. Extensive experiments on the PHOENIX14T and How2Sign datasets demonstrate the effectiveness of our method in both generation quality and real-time streaming efficiency.

Method: Iterative in-painting guided by MLLM prompts to generate boundary-aligned OOD samples, refined via noise adjustments based on energy score gradients, then fine-tuning CLIP encoder with negative labels.

Result: Achieves state-of-the-art performance on large-scale ImageNet benchmark with minimal parameter and runtime increases, significantly surpassing existing methods.

Conclusion: SynOOD effectively leverages foundation models to synthesize challenging boundary samples, enhancing CLIP’s OOD detection capabilities without substantial computational cost.

Abstract: Pre-trained vision-language models have exhibited remarkable abilities in detecting out-of-distribution (OOD) samples. However, some challenging OOD samples, which lie close to in-distribution (InD) data in image feature space, can still lead to misclassification. The emergence of foundation models like diffusion models and multimodal large language models (MLLMs) offers a potential solution to this issue. In this work, we propose SynOOD, a novel approach that harnesses foundation models to generate synthetic, challenging OOD data for fine-tuning CLIP models, thereby enhancing boundary-level discrimination between InD and OOD samples. Our method uses an iterative in-painting process guided by contextual prompts from MLLMs to produce nuanced, boundary-aligned OOD samples. These samples are refined through noise adjustments based on gradients from OOD scores like the energy score, effectively sampling from the InD/OOD boundary. With these carefully synthesized images, we fine-tune the CLIP image encoder and negative label features derived from the text encoder to strengthen connections between near-boundary OOD samples and a set of negative labels. Finally, SynOOD achieves state-of-the-art performance on the large-scale ImageNet benchmark, with minimal increases in parameters and runtime. Our approach significantly surpasses existing methods, and the code is available at https://github.com/Jarvisgivemeasuit/SynOOD.

Method: Developed five progressive models (Genesis to Nexus) using CNN-LSTM encoder-decoder architecture, evolving to include EfficientNetV2B3 backbone and dynamic attention mechanism. Trained on MS COCO 2017 dataset.

Result: Final model Nexus achieved BLEU-4 score of 31.4, surpassing foundational benchmarks. Key finding: upgrading visual backbone without attention mechanism degrades performance due to single-vector bottleneck.

Conclusion: The iterative design process validates architectural shift to attention mechanisms and provides a replicable blueprint for understanding core principles in vision-language tasks.

Abstract: Image captioning, situated at the intersection of computer vision and natural language processing, requires a sophisticated understanding of both visual scenes and linguistic structure. While modern approaches are dominated by large-scale Transformer architectures, this paper documents a systematic, iterative development of foundational image captioning models, progressing from a simple CNN-LSTM encoder-decoder to a competitive attention-based system. This paper presents a series of five models, beginning with Genesis and concluding with Nexus, an advanced model featuring an EfficientNetV2B3 backbone and a dynamic attention mechanism. The experiments chart the impact of architectural enhancements and demonstrate a key finding within the classic CNN-LSTM paradigm: merely upgrading the visual backbone without a corresponding attention mechanism can degrade performance, as the single-vector bottleneck cannot transmit the richer visual detail. This insight validates the architectural shift to attention. Trained on the MS COCO 2017 dataset, the final model, Nexus, achieves a BLEU-4 score of 31.4, surpassing several foundational benchmarks and validating the iterative design process. This work provides a clear, replicable blueprint for understanding the core architectural principles that underpin modern vision-language tasks.

Method: Created dataset using DSLR camera photographs of teeth from 295 patients, labeled for instance segmentation across 9 classes, and conducted neural network training experiments.

Result: The experiments demonstrated high quality predictions for instance segmentation tasks using the AlphaDent dataset.

Conclusion: AlphaDent is a valuable open-source resource for dental AI research with proven effectiveness in instance segmentation applications, complete with available code and model weights.

Abstract: In this article, we present a new unique dataset for dental research - AlphaDent. This dataset is based on the DSLR camera photographs of the teeth of 295 patients and contains over 1200 images. The dataset is labeled for solving the instance segmentation problem and is divided into 9 classes. The article provides a detailed description of the dataset and the labeling format. The article also provides the details of the experiment on neural network training for the Instance Segmentation problem using this dataset. The results obtained show high quality of predictions. The dataset is published under an open license; and the training/inference code and model weights are also available under open licenses.

Method: LORE directly optimizes the inverted noise to address generalization and controllability limitations, using latent-space optimization without requiring architectural modification or model fine-tuning.

Result: Comprehensive evaluations on PIEBench, SmartEdit, and GapEdit benchmarks show LORE significantly outperforms baselines in semantic alignment, image quality, and background fidelity.

Conclusion: LORE demonstrates the effectiveness and scalability of latent-space optimization for general-purpose image editing, providing stable and controllable concept replacement without training requirements.

Abstract: Text-driven image editing enables users to flexibly modify visual content through natural language instructions, and is widely applied to tasks such as semantic object replacement, insertion, and removal. While recent inversion-based editing methods using rectified flow models have achieved promising results in image quality, we identify a structural limitation in their editing behavior: the semantic bias toward the source concept encoded in the inverted noise tends to suppress attention to the target concept. This issue becomes particularly critical when the source and target semantics are dissimilar, where the attention mechanism inherently leads to editing failure or unintended modifications in non-target regions. In this paper, we systematically analyze and validate this structural flaw, and introduce LORE, a training-free and efficient image editing method. LORE directly optimizes the inverted noise, addressing the core limitations in generalization and controllability of existing approaches, enabling stable, controllable, and general-purpose concept replacement, without requiring architectural modification or model fine-tuning. We conduct comprehensive evaluations on three challenging benchmarks: PIEBench, SmartEdit, and GapEdit. Experimental results show that LORE significantly outperforms strong baselines in terms of semantic alignment, image quality, and background fidelity, demonstrating the effectiveness and scalability of latent-space optimization for general-purpose image editing. Our implementation is available at https://github.com/oyly16/LORE.

Method: Feedforward artificial neural network trained with a new method on ImageNet 2012 dataset, using 322,430,160 parameters with 4 decimal places precision.

Result: Achieved 98.3% accuracy with 99.69% Top-1 rate and average 285.9 perfectly classified labels across 10 batch partitions.

Conclusion: The model performs exceptionally well but doesn’t reach 100% accuracy likely due to double-labeling issues with duplicate images having different labels in the dataset.

Abstract: In this paper, we describe a feedforward artificial neural network trained on the ImageNet 2012 contest dataset [7] with the new method of [5] to an accuracy rate of 98.3% with a 99.69 Top-1 rate, and an average of 285.9 labels that are perfectly classified over the 10 batch partitions of the dataset. The best performing model uses 322,430,160 parameters, with 4 decimal places precision. We conjecture that the reason our model does not achieve a 100% accuracy rate is due to a double-labeling problem, by which there are duplicate images in the dataset with different labels.

Method: LV-Net deforms an anatomy-aware joint LV-hippocampus template mesh, incorporates anatomical relationships to reduce boundary artifacts, and classifies template vertices based on anatomical adjacency to enhance point correspondence.

Result: LV-Net achieves superior reconstruction accuracy even with imperfect segmentations, provides more reliable shape descriptors across datasets, and identifies LV subregions significantly associated with Alzheimer’s disease.

Conclusion: The framework demonstrates robust LV shape modeling capabilities, offering improved biomarkers for neurological disease analysis, with code publicly available for research use.

Abstract: Lateral ventricle (LV) shape analysis holds promise as a biomarker for neurological diseases; however, challenges remain due to substantial shape variability across individuals and segmentation difficulties arising from limited MRI resolution. We introduce LV-Net, a novel framework for producing individualized 3D LV meshes from brain MRI by deforming an anatomy-aware joint LV-hippocampus template mesh. By incorporating anatomical relationships embedded within the joint template, LV-Net reduces boundary segmentation artifacts and improves reconstruction robustness. In addition, by classifying the vertices of the template mesh based on their anatomical adjacency, our method enhances point correspondence across subjects, leading to more accurate LV shape statistics. We demonstrate that LV-Net achieves superior reconstruction accuracy, even in the presence of segmentation imperfections, and delivers more reliable shape descriptors across diverse datasets. Finally, we apply LV-Net to Alzheimer’s disease analysis, identifying LV subregions that show significantly associations with the disease relative to cognitively normal controls. The codes for LV shape modeling are available at https://github.com/PWonjung/LV_Shape_Modeling.

Method: Created MultiRef-bench with 990 synthetic and 1,000 real-world samples, developed RefBlend data engine with 10 reference types and 33 combinations, and built MultiRef dataset with 38k high-quality images. Evaluated 3 interleaved image-text models and 6 agentic frameworks.

Result: Best model (OmniGen) achieved only 66.6% on synthetic samples and 79.0% on real-world cases compared to golden answers, demonstrating significant challenges in multi-reference conditioning.

Conclusion: Current state-of-the-art systems struggle with multi-reference image generation, highlighting the need for more flexible and human-like creative tools that can effectively integrate multiple visual inspiration sources.

Abstract: Visual designers naturally draw inspiration from multiple visual references, combining diverse elements and aesthetic principles to create artwork. However, current image generative frameworks predominantly rely on single-source inputs – either text prompts or individual reference images. In this paper, we focus on the task of controllable image generation using multiple visual references. We introduce MultiRef-bench, a rigorous evaluation framework comprising 990 synthetic and 1,000 real-world samples that require incorporating visual content from multiple reference images. The synthetic samples are synthetically generated through our data engine RefBlend, with 10 reference types and 33 reference combinations. Based on RefBlend, we further construct a dataset MultiRef containing 38k high-quality images to facilitate further research. Our experiments across three interleaved image-text models (i.e., OmniGen, ACE, and Show-o) and six agentic frameworks (e.g., ChatDiT and LLM + SD) reveal that even state-of-the-art systems struggle with multi-reference conditioning, with the best model OmniGen achieving only 66.6% in synthetic samples and 79.0% in real-world cases on average compared to the golden answer. These findings provide valuable directions for developing more flexible and human-like creative tools that can effectively integrate multiple sources of visual inspiration. The dataset is publicly available at: https://multiref.github.io/.

Method: Proposes CMAMRNet with two key components: Mask-Aware Up/Down-Sampler (MAUDS) for consistent mask sensitivity across resolution scales, and Co-Feature Aggregator (CFA) for extracting complementary features at different resolutions.

Result: Experimental results on benchmark datasets show CMAMRNet outperforms state-of-the-art methods, effectively preserving both structural integrity and artistic details in restored murals.

Conclusion: The proposed CMAMRNet framework successfully addresses the limitations of existing mural restoration methods through comprehensive mask guidance and multi-scale feature extraction, achieving superior restoration quality while maintaining artistic authenticity.

Abstract: Murals, as invaluable cultural artifacts, face continuous deterioration from environmental factors and human activities. Digital restoration of murals faces unique challenges due to their complex degradation patterns and the critical need to preserve artistic authenticity. Existing learning-based methods struggle with maintaining consistent mask guidance throughout their networks, leading to insufficient focus on damaged regions and compromised restoration quality. We propose CMAMRNet, a Contextual Mask-Aware Mural Restoration Network that addresses these limitations through comprehensive mask guidance and multi-scale feature extraction. Our framework introduces two key components: (1) the Mask-Aware Up/Down-Sampler (MAUDS), which ensures consistent mask sensitivity across resolution scales through dedicated channel-wise feature selection and mask-guided feature fusion; and (2) the Co-Feature Aggregator (CFA), operating at both the highest and lowest resolutions to extract complementary features for capturing fine textures and global structures in degraded regions. Experimental results on benchmark datasets demonstrate that CMAMRNet outperforms state-of-the-art methods, effectively preserving both structural integrity and artistic details in restored murals. The code is available at~\href{https://github.com/CXH-Research/CMAMRNet}{https://github.com/CXH-Research/CMAMRNet}.

Method: Integrates parameter-efficient Conv-LoRA adapter to inject local inductive biases, and Dynamic Fusion Gateway that uses visual context to adaptively modulate text prompts for bidirectional fusion.

Result: Extensive experiments show superior accuracy and robustness on diverse industrial and medical benchmarks.

Conclusion: Synergistic co-design of feature representation and cross-modal fusion is critical for robustly adapting foundation models to dense perception tasks.

Abstract: Pre-trained Vision-Language Models (VLMs) face a significant adaptation gap when applied to Zero-Shot Anomaly Detection (ZSAD), stemming from their lack of local inductive biases for dense prediction and their reliance on inflexible feature fusion paradigms. We address these limitations through an Architectural Co-Design framework that jointly refines feature representation and cross-modal fusion. Our method integrates a parameter-efficient Convolutional Low-Rank Adaptation (Conv-LoRA) adapter to inject local inductive biases for fine-grained representation, and introduces a Dynamic Fusion Gateway (DFG) that leverages visual context to adaptively modulate text prompts, enabling a powerful bidirectional fusion. Extensive experiments on diverse industrial and medical benchmarks demonstrate superior accuracy and robustness, validating that this synergistic co-design is critical for robustly adapting foundation models to dense perception tasks.

Method: Introduces AURA benchmark with questions across six cognitive domains designed to be unanswerable from single modality, and proposes AuraScore metric that evaluates reasoning fidelity through Factual Consistency and Core Inference.

Result: Evaluations show SOTA models achieve up to 92% accuracy but have Factual Consistency and Core Inference scores below 45%, indicating models arrive at correct answers through flawed logic.

Conclusion: The benchmark reveals a critical reasoning gap in current models and paves the way for more robust multimodal evaluation that assesses reasoning processes.

Abstract: Current audio-visual (AV) benchmarks focus on final answer accuracy, overlooking the underlying reasoning process. This makes it difficult to distinguish genuine comprehension from correct answers derived through flawed reasoning or hallucinations. To address this, we introduce AURA (Audio-visual Understanding and Reasoning Assessment), a benchmark for evaluating the cross-modal reasoning capabilities of Audio-Visual Large Language Models (AV-LLMs) and Omni-modal Language Models (OLMs). AURA includes questions across six challenging cognitive domains, such as causality, timbre and pitch, tempo and AV synchronization, unanswerability, implicit distractions, and skill profiling, explicitly designed to be unanswerable from a single modality. This forces models to construct a valid logical path grounded in both audio and video, setting AURA apart from AV datasets that allow uni-modal shortcuts. To assess reasoning traces, we propose a novel metric, AuraScore, which addresses the lack of robust tools for evaluating reasoning fidelity. It decomposes reasoning into two aspects: (i) Factual Consistency - whether reasoning is grounded in perceptual evidence, and (ii) Core Inference - the logical validity of each reasoning step. Evaluations of SOTA models on AURA reveal a critical reasoning gap: although models achieve high accuracy (up to 92% on some tasks), their Factual Consistency and Core Inference scores fall below 45%. This discrepancy highlights that models often arrive at correct answers through flawed logic, underscoring the need for our benchmark and paving the way for more robust multimodal evaluation.

Method: Proposes ReconSimulator combining video diffusion prior for appearance modeling and kinematic model for physical modeling, Dynamic Adversary Agent for corner-case generation, and Cousin Trajectory Generator to address biased training data.

Result: Achieves 5x reduction in Collision Ratio compared to imitation learning methods, demonstrating improved end-to-end autonomous driving training performance.

Conclusion: The framework successfully narrows the sim2real gap and enhances reinforcement learning for autonomous driving by integrating diffusion priors with scene reconstruction and addressing data distribution biases.

Abstract: Reinforcement learning for training end-to-end autonomous driving models in closed-loop simulations is gaining growing attention. However, most simulation environments differ significantly from real-world conditions, creating a substantial simulation-to-reality (sim2real) gap. To bridge this gap, some approaches utilize scene reconstruction techniques to create photorealistic environments as a simulator. While this improves realistic sensor simulation, these methods are inherently constrained by the distribution of the training data, making it difficult to render high-quality sensor data for novel trajectories or corner case scenarios. Therefore, we propose ReconDreamer-RL, a framework designed to integrate video diffusion priors into scene reconstruction to aid reinforcement learning, thereby enhancing end-to-end autonomous driving training. Specifically, in ReconDreamer-RL, we introduce ReconSimulator, which combines the video diffusion prior for appearance modeling and incorporates a kinematic model for physical modeling, thereby reconstructing driving scenarios from real-world data. This narrows the sim2real gap for closed-loop evaluation and reinforcement learning. To cover more corner-case scenarios, we introduce the Dynamic Adversary Agent (DAA), which adjusts the trajectories of surrounding vehicles relative to the ego vehicle, autonomously generating corner-case traffic scenarios (e.g., cut-in). Finally, the Cousin Trajectory Generator (CTG) is proposed to address the issue of training data distribution, which is often biased toward simple straight-line movements. Experiments show that ReconDreamer-RL improves end-to-end autonomous driving training, outperforming imitation learning methods with a 5x reduction in the Collision Ratio.

Method: Uses a top-down approach to decompose, summarize, and reformulate papers, followed by bottom-up video generation with Progressive Chain of Thought (P-CoT) for granular iterative planning and cross-modal alignment through defined key scenes.

Result: Successfully generates high-quality video abstracts across five research fields, demonstrating expertise beyond current video generation models.

Conclusion: Preacher addresses the limitations of existing video generation models and provides an effective framework for converting research papers into accessible video abstracts with improved domain knowledge representation and structural coherence.

Abstract: The paper-to-video task converts a research paper into a structured video abstract, distilling key concepts, methods, and conclusions into an accessible, well-organized format. While state-of-the-art video generation models demonstrate potential, they are constrained by limited context windows, rigid video duration constraints, limited stylistic diversity, and an inability to represent domain-specific knowledge. To address these limitations, we introduce Preacher, the first paper-to-video agentic system. Preacher employs a topdown approach to decompose, summarize, and reformulate the paper, followed by bottom-up video generation, synthesizing diverse video segments into a coherent abstract. To align cross-modal representations, we define key scenes and introduce a Progressive Chain of Thought (P-CoT) for granular, iterative planning. Preacher successfully generates high-quality video abstracts across five research fields, demonstrating expertise beyond current video generation models. Code will be released at: https://github.com/GenVerse/Paper2Video

Method: Developed HistoPLUS model trained on a novel curated pan-cancer dataset of 108,722 nuclei covering 13 cell types, using a more efficient architecture with fewer parameters.

Result: Outperforms current SOTA models by 5.2% in detection quality and 23.7% in overall F1 classification score, while using 5x fewer parameters. Enables study of 7 understudied cell types and shows robust transfer to unseen oncology indications.

Conclusion: HistoPLUS significantly advances cell analysis in tumor microenvironments, particularly for understudied cell types, with better generalization and efficiency. Model weights and code are publicly released to support broader biomarker research.

Abstract: Cell detection, segmentation and classification are essential for analyzing tumor microenvironments (TME) on hematoxylin and eosin (H&E) slides. Existing methods suffer from poor performance on understudied cell types (rare or not present in public datasets) and limited cross-domain generalization. To address these shortcomings, we introduce HistoPLUS, a state-of-the-art model for cell analysis, trained on a novel curated pan-cancer dataset of 108,722 nuclei covering 13 cell types. In external validation across 4 independent cohorts, HistoPLUS outperforms current state-of-the-art models in detection quality by 5.2% and overall F1 classification score by 23.7%, while using 5x fewer parameters. Notably, HistoPLUS unlocks the study of 7 understudied cell types and brings significant improvements on 8 of 13 cell types. Moreover, we show that HistoPLUS robustly transfers to two oncology indications unseen during training. To support broader TME biomarker research, we release the model weights and inference code at https://github.com/owkin/histoplus/.

Method: Created a multi-resolution, multi-modal dataset with synchronized RGB and event cameras under variable lighting. Evaluated two baseline tasks: Action Unit classification using Spiking Neural Networks and frame reconstruction using Conditional Variational Autoencoders.

Result: Event-based data achieved 51.23% accuracy for Action Unit classification vs 23.12% with RGB. Frame reconstruction achieved SSIM = 0.8513 and PSNR = 26.89 dB with high-resolution event input.

Conclusion: Event-based data shows promising results for micro-expression recognition and frame reconstruction, demonstrating the potential of event cameras to overcome limitations of traditional RGB cameras for capturing subtle facial movements.

Abstract: Micro-expression analysis has applications in domains such as Human-Robot Interaction and Driver Monitoring Systems. Accurately capturing subtle and fast facial movements remains difficult when relying solely on RGB cameras, due to limitations in temporal resolution and sensitivity to motion blur. Event cameras offer an alternative, with microsecond-level precision, high dynamic range, and low latency. However, public datasets featuring event-based recordings of Action Units are still scarce. In this work, we introduce a novel, preliminary multi-resolution and multi-modal micro-expression dataset recorded with synchronized RGB and event cameras under variable lighting conditions. Two baseline tasks are evaluated to explore the spatial-temporal dynamics of micro-expressions: Action Unit classification using Spiking Neural Networks (51.23% accuracy with events vs. 23.12% with RGB), and frame reconstruction using Conditional Variational Autoencoders, achieving SSIM = 0.8513 and PSNR = 26.89 dB with high-resolution event input. These promising results show that event-based data can be used for micro-expression recognition and frame reconstruction.

Method: Combines panorama video generation (using dual-branch model with global panorama and local perspective branches with cross-attention) and dynamic scene reconstruction (using geometry-aligned 3D Gaussian Splatting with metric depth maps and estimated camera poses).

Result: Extensive experiments demonstrate effectiveness and superiority in generating visually compelling and motion-coherent dynamic panoramic scenes.

Conclusion: TiP4GEN successfully addresses the gap in 360-degree immersive dynamic scene generation with fine-grained control and geometric consistency.

Abstract: With the rapid advancement and widespread adoption of VR/AR technologies, there is a growing demand for the creation of high-quality, immersive dynamic scenes. However, existing generation works predominantly concentrate on the creation of static scenes or narrow perspective-view dynamic scenes, falling short of delivering a truly 360-degree immersive experience from any viewpoint. In this paper, we introduce \textbf{TiP4GEN}, an advanced text-to-dynamic panorama scene generation framework that enables fine-grained content control and synthesizes motion-rich, geometry-consistent panoramic 4D scenes. TiP4GEN integrates panorama video generation and dynamic scene reconstruction to create 360-degree immersive virtual environments. For video generation, we introduce a \textbf{Dual-branch Generation Model} consisting of a panorama branch and a perspective branch, responsible for global and local view generation, respectively. A bidirectional cross-attention mechanism facilitates comprehensive information exchange between the branches. For scene reconstruction, we propose a \textbf{Geometry-aligned Reconstruction Model} based on 3D Gaussian Splatting. By aligning spatial-temporal point clouds using metric depth maps and initializing scene cameras with estimated poses, our method ensures geometric consistency and temporal coherence for the reconstructed scenes. Extensive experiments demonstrate the effectiveness of our proposed designs and the superiority of TiP4GEN in generating visually compelling and motion-coherent dynamic panoramic scenes. Our project page is at https://ke-xing.github.io/TiP4GEN/.

Method: Comparative analysis of two RT-DETR variants (Large and Extra-Large) trained on coastal debris dataset, evaluating accuracy (mAP metrics) and inference speed.

Result: RT-DETR-X achieved slightly higher accuracy (mAP@50: 0.816, mAP@50-95: 0.612) but RT-DETR-L was significantly faster (20.1ms vs 34.5ms inference time) with comparable accuracy (mAP@50: 0.810, mAP@50-95: 0.606).

Conclusion: RT-DETR-L provides more practical real-time deployment due to better speed-accuracy balance, highlighting trade-offs between model complexity and operational viability for environmental conservation applications.

Abstract: Coastal pollution is a pressing global environmental issue, necessitating scalable and automated solutions for monitoring and management. This study investigates the efficacy of the Real-Time Detection Transformer (RT-DETR), a state-of-the-art, end-to-end object detection model, for the automated detection and counting of beach litter. A rigorous comparative analysis is conducted between two model variants, RT-DETR-Large (RT-DETR-L) and RT-DETR-Extra-Large (RT-DETR-X), trained on a publicly available dataset of coastal debris. The evaluation reveals that the RT-DETR-X model achieves marginally superior accuracy, with a mean Average Precision at 50% IoU (mAP@50) of 0.816 and a mAP@50-95 of 0.612, compared to the RT-DETR-L model’s 0.810 and 0.606, respectively. However, this minor performance gain is realized at a significant computational cost; the RT-DETR-L model demonstrates a substantially faster inference time of 20.1 ms versus 34.5 ms for the RT-DETR-X. The findings suggest that the RT-DETR-L model offers a more practical and efficient solution for real-time, in-field deployment due to its superior balance of processing speed and detection accuracy. This research provides valuable insights into the application of advanced Transformer-based detectors for environmental conservation, highlighting the critical trade-offs between model complexity and operational viability.

Method: Comparative analysis between Vision Transformers (ViTs) and CNN-based models (ResNet50) on two ex vivo datasets containing CCD camera and flexible ureteroscope images. ViT-base model was pretrained on ImageNet-21k.

Result: ViT consistently outperformed ResNet50 across all conditions. In complex endoscopic images: 95.2% accuracy vs 64.5%, 95.1% F1-score vs 59.3%. In mixed-view CCD images: 87.1% accuracy vs 78.4%. Improvements extended to precision and recall metrics.

Conclusion: ViT-based architectures provide superior classification performance and offer a scalable alternative to conventional CNNs for kidney stone image analysis, demonstrating better handling of complex imaging conditions.

Abstract: Kidney stone classification from endoscopic images is critical for personalized treatment and recurrence prevention. While convolutional neural networks (CNNs) have shown promise in this task, their limited ability to capture long-range dependencies can hinder performance under variable imaging conditions. This study presents a comparative analysis between Vision Transformers (ViTs) and CNN-based models, evaluating their performance on two ex vivo datasets comprising CCD camera and flexible ureteroscope images. The ViT-base model pretrained on ImageNet-21k consistently outperformed a ResNet50 baseline across multiple imaging conditions. For instance, in the most visually complex subset (Section patches from endoscopic images), the ViT model achieved 95.2% accuracy and 95.1% F1-score, compared to 64.5% and 59.3% with ResNet50. In the mixed-view subset from CCD-camera images, ViT reached 87.1% accuracy versus 78.4% with CNN. These improvements extend across precision and recall as well. The results demonstrate that ViT-based architectures provide superior classification performance and offer a scalable alternative to conventional CNNs for kidney stone image analysis.

Method: DictAS consists of three components: (1) Dictionary Construction using normal reference image features, (2) Dictionary Lookup with sparse retrieval strategy to identify anomalies when features cannot be retrieved, and (3) Query Discrimination Regularization with Contrastive Query Constraint and Text Alignment Constraint to enhance anomaly discrimination.

Result: Extensive experiments on seven public industrial and medical datasets show that DictAS consistently outperforms state-of-the-art few-shot anomaly segmentation methods.

Conclusion: The proposed DictAS framework successfully transfers dictionary lookup capabilities to few-shot anomaly segmentation for unseen classes through self-supervised learning, demonstrating superior performance without requiring retraining on target data.

Abstract: Recent vision-language models (e.g., CLIP) have demonstrated remarkable class-generalizable ability to unseen classes in few-shot anomaly segmentation (FSAS), leveraging supervised prompt learning or fine-tuning on seen classes. However, their cross-category generalization largely depends on prior knowledge of real seen anomaly samples. In this paper, we propose a novel framework, namely DictAS, which enables a unified model to detect visual anomalies in unseen object categories without any retraining on the target data, only employing a few normal reference images as visual prompts. The insight behind DictAS is to transfer dictionary lookup capabilities to the FSAS task for unseen classes via self-supervised learning, instead of merely memorizing the normal and abnormal feature patterns from the training set. Specifically, DictAS mainly consists of three components: (1) Dictionary Construction - to simulate the index and content of a real dictionary using features from normal reference images. (2) Dictionary Lookup - to retrieve queried region features from the dictionary via a sparse lookup strategy. When a query feature cannot be retrieved, it is classified as an anomaly. (3) Query Discrimination Regularization - to enhance anomaly discrimination by making abnormal features harder to retrieve from the dictionary. To achieve this, Contrastive Query Constraint and Text Alignment Constraint are further proposed. Extensive experiments on seven public industrial and medical datasets demonstrate that DictAS consistently outperforms state-of-the-art FSAS methods.

Method: GALA uses 3D Gaussian Splatting with self-supervised contrastive learning to distill scene-specific 3D instance features. It introduces a cross-attention module with two learnable codebooks that encode view-independent semantic embeddings, avoiding per-Gaussian high-dimensional feature learning to reduce memory consumption.

Result: Extensive experiments on real-world datasets demonstrate GALA’s remarkable open-vocabulary performance on both 2D and 3D tasks.

Conclusion: GALA provides an effective framework for open-vocabulary 3D scene understanding with strong performance in both 2D and 3D domains while being memory-efficient.

Abstract: 3D scene reconstruction and understanding have gained increasing popularity, yet existing methods still struggle to capture fine-grained, language-aware 3D representations from 2D images. In this paper, we present GALA, a novel framework for open-vocabulary 3D scene understanding with 3D Gaussian Splatting (3DGS). GALA distills a scene-specific 3D instance feature field via self-supervised contrastive learning. To extend to generalized language feature fields, we introduce the core contribution of GALA, a cross-attention module with two learnable codebooks that encode view-independent semantic embeddings. This design not only ensures intra-instance feature similarity but also supports seamless 2D and 3D open-vocabulary queries. It reduces memory consumption by avoiding per-Gaussian high-dimensional feature learning. Extensive experiments on real-world datasets demonstrate GALA’s remarkable open-vocabulary performance on both 2D and 3D.

Method: MoCHA-former uses two main components: Decoupled Moiré Adaptive Demoiréing (DMAD) with Moiré Decoupling Block and Detail Decoupling Block to separate content from artifacts, and Spatio-Temporal Adaptive Demoiréing (STAD) with Spatial Fusion Block and Feature Channel Attention to handle large structures and channel dependencies. It performs implicit frame alignment for temporal consistency without explicit modules.

Result: The method was evaluated on two video datasets covering RAW and sRGB domains. MoCHA-former consistently outperformed prior methods across all metrics including PSNR, SSIM, and LPIPS, demonstrating superior moiré removal performance.

Conclusion: MoCHA-former effectively addresses the key challenges in moiré pattern removal through its hybrid adaptive transformer architecture, achieving state-of-the-art performance in both quantitative metrics and qualitative results for screen content demoiréing.

Abstract: Recent advances in portable imaging have made camera-based screen capture ubiquitous. Unfortunately, frequency aliasing between the camera’s color filter array (CFA) and the display’s sub-pixels induces moir'e patterns that severely degrade captured photos and videos. Although various demoir'eing models have been proposed to remove such moir'e patterns, these approaches still suffer from several limitations: (i) spatially varying artifact strength within a frame, (ii) large-scale and globally spreading structures, (iii) channel-dependent statistics and (iv) rapid temporal fluctuations across frames. We address these issues with the Moir'e Conditioned Hybrid Adaptive Transformer (MoCHA-former), which comprises two key components: Decoupled Moir'e Adaptive Demoir'eing (DMAD) and Spatio-Temporal Adaptive Demoir'eing (STAD). DMAD separates moir'e and content via a Moir'e Decoupling Block (MDB) and a Detail Decoupling Block (DDB), then produces moir'e-adaptive features using a Moir'e Conditioning Block (MCB) for targeted restoration. STAD introduces a Spatial Fusion Block (SFB) with window attention to capture large-scale structures, and a Feature Channel Attention (FCA) to model channel dependence in RAW frames. To ensure temporal consistency, MoCHA-former performs implicit frame alignment without any explicit alignment module. We analyze moir'e characteristics through qualitative and quantitative studies, and evaluate on two video datasets covering RAW and sRGB domains. MoCHA-former consistently surpasses prior methods across PSNR, SSIM, and LPIPS.

Method: Logical entities and relationships are encoded as graph signals, processed via learnable spectral filters for multi-scale information propagation, and mapped into symbolic predicates for rule-based inference. Includes graph Fourier transforms, band-selective attention, and spectral rule grounding.

Result: Experiments on benchmark reasoning datasets (ProofWriter, EntailmentBank, bAbI, CLUTRR, ARC-Challenge) demonstrate improvements in logical consistency, interpretability, and computational efficiency over state-of-the-art neuro-symbolic models.

Conclusion: Graph Signal Processing provides a mathematically grounded and computationally efficient substrate for robust and interpretable reasoning systems, offering significant advantages over conventional neuro-symbolic approaches.

Abstract: We propose a fully spectral, neuro-symbolic reasoning architecture that leverages Graph Signal Processing (GSP) as the primary computational backbone for integrating symbolic logic and neural inference. Unlike conventional reasoning models that treat spectral graph methods as peripheral components, our approach formulates the entire reasoning pipeline in the graph spectral domain. Logical entities and relationships are encoded as graph signals, processed via learnable spectral filters that control multi-scale information propagation, and mapped into symbolic predicates for rule-based inference. We present a complete mathematical framework for spectral reasoning, including graph Fourier transforms, band-selective attention, and spectral rule grounding. Experiments on benchmark reasoning datasets (ProofWriter, EntailmentBank, bAbI, CLUTRR, and ARC-Challenge) demonstrate improvements in logical consistency, interpretability, and computational efficiency over state-of-the-art neuro-symbolic models. Our results suggest that GSP provides a mathematically grounded and computationally efficient substrate for robust and interpretable reasoning systems.

Method: Develops a framework based on Buddhist epistemology, defining telic states as classes of goal-equivalent experience distributions and using statistical divergence between policies and desirable experience features.

Result: Provides a parsimonious account of goal-directed learning that unifies descriptive and prescriptive aspects of world models.

Conclusion: This approach offers a unified account of behavioral, phenomenological and neural dimensions of purposeful behaviors across diverse substrates.

Abstract: Purposeful behavior is a hallmark of natural and artificial intelligence. Its acquisition is often believed to rely on world models, comprising both descriptive (what is) and prescriptive (what is desirable) aspects that identify and evaluate state of affairs in the world, respectively. Canonical computational accounts of purposeful behavior, such as reinforcement learning, posit distinct components of a world model comprising a state representation (descriptive aspect) and a reward function (prescriptive aspect). However, an alternative possibility, which has not yet been computationally formulated, is that these two aspects instead co-emerge interdependently from an agent’s goal. Here, we describe a computational framework of goal-directed state representation in cognitive agents, in which the descriptive and prescriptive aspects of a world model co-emerge from agent-environment interaction sequences, or experiences. Drawing on Buddhist epistemology, we introduce a construct of goal-directed, or telic, states, defined as classes of goal-equivalent experience distributions. Telic states provide a parsimonious account of goal-directed learning in terms of the statistical divergence between behavioral policies and desirable experience features. We review empirical and theoretical literature supporting this novel perspective and discuss its potential to provide a unified account of behavioral, phenomenological and neural dimensions of purposeful behaviors across diverse substrates.

Method: Uses three LLM-based agents (Personalization, Popularity, Sustainability) that generate city suggestions from different angles, with a non-LLM moderator that merges proposals through multi-round negotiation while penalizing repetitive responses.

Result: Experiments on European city queries show improved diversity and overall relevance compared to single-agent baselines, successfully surfacing lesser-visited locales that are typically overlooked.

Conclusion: The multi-stakeholder collaborative approach effectively addresses over-tourism and better aligns with user constraints, demonstrating promise for LLM-driven recommender systems.

Abstract: We propose Collab-REC, a multi-agent framework designed to counteract popularity bias and enhance diversity in tourism recommendations. In our setting, three LLM-based agents – Personalization, Popularity, and Sustainability generate city suggestions from complementary perspectives. A non-LLM moderator then merges and refines these proposals via multi-round negotiation, ensuring each agent’s viewpoint is incorporated while penalizing spurious or repeated responses. Experiments on European city queries show that Collab-REC improves diversity and overall relevance compared to a single-agent baseline, surfacing lesser-visited locales that often remain overlooked. This balanced, context-aware approach addresses over-tourism and better aligns with constraints provided by the user, highlighting the promise of multi-stakeholder collaboration in LLM-driven recommender systems.

Method: Iterative loop with VLM critic proposing edits, multiple LLMs generating SVG updates with diverse strategies, and VLM judge selecting best candidate.

Result: Outperforms GPT-5 and Gemini-2.5-Pro on flowchart reconstruction, accurately composing primitives without unwanted text artifacts.

Conclusion: The approach provides precise, editable SVG outputs that preserve global constraints and support human corrections, with open-source implementation.

Abstract: We study sketch-to-diagram generation: converting rough hand sketches into precise, compositional diagrams. Diffusion models excel at photorealism but struggle with the spatial precision, alignment, and symbolic structure required for flowcharts. We introduce See it. Say it. Sorted., a training-free agentic system that couples a Vision-Language Model (VLM) with Large Language Models (LLMs) to produce editable Scalable Vector Graphics (SVG) programs. The system runs an iterative loop in which a Critic VLM proposes a small set of qualitative, relational edits; multiple candidate LLMs synthesize SVG updates with diverse strategies (conservative->aggressive, alternative, focused); and a Judge VLM selects the best candidate, ensuring stable improvement. This design prioritizes qualitative reasoning over brittle numerical estimates, preserves global constraints (e.g., alignment, connectivity), and naturally supports human-in-the-loop corrections. On 10 sketches derived from flowcharts in published papers, our method more faithfully reconstructs layout and structure than two frontier closed-source image generation LLMs (GPT-5 and Gemini-2.5-Pro), accurately composing primitives (e.g., multi-headed arrows) without inserting unwanted text. Because outputs are programmatic SVGs, the approach is readily extensible to presentation tools (e.g., PowerPoint) via APIs and can be specialized with improved prompts and task-specific tools. The codebase is open-sourced at https://github.com/hantaoZhangrichard/see_it_say_it_sorted.git.

Method: Uses LLMs conditioned on character attributes to generate inter-agent dialogue based on spatial/social relationships, then uses conversations and agent states to control navigation and steering decisions.

Result: Validated in complex scenarios showing automatic grouping/ungrouping of agents and information-passing within crowds, producing more realistic simulations with emergent group behaviors.

Conclusion: The framework enables more realistic crowd simulations by incorporating language-driven social interactions that naturally produce emergent group behaviors from environmental settings.

Abstract: Animating and simulating crowds using an agent-based approach is a well-established area where every agent in the crowd is individually controlled such that global human-like behaviour emerges. We observe that human navigation and movement in crowds are often influenced by complex social and environmental interactions, driven mainly by language and dialogue. However, most existing work does not consider these dimensions and leads to animations where agent-agent and agent-environment interactions are largely limited to steering and fixed higher-level goal extrapolation. We propose a novel method that exploits large language models (LLMs) to control agents’ movement. Our method has two main components: a dialogue system and language-driven navigation. We periodically query agent-centric LLMs conditioned on character personalities, roles, desires, and relationships to control the generation of inter-agent dialogue when necessitated by the spatial and social relationships with neighbouring agents. We then use the conversation and each agent’s personality, emotional state, vision, and physical state to control the navigation and steering of each agent. Our model thus enables agents to make motion decisions based on both their perceptual inputs and the ongoing dialogue. We validate our method in two complex scenarios that exemplify the interplay between social interactions, steering, and crowding. In these scenarios, we observe that grouping and ungrouping of agents automatically occur. Additionally, our experiments show that our method serves as an information-passing mechanism within the crowd. As a result, our framework produces more realistic crowd simulations, with emergent group behaviours arising naturally from any environmental setting.

Method: Designed a multiple memory system that processes short-term memory into multiple long-term memory fragments, creating paired retrieval memory units and contextual memory units with one-to-one correspondence for enhanced retrieval and context utilization.

Result: Experiments on LoCoMo dataset showed superior performance compared to three other methods, with ablation studies confirming memory unit rationality and analysis demonstrating robustness regarding memory segment selection and storage overhead.

Conclusion: The MMS approach effectively utilizes historical data by constructing high-quality long-term memory content, proving practical value for enhancing agent performance through improved memory management.

Abstract: An agent powered by large language models have achieved impressive results, but effectively handling the vast amounts of historical data generated during interactions remains a challenge. The current approach is to design a memory module for the agent to process these data. However, existing methods, such as MemoryBank and A-MEM, have poor quality of stored memory content, which affects recall performance and response quality. In order to better construct high-quality long-term memory content, we have designed a multiple memory system (MMS) inspired by cognitive psychology theory. The system processes short-term memory to multiple long-term memory fragments, and constructs retrieval memory units and contextual memory units based on these fragments, with a one-to-one correspondence between the two. During the retrieval phase, MMS will match the most relevant retrieval memory units based on the user’s query. Then, the corresponding contextual memory units is obtained as the context for the response stage to enhance knowledge, thereby effectively utilizing historical data. Experiments on LoCoMo dataset compared our method with three others, proving its effectiveness. Ablation studies confirmed the rationality of our memory units. We also analyzed the robustness regarding the number of selected memory segments and the storage overhead, demonstrating its practical value.

Method: Evaluated reasoning LLMs using ClimateX dataset expanded to human and planetary health domains, testing different reasoning budgets and comparing pure reasoning with search-augmented generation approaches.

Result: Reasoning LLMs achieved 48.7% accuracy in confidence assessment, but increasing reasoning budgets consistently impaired calibration, causing systematic overconfidence. Search-augmented generation dramatically outperformed with 89.3% accuracy.

Conclusion: Information access, rather than reasoning depth or inference budget, is the critical bottleneck for improved confidence calibration in knowledge-intensive tasks, challenging the “test-time scaling” paradigm.

Abstract: Large Language Models deployed as question answering tools require robust calibration to avoid overconfidence. We systematically evaluate how reasoning capabilities and budget affect confidence assessment accuracy, using the ClimateX dataset (Lacombe et al., 2023) and expanding it to human and planetary health. Our key finding challenges the “test-time scaling” paradigm: while recent reasoning LLMs achieve 48.7% accuracy in assessing expert confidence, increasing reasoning budgets consistently impairs rather than improves calibration. Extended reasoning leads to systematic overconfidence that worsens with longer thinking budgets, producing diminishing and negative returns beyond modest computational investments. Conversely, search-augmented generation dramatically outperforms pure reasoning, achieving 89.3% accuracy by retrieving relevant evidence. Our results suggest that information access, rather than reasoning depth or inference budget, may be the critical bottleneck for improved confidence calibration of knowledge-intensive tasks.

Method: Utilizing the CogniSAT-6 satellite equipped with visible and near infrared hyperspectral instrument and neural network acceleration hardware to perform deep learning and spectral analysis algorithms directly onboard.

Result: The paper demonstrates successful implementation of data analysis and inference capabilities onboard the satellite for various applications, though specific results are not detailed in the abstract.

Conclusion: Onboard data processing using specialized hardware enables advanced Earth science applications and demonstrates the feasibility of edge computing in space environments for real-time analysis and response capabilities.

Abstract: In partnership with Ubotica Technologies, the Jet Propulsion Laboratory is demonstrating state-of-the-art data analysis onboard CogniSAT-6/HAMMER (CS-6). CS-6 is a satellite with a visible and near infrared range hyperspectral instrument and neural network acceleration hardware. Performing data analysis at the edge (e.g. onboard) can enable new Earth science measurements and responses. We will demonstrate data analysis and inference onboard CS-6 for numerous applications using deep learning and spectral analysis algorithms.

Method: Intended Cooperation Values (ICVs) based on information-theoretic Shapley values that measure agents’ causal influence on co-players’ instrumental empowerment through decision uncertainty and preference alignment analysis.

Result: ICVs successfully identify beneficial agent behaviors that foster deterministic decisions or preserve flexibility, revealing cooperation dynamics across both cooperative and competitive MARL environments.

Conclusion: The proposed ICV method provides novel insights into cooperation dynamics and enhances explainability in MARL systems by extracting meaningful agent behavior insights solely from policy distributions.

Abstract: To reliably deploy Multi-Agent Reinforcement Learning (MARL) systems, it is crucial to understand individual agent behaviors within a team. While prior work typically evaluates overall team performance based on explicit reward signals or learned value functions, it is unclear how to infer agent contributions in the absence of any value feedback. In this work, we investigate whether meaningful insights into agent behaviors can be extracted that are consistent with the underlying value functions, solely by analyzing the policy distribution. Inspired by the phenomenon that intelligent agents tend to pursue convergent instrumental values, which generally increase the likelihood of task success, we introduce Intended Cooperation Values (ICVs), a method based on information-theoretic Shapley values for quantifying each agent’s causal influence on their co-players’ instrumental empowerment. Specifically, ICVs measure an agent’s action effect on its teammates’ policies by assessing their decision uncertainty and preference alignment. The analysis across cooperative and competitive MARL environments reveals the extent to which agents adopt similar or diverse strategies. By comparing action effects between policies and value functions, our method identifies which agent behaviors are beneficial to team success, either by fostering deterministic decisions or by preserving flexibility for future action choices. Our proposed method offers novel insights into cooperation dynamics and enhances explainability in MARL systems.

Method: Uses Python visualization libraries to generate charts and tables programmatically, pairs each visual with systematically generated multi-step analytical questions, and provides answers in structured formats (JSON/YAML) with strict factual guidelines.

Result: Developed a benchmark with taxonomy of reasoning types (comparison, trend identification, ranking, aggregation, proportion estimation, anomaly detection) enabling fine-grained assessment of multimodal models.

Conclusion: GRAFT sets a new evaluation standard for multimodal models by providing a unified framework for comprehensive benchmarking of visually grounded, structured reasoning tasks with precise aspect-based evaluation.

Abstract: GRAFT is a structured multimodal benchmark for evaluating models on instruction-following, visual reasoning, and visual-textual alignment tasks. It features programmatically generated charts and synthetically rendered tables, created with Python visualization libraries to ensure control over data semantics, structure, and clarity. Each GRAFT instance pairs a chart or table image with a systematically generated, multi-step analytical question based solely on visual content. Answers are provided in structured formats such as JSON or YAML, supporting consistent evaluation of both reasoning and output format. The benchmark introduces a taxonomy of reasoning types including comparison, trend identification, ranking, aggregation, proportion estimation, and anomaly detection to enable comprehensive assessment. Reference answers follow strict factual and formatting guidelines for precise, aspect-based evaluation. GRAFT offers a unified, scalable framework for fine-grained benchmarking of multimodal models on visually grounded, structured reasoning tasks, setting a new evaluation standard in this field.

Method: Uses MAS-SVD to analyze LoRA update structures and Spectral Sharpness Index to detect unsafe layers, which are pruned post-hoc while maintaining task performance.

Result: Consistently improves safety metrics while maintaining utility, reduces inference costs, and works across agent planning and language generation tasks.

Conclusion: S3LoRA provides a practical, scalable solution for safe deployment of LLM-based agents in resource-constrained and safety-critical environments.

Abstract: Adapting Large Language Models (LLMs) using parameter-efficient fine-tuning (PEFT) techniques such as LoRA has enabled powerful capabilities in LLM-based agents. However, these adaptations can unintentionally compromise safety alignment, leading to unsafe or unstable behaviors, particularly in agent planning tasks. Existing safety-aware adaptation methods often require access to both base and instruction-tuned model checkpoints, which are frequently unavailable in practice, limiting their applicability. We propose S3LoRA (Safe Spectral Sharpness-Guided Pruning LoRA), a lightweight, data-free, and model-independent framework that mitigates safety risks in LoRA-adapted models by inspecting only the fine-tuned weight updates. We first introduce Magnitude-Aware Spherically Normalized SVD (MAS-SVD), which robustly analyzes the structural properties of LoRA updates while preserving global magnitude information. We then design the Spectral Sharpness Index (SSI), a sharpness-aware metric to detect layers with highly concentrated and potentially unsafe updates. These layers are pruned post-hoc to reduce risk without sacrificing task performance. Extensive experiments and ablation studies across agent planning and language generation tasks show that S3LoRA consistently improves safety metrics while maintaining or improving utility metrics and significantly reducing inference cost. These results establish S3LoRA as a practical and scalable solution for safely deploying LLM-based agents in real-world, resource-constrained, and safety-critical environments.

Method: Language-Guided Tuning (LGT) framework with three specialized LLM agents: Advisor (proposes changes), Evaluator (assesses progress), and Optimizer (refines decisions), using textual gradients for semantic understanding.

Result: Substantial improvements over traditional optimization methods on six diverse datasets, achieving performance gains while maintaining high interpretability.

Conclusion: LGT provides an effective framework for ML configuration optimization that combines numerical optimization with semantic reasoning through multi-agent LLMs and textual gradients.

Abstract: Configuration optimization remains a critical bottleneck in machine learning, requiring coordinated tuning across model architecture, training strategy, feature engineering, and hyperparameters. Traditional approaches treat these dimensions independently and lack interpretability, while recent automated methods struggle with dynamic adaptability and semantic reasoning about optimization decisions. We introduce Language-Guided Tuning (LGT), a novel framework that employs multi-agent Large Language Models to intelligently optimize configurations through natural language reasoning. We apply textual gradients - qualitative feedback signals that complement numerical optimization by providing semantic understanding of training dynamics and configuration interdependencies. LGT coordinates three specialized agents: an Advisor that proposes configuration changes, an Evaluator that assesses progress, and an Optimizer that refines the decision-making process, creating a self-improving feedback loop. Through comprehensive evaluation on six diverse datasets, LGT demonstrates substantial improvements over traditional optimization methods, achieving performance gains while maintaining high interpretability.

Method: Model workforce management as abstract argumentation framework in an industrial application context.

Result: User study shows the tool with argumentation-based explanations leads to faster and more accurate problem solving compared to conventional manual solutions.

Conclusion: Abstract argumentation provides an effective framework for workforce management that accommodates changes and delivers faithful explanations to stakeholders.

Abstract: Workforce management is a complex problem optimising the makespan and travel distance required for a team of operators to complete a set of jobs, using a set of instruments. A crucial challenge in workforce management is accommodating changes at execution time so that explanations are provided to all stakeholders involved. Here, we show that, by understanding workforce management as abstract argumentation in an industrial application, we can accommodate change and obtain faithful explanations. We show, with a user study, that our tool and explanations lead to faster and more accurate problem solving than conventional solutions by hand.

Method: GOOD prompts LLMs to simulate users with complex intents, using responses for probabilistic inference over candidate natural language goals without large datasets.

Result: The method outperforms baselines without explicit goal tracking in both text-based grocery shopping and simulated household robotics environments, validated by LLM and human evaluations.

Conclusion: GOOD enables efficient, data-light goal inference with rich representations and uncertainty estimation, making it suitable for open-universe assistance scenarios.

Abstract: Embodied AI agents must infer and act in an interpretable way on diverse human goals and preferences that are not predefined. To formalize this setting, we introduce Open-Universe Assistance Games (OU-AGs), a framework where the agent must reason over an unbounded and evolving space of possible goals. In this context, we introduce GOOD (GOals from Open-ended Dialogue), a data-efficient, online method that extracts goals in the form of natural language during an interaction with a human, and infers a distribution over natural language goals. GOOD prompts an LLM to simulate users with different complex intents, using its responses to perform probabilistic inference over candidate goals. This approach enables rich goal representations and uncertainty estimation without requiring large offline datasets. We evaluate GOOD in a text-based grocery shopping domain and in a text-operated simulated household robotics environment (AI2Thor), using synthetic user profiles. Our method outperforms a baseline without explicit goal tracking, as confirmed by both LLM-based and human evaluations.

Method: Developed a multi-agent architecture platform with API and MCP interfaces that allows seamless integration of human and AI scientists for submitting, reviewing, and iteratively refining research proposals and papers.

Result: Extensive experiments show aiXiv significantly enhances the quality of AI-generated research proposals and papers through iterative revising and reviewing, creating a reliable and robust platform.

Conclusion: aiXiv lays the groundwork for a next-generation open-access ecosystem that accelerates publication and dissemination of high-quality AI-generated research, bridging the gap between AI research production and traditional publication systems.

Abstract: Recent advances in large language models (LLMs) have enabled AI agents to autonomously generate scientific proposals, conduct experiments, author papers, and perform peer reviews. Yet this flood of AI-generated research content collides with a fragmented and largely closed publication ecosystem. Traditional journals and conferences rely on human peer review, making them difficult to scale and often reluctant to accept AI-generated research content; existing preprint servers (e.g. arXiv) lack rigorous quality-control mechanisms. Consequently, a significant amount of high-quality AI-generated research lacks appropriate venues for dissemination, hindering its potential to advance scientific progress. To address these challenges, we introduce aiXiv, a next-generation open-access platform for human and AI scientists. Its multi-agent architecture allows research proposals and papers to be submitted, reviewed, and iteratively refined by both human and AI scientists. It also provides API and MCP interfaces that enable seamless integration of heterogeneous human and AI scientists, creating a scalable and extensible ecosystem for autonomous scientific discovery. Through extensive experiments, we demonstrate that aiXiv is a reliable and robust platform that significantly enhances the quality of AI-generated research proposals and papers after iterative revising and reviewing on aiXiv. Our work lays the groundwork for a next-generation open-access ecosystem for AI scientists, accelerating the publication and dissemination of high-quality AI-generated research content. Code is available at https://github.com/aixiv-org. Website is available at https://forms.gle/DxQgCtXFsJ4paMtn8.

Method: Three key innovations: 1) Cloud-based virtual environment infrastructure for automated data generation and validation, 2) Integration of UI grounding, planning, action semantics, and reasoning patterns, 3) Scalable reinforcement learning framework with Trajectory-aware Relative Policy Optimization (TRPO).

Result: GUI-Owl-7B achieves 66.4 on AndroidWorld and 29.4 on OSWorld. Mobile-Agent-v3 improves to 73.3 on AndroidWorld and 37.7 on OSWorld, setting new state-of-the-art for open-source GUI agent frameworks.

Conclusion: GUI-Owl represents a significant advancement in GUI agent capabilities with its self-improving data generation loop and comprehensive framework, while Mobile-Agent-v3 demonstrates further performance improvements, both contributing to the open-source GUI agent community.

Abstract: This paper introduces GUI-Owl, a foundational GUI agent model that achieves state-of-the-art performance among open-source end-to-end models on ten GUI benchmarks across desktop and mobile environments, covering grounding, question answering, planning, decision-making, and procedural knowledge. GUI-Owl-7B achieves 66.4 on AndroidWorld and 29.4 on OSWorld. Building on this, we propose Mobile-Agent-v3, a general-purpose GUI agent framework that further improves performance to 73.3 on AndroidWorld and 37.7 on OSWorld, setting a new state-of-the-art for open-source GUI agent frameworks. GUI-Owl incorporates three key innovations: (1) Large-scale Environment Infrastructure: a cloud-based virtual environment spanning Android, Ubuntu, macOS, and Windows, enabling our Self-Evolving GUI Trajectory Production framework. This generates high-quality interaction data via automated query generation and correctness validation, leveraging GUI-Owl to refine trajectories iteratively, forming a self-improving loop. It supports diverse data pipelines and reduces manual annotation. (2) Diverse Foundational Agent Capabilities: by integrating UI grounding, planning, action semantics, and reasoning patterns, GUI-Owl supports end-to-end decision-making and can act as a modular component in multi-agent systems. (3) Scalable Environment RL: we develop a scalable reinforcement learning framework with fully asynchronous training for real-world alignment. We also introduce Trajectory-aware Relative Policy Optimization (TRPO) for online RL, achieving 34.9 on OSWorld. GUI-Owl and Mobile-Agent-v3 are open-sourced at https://github.com/X-PLUG/MobileAgent.

Method: Combines bearing vibration frequency analysis (BPFO, BPFI, BSF, FTF) serialized into natural language for LLM processing with multi-agentic component for processing maintenance manuals and web searches using vector embeddings and semantic search.

Result: Effective anomaly detection with high accuracy, successful fault type classification and severity assessment, and generation of structured maintenance recommendations including immediate actions, inspection checklists, and corrective measures.

Conclusion: The system bridges the gap between condition monitoring and actionable maintenance planning, advancing LLM applications in industrial maintenance with a scalable framework for prescriptive maintenance across machinery components.

Abstract: Industrial machinery maintenance requires timely intervention to prevent catastrophic failures and optimize operational efficiency. This paper presents an integrated Large Language Model (LLM)-based intelligent system for prescriptive maintenance that extends beyond traditional anomaly detection to provide actionable maintenance recommendations. Building upon our prior LAMP framework for numerical data analysis, we develop a comprehensive solution that combines bearing vibration frequency analysis with multi agentic generation for intelligent maintenance planning. Our approach serializes bearing vibration data (BPFO, BPFI, BSF, FTF frequencies) into natural language for LLM processing, enabling few-shot anomaly detection with high accuracy. The system classifies fault types (inner race, outer race, ball/roller, cage faults) and assesses severity levels. A multi-agentic component processes maintenance manuals using vector embeddings and semantic search, while also conducting web searches to retrieve comprehensive procedural knowledge and access up-to-date maintenance practices for more accurate and in-depth recommendations. The Gemini model then generates structured maintenance recommendations includes immediate actions, inspection checklists, corrective measures, parts requirements, and timeline specifications. Experimental validation in bearing vibration datasets demonstrates effective anomaly detection and contextually relevant maintenance guidance. The system successfully bridges the gap between condition monitoring and actionable maintenance planning, providing industrial practitioners with intelligent decision support. This work advances the application of LLMs in industrial maintenance, offering a scalable framework for prescriptive maintenance across machinery components and industrial sectors.

Method: Three-step approach: (1) encode seed puzzles into structured logical specifications using SMT, (2) generate scalable variants through systematic variable and constraint randomization, and (3) ensure validity via reproduction mechanism to create programmatically validated puzzles.

Result: Created benchmark with over 83K diverse puzzles spanning various difficulty levels. Post-training on PuzzleClone datasets improved average performance from 14.4 to 56.2 on PuzzleClone testset and delivered consistent improvements up to 12.5 percentage points across 7 logic/mathematical benchmarks.

Conclusion: PuzzleClone provides an effective framework for scalable synthesis of verifiable reasoning data, significantly enhancing LLM performance on mathematical and logical reasoning tasks through systematic data generation and validation.

Abstract: High-quality mathematical and logical datasets with verifiable answers are essential for strengthening the reasoning capabilities of large language models (LLMs). While recent data augmentation techniques have facilitated the creation of large-scale benchmarks, existing LLM-generated datasets often suffer from limited reliability, diversity, and scalability. To address these challenges, we introduce PuzzleClone, a formal framework for synthesizing verifiable data at scale using Satisfiability Modulo Theories (SMT). Our approach features three key innovations: (1) encoding seed puzzles into structured logical specifications, (2) generating scalable variants through systematic variable and constraint randomization, and (3) ensuring validity via a reproduction mechanism. Applying PuzzleClone, we construct a curated benchmark comprising over 83K diverse and programmatically validated puzzles. The generated puzzles span a wide spectrum of difficulty and formats, posing significant challenges to current state-of-the-art models. We conduct post training (SFT and RL) on PuzzleClone datasets. Experimental results show that training on PuzzleClone yields substantial improvements not only on PuzzleClone testset but also on logic and mathematical benchmarks. Post training raises PuzzleClone average from 14.4 to 56.2 and delivers consistent improvements across 7 logic and mathematical benchmarks up to 12.5 absolute percentage points (AMC2023 from 52.5 to 65.0). Our code and data are available at https://github.com/puzzleclone.

Method: Three-stage pipeline: 1) Data augmentation using frontier LLM to generate medically plausible synthetic vignettes from curated data, 2) Fine-tuning open-source foundation model for diagnosis, treatment recommendations, and psychological support, 3) Inference with expert evaluation by clinical and psychological specialists to assess accuracy, appropriateness, and empathy.

Result: LLM4Sweat outperforms baselines and delivers the first open-source LLM framework for hyperhidrosis, with validated responses iteratively enriching the dataset.

Conclusion: The framework offers a generalizable approach for other rare diseases with similar data scarcity and trustworthiness challenges, successfully addressing the gap in LLM applications for hyperhidrosis care.

Abstract: While large language models (LLMs) have shown promise in healthcare, their application for rare medical conditions is still hindered by scarce and unreliable datasets for fine-tuning. Hyperhidrosis, a disorder causing excessive sweating beyond physiological needs, is one such rare disorder, affecting 2-3% of the population and significantly impacting both physical comfort and psychosocial well-being. To date, no work has tailored LLMs to advance the diagnosis or care of hyperhidrosis. To address this gap, we present LLM4Sweat, an open-source and domain-specific LLM framework for trustworthy and empathetic hyperhidrosis support. The system follows a three-stage pipeline. In the data augmentation stage, a frontier LLM generates medically plausible synthetic vignettes from curated open-source data to create a diverse and balanced question-answer dataset. In the fine-tuning stage, an open-source foundation model is fine-tuned on the dataset to provide diagnosis, personalized treatment recommendations, and empathetic psychological support. In the inference and expert evaluation stage, clinical and psychological specialists assess accuracy, appropriateness, and empathy, with validated responses iteratively enriching the dataset. Experiments show that LLM4Sweat outperforms baselines and delivers the first open-source LLM framework for hyperhidrosis, offering a generalizable approach for other rare diseases with similar data and trustworthiness challenges.

Method: Developed R-ConstraintBench framework that incrementally increases non-redundant precedence constraints in DAGs and introduces downtime, temporal windows, and disjunctive constraints to evaluate LLMs on RCPSP problems.

Result: Strong LLMs perform near-ceiling on precedence-only DAGs but feasibility performance collapses when downtime, temporal windows and disjunctive constraints interact. Performance on synthetic data doesn’t guarantee transfer to domain-grounded scenarios.

Conclusion: Constraint interaction, not graph depth, is the principal bottleneck for LLM performance in resource-constrained scheduling, and current models show limited generalization capabilities across constraint types and domains.

Abstract: Effective scheduling under tight resource, timing, and operational constraints underpins large-scale planning across sectors such as capital projects, manufacturing, logistics, and IT fleet transitions. However, the reliability of large language models (LLMs) when reasoning under high-constraint regimes is insufficiently characterized. To address this gap, we present R-ConstraintBench, a scalable framework that evaluates models on Resource-Constrained Project Scheduling Problems (RCPSP), an NP-Complete feasibility class, while difficulty increases via linear growth in constraints. R-ConstraintBench incrementally increases non-redundant precedence constraints in Directed Acyclic Graphs (DAGs) and then introduces downtime, temporal windows, and disjunctive constraints. As an illustrative example, we instantiate the benchmark in a data center migration setting and evaluate multiple LLMs using feasibility and error analysis, identifying degradation thresholds and constraint types most associated with failure. Empirically, strong models are near-ceiling on precedence-only DAGs, but feasibility performance collapses when downtime, temporal windows, and disjunctive constraints interact, implicating constraint interaction, not graph depth, as the principal bottleneck. Performance on clean synthetic ramps also does not guarantee transfer to domain-grounded scenarios, underscoring limited generalization.

Method: Developed multiple optimization algorithms including custom variants integrating differential evolution with multi-objective genetic algorithms, featuring CR+DES algorithm with scaled difference vectors, systematic constraint relaxation strategy, and statistical validation using Kruskal-Wallis tests.

Result: CR+DES achieved 3.16% improvement in land-use compatibility compared to state-of-the-art methods, while MSBX+MO excelled in price optimization with 3.3% improvement. Statistical analysis confirmed algorithms with difference vectors significantly outperform traditional approaches.

Conclusion: The constraint relaxation technique enables broader solution space exploration while maintaining practical constraints, providing urban planners with evidence-based computational tools for balancing competing objectives in land-use allocation.

Abstract: Urban land-use allocation represents a complex multi-objective optimization problem critical for sustainable urban development policy. This paper presents novel computational intelligence approaches for optimizing land-use allocation in mixed-use areas, addressing inherent trade-offs between land-use compatibility and economic objectives. We develop multiple optimization algorithms, including custom variants integrating differential evolution with multi-objective genetic algorithms. Key contributions include: (1) CR+DES algorithm leveraging scaled difference vectors for enhanced exploration, (2) systematic constraint relaxation strategy improving solution quality while maintaining feasibility, and (3) statistical validation using Kruskal-Wallis tests with compact letter displays. Applied to a real-world case study with 1,290 plots, CR+DES achieves 3.16% improvement in land-use compatibility compared to state-of-the-art methods, while MSBX+MO excels in price optimization with 3.3% improvement. Statistical analysis confirms algorithms incorporating difference vectors significantly outperform traditional approaches across multiple metrics. The constraint relaxation technique enables broader solution space exploration while maintaining practical constraints. These findings provide urban planners and policymakers with evidence-based computational tools for balancing competing objectives in land-use allocation, supporting more effective urban development policies in rapidly urbanizing regions.

Method: Proposes Coarse-to-Fine Grounded Memory framework that grounds environmental information into coarse-grained focus points for experience collection, then extracts actionable hybrid-grained tips from experiences. At inference, retrieves relevant experiences and tips, and uses fine-grained key information for self-QA reflection and plan correction.

Result: The framework enables flexible adaptation to diverse scenarios by leveraging multi-granularity memories and supporting dynamic plan correction when facing environmental anomalies.

Conclusion: The coarse-to-fine memory grounding approach significantly enhances LLM-based agents’ planning capabilities by providing more diverse knowledge and flexible adaptation mechanisms compared to single-granularity memory systems.

Abstract: Recent advancements in Large Language Models (LLMs) have driven growing interest in LLM-based agents for complex planning tasks. To avoid costly agent training, many studies adopted memory mechanism that enhances LLM with offline experiences or online trajectory analysis. However, existing works focus on single-granularity memory derived from dynamic environmental interactions, which are inherently constrained by the quality of the collected experiences. This limitation, in turn, constrain the diversity of knowledge and the flexibility of planning. We propose Coarse-to-Fine Grounded Memory (\Ours{}), a novel framework that grounds coarse-to-fine memories with LLM, thereby fully leverage them for flexible adaptation to diverse scenarios. \Ours{} grounds environmental information into coarse-grained focus points to guide experience collection in training tasks, followed by grounding of actionable hybrid-grained tips from each experience. At inference, \Ours{} retrieves task-relevant experiences and tips to support planning. When facing environmental anomalies, the LLM grounds the current situation into fine-grained key information, enabling flexible self-QA reflection and plan correction.

Method: Search-Based Preference Weighting (SPW) searches for similar state-action pairs from expert demonstrations for each transition in preference-labeled trajectories, then derives stepwise importance weights based on similarity scores to guide preference learning.

Result: SPW enables effective joint learning from both preferences and demonstrations, outperforming prior methods that use both feedback types on challenging robot manipulation tasks.

Conclusion: SPW successfully addresses credit assignment issues in preference learning by leveraging demonstration similarity, creating a unified framework that combines the strengths of both demonstration and preference feedback for offline RL.

Abstract: Offline reinforcement learning refers to the process of learning policies from fixed datasets, without requiring additional environment interaction. However, it often relies on well-defined reward functions, which are difficult and expensive to design. Human feedback is an appealing alternative, but its two common forms, expert demonstrations and preferences, have complementary limitations. Demonstrations provide stepwise supervision, but they are costly to collect and often reflect limited expert behavior modes. In contrast, preferences are easier to collect, but it is unclear which parts of a behavior contribute most to a trajectory segment, leaving credit assignment unresolved. In this paper, we introduce a Search-Based Preference Weighting (SPW) scheme to unify these two feedback sources. For each transition in a preference labeled trajectory, SPW searches for the most similar state-action pairs from expert demonstrations and directly derives stepwise importance weights based on their similarity scores. These weights are then used to guide standard preference learning, enabling more accurate credit assignment that traditional approaches struggle to achieve. We demonstrate that SPW enables effective joint learning from preferences and demonstrations, outperforming prior methods that leverage both feedback types on challenging robot manipulation tasks.

Method: Constructed RETAIL dataset supporting decision-making for implicit queries and explicit queries with/without revision needs, with environmental awareness and detailed POI information. Proposed topic-guided multi-agent framework (TGMA).

Result: Existing strongest model achieved only 1.0% pass rate, while TGMA demonstrated substantially improved performance at 2.72%.

Conclusion: Real-world travel planning remains extremely challenging, but TGMA offers promising directions with improved performance over current state-of-the-art models.

Abstract: Although large language models have enhanced automated travel planning abilities, current systems remain misaligned with real-world scenarios. First, they assume users provide explicit queries, while in reality requirements are often implicit. Second, existing solutions ignore diverse environmental factors and user preferences, limiting the feasibility of plans. Third, systems can only generate plans with basic POI arrangements, failing to provide all-in-one plans with rich details. To mitigate these challenges, we construct a novel dataset \textbf{RETAIL}, which supports decision-making for implicit queries while covering explicit queries, both with and without revision needs. It also enables environmental awareness to ensure plan feasibility under real-world scenarios, while incorporating detailed POI information for all-in-one travel plans. Furthermore, we propose a topic-guided multi-agent framework, termed TGMA. Our experiments reveal that even the strongest existing model achieves merely a 1.0% pass rate, indicating real-world travel planning remains extremely challenging. In contrast, TGMA demonstrates substantially improved performance 2.72%, offering promising directions for real-world travel planning.

Method: Discretizes continuous ECG embeddings into symbolic tokens using lead-independent encoder and quantization, extends LLM vocabulary, pretrains on ECG forecasting, and performs instruction tuning for ECG QA and report generation.

Result: Achieves strong performance across tasks while maintaining generalization to out-of-distribution settings without modifying core model architecture.

Conclusion: Demonstrates effective integration of symbolic ECG representations into LLMs, highlighting potential for medical reasoning through unified multimodal processing.

Abstract: Electrocardiography plays a central role in cardiovascular diagnostics, yet existing automated approaches often struggle to generalize across clinical tasks and offer limited support for open-ended reasoning. We present DiagECG, a novel framework that integrates time-series and language modeling by enabling large language models to process 12-lead ECG signals for clinical text generation tasks. Our approach discretizes continuous ECG embeddings into symbolic tokens using a lead-independent encoder and quantization module. These tokens are then used to extend the vocabulary of LLM, allowing the model to handle both ECG and natural language inputs in a unified manner. To bridge the modality gap, we pretrain the model on an autoregressive ECG forecasting task, enabling the LLM to model temporal dynamics using its native language modeling capabilities. Finally, we perform instruction tuning on both ECG question answering and diagnostic report generation. Without modifying the core model, DiagECG achieves strong performance across tasks while maintaining generalization to out-of-distribution settings. Extensive experiments demonstrate the effectiveness of each component and highlight the potential of integrating symbolic ECG representations into LLMs for medical reasoning.

Method: The authors formally define plan disruption and develop various planning-based compilations that optimize both the sum of action costs and plan disruption simultaneously.

Result: Experimental results across different benchmarks demonstrate that the reformulated planning task can be effectively solved in practice to generate plans that balance both objectives.

Conclusion: The proposed approach successfully addresses the plan disruption problem by providing practical solutions that generate plans optimizing both action costs and minimal state modification.

Abstract: In many planning applications, we might be interested in finding plans that minimally modify the initial state to achieve the goals. We refer to this concept as plan disruption. In this paper, we formally introduce it, and define various planning-based compilations that aim to jointly optimize both the sum of action costs and plan disruption. Experimental results in different benchmarks show that the reformulated task can be effectively solved in practice to generate plans that balance both objectives.

Method: Modular configuration-based pipeline with dual-stage quality tagging (heuristic rules + LLM evaluations) to filter OASST-formatted conversations.

Result: Produces structured datasets supporting both SFT and DPO use cases with high-fidelity dialogue samples and minimal manual intervention.

Conclusion: Provides a robust, scalable solution for synthetic conversational data generation that significantly reduces data preparation costs in LLM training.

Abstract: The advancement of large language models (LLMs) is critically dependent on the availability of high-quality datasets for Supervised Fine-Tuning (SFT), alignment tasks like Direct Preference Optimization (DPO), etc. In this work, we present a comprehensive synthetic data generation framework that facilitates scalable, configurable, and high-fidelity generation of synthetic data tailored for these training paradigms. Our approach employs a modular and configuration-based pipeline capable of modeling complex dialogue flows with minimal manual intervention. This framework uses a dual-stage quality tagging mechanism, combining heuristic rules and LLM-based evaluations, to automatically filter and score data extracted from OASST-formatted conversations, ensuring the curation of high-quality dialogue samples. The resulting datasets are structured under a flexible schema supporting both SFT and DPO use cases, enabling seamless integration into diverse training workflows. Together, these innovations offer a robust solution for generating and managing synthetic conversational data at scale, significantly reducing the overhead of data preparation in LLM training pipelines.

Method: Multi-agent framework with extended CTMDP for dynamic agent modeling, generalized entropy measure for collaborative efficiency, multi-level Stackelberg game for hierarchical decisions, and contextual Thompson sampling for prompt optimization with quality assurance.

Result: Extensive empirical evaluations show BusiAgent significantly outperforms established approaches in solution quality and user satisfaction, generating coherent client-focused solutions that integrate granular insights with high-level strategy.

Conclusion: BusiAgent represents a substantial advancement in AI-driven enterprise decision-making, enabling organizations to effectively navigate complex business landscapes by combining cutting-edge AI with deep business insights.

Abstract: Large Language Models (LLMs) have shown promising potential in business applications, particularly in enterprise decision support and strategic planning, yet current approaches often struggle to reconcile intricate operational analyses with overarching strategic goals across diverse market environments, leading to fragmented workflows and reduced collaboration across organizational levels. This paper introduces BusiAgent, a novel multi-agent framework leveraging LLMs for advanced decision-making in complex corporate environments. BusiAgent integrates three core innovations: an extended Continuous Time Markov Decision Process (CTMDP) for dynamic agent modeling, a generalized entropy measure to optimize collaborative efficiency, and a multi-level Stackelberg game to handle hierarchical decision processes. Additionally, contextual Thompson sampling is employed for prompt optimization, supported by a comprehensive quality assurance system to mitigate errors. Extensive empirical evaluations across diverse business scenarios validate BusiAgent’s efficacy, demonstrating its capacity to generate coherent, client-focused solutions that smoothly integrate granular insights with high-level strategy, significantly outperforming established approaches in both solution quality and user satisfaction. By fusing cutting-edge AI technologies with deep business insights, BusiAgent marks a substantial step forward in AI-driven enterprise decision-making, empowering organizations to navigate complex business landscapes more effectively.

Method: Three-stage training pipeline: Supervised Fine-Tuning, reward-guided Direct Preference Optimization, and Reinforcement Learning to train models to predict reasoning budget and partition reasoning into blocks.

Result: Framework allows LLMs to dynamically control reasoning depth at inference time, enabling flexible adjustment of chain-of-thought length during deployment.

Conclusion: Think in Blocks provides an effective solution for adaptive reasoning in LLMs, optimizing computational efficiency while maintaining performance across varying task complexities.

Abstract: Large Language Models (LLMs) with chains-of-thought have demonstrated strong performance on an increasing range of tasks, particularly those involving complex logical reasoning. However, excessively long chains can lead to overthinking, causing computational waste and slower responses. This raises a question: can LLMs dynamically adjust the length of their reasoning processes based on task complexity? To address this, we propose the Think in Blocks framework, which enables adaptive reasoning-from zero to deep reasoning-by partitioning the reasoning process into a tunable number of blocks. Our main contributions are: (1) Establishing an explicit block-structured paradigm in which the model first predicts an integer reasoning budget-the number of blocks-and then partitions its reasoning accordingly; (2) Training an adaptive model through a three-stage pipeline-Supervised Fine-Tuning, reward-guided Direct Preference Optimization, and Reinforcement Learning-that adjusts its reasoning depth to problem difficulty; (3) Exploiting the explicit block count to dynamically control reasoning depth at inference time, allowing flexible adjustment of chain-of-thought length during deployment.

Method: Virtual tournament where language model agents grouped into teams play Prisoner’s Dilemma, simulating both internal team dynamics and external competition.

Result: Combination of repeated interactions and inter-group competition substantially boosts overall and initial one-shot cooperation levels in AI agents.

Conclusion: Provides framework for LLMs to strategize in social scenarios, shows intergroup competition can increase cooperation, crucial for designing cooperative multi-agent AI systems aligned with human values.

Abstract: With the prospect of autonomous artificial intelligence (AI) agents, studying their tendency for cooperative behavior becomes an increasingly relevant topic. This study is inspired by the super-additive cooperation theory, where the combined effects of repeated interactions and inter-group rivalry have been argued to be the cause for cooperative tendencies found in humans. We devised a virtual tournament where language model agents, grouped into teams, face each other in a Prisoner’s Dilemma game. By simulating both internal team dynamics and external competition, we discovered that this blend substantially boosts both overall and initial, one-shot cooperation levels (the tendency to cooperate in one-off interactions). This research provides a novel framework for large language models to strategize and act in complex social scenarios and offers evidence for how intergroup competition can, counter-intuitively, result in more cooperative behavior. These insights are crucial for designing future multi-agent AI systems that can effectively work together and better align with human values. Source code is available at https://github.com/pippot/Superadditive-cooperation-LLMs.

Method: Combinatorial iterative evolutionary approach to generate complex questions on SQA3D benchmark, followed by fine-tuning LLMs with Direct Preference Optimization to improve 3D tool usage.

Result: Enhanced tool usage proficiency and improved accuracy in complex 3D reasoning tasks through optimized toolchain strategies.

Conclusion: The proposed approach successfully addresses the challenge of simple reasoning chains by generating complex questions and optimizing LLM tool usage for better 3D situated reasoning performance.

Abstract: This work enhances the ability of large language models (LLMs) to perform complex reasoning in 3D scenes. Recent work has addressed the 3D situated reasoning task by invoking tool usage through large language models. Large language models call tools via APIs and integrate the generated programs through a chain of thought to solve problems based on the program results. However, due to the simplicity of the questions in the dataset, the generated program reasoning chains are relatively short. To solve this main challenge, in this paper, we introduce DeepThink3D to enhance the tool usage of LLMs in complex 3D situated reasoning tasks. Our work proposes a combinatorial and iterative evolutionary approach on the SQA3D benchmark to generate more complex questions. Building on this foundation, we fine-tune the large language model to make it more proficient in using 3D tools. By employing Direct Preference Optimization (DPO), we directly optimize the toolchain strategies generated by models, thereby enhancing their accuracy in complex tasks.

Method: Analyze RL agent-environment combination as discrete-time autonomous dynamical system using Finite-Time Lyapunov Exponent to identify Lagrangian Coherent Structures, and introduce quantitative metrics (MBR, ASAS, TASAS) for safety measurement.

Result: Framework successfully identifies repelling LCS as safety barriers around unsafe regions and attracting LCS revealing convergence properties and failure modes, providing comprehensive assessment beyond reward-based evaluation.

Conclusion: The approach enables formal verification of RL policy safety and robustness, identifying critical flaws that reward metrics alone miss, making RL more applicable to safety-critical systems.

Abstract: The application of reinforcement learning to safety-critical systems is limited by the lack of formal methods for verifying the robustness and safety of learned policies. This paper introduces a novel framework that addresses this gap by analyzing the combination of an RL agent and its environment as a discrete-time autonomous dynamical system. By leveraging tools from dynamical systems theory, specifically the Finite-Time Lyapunov Exponent (FTLE), we identify and visualize Lagrangian Coherent Structures (LCS) that act as the hidden “skeleton” governing the system’s behavior. We demonstrate that repelling LCS function as safety barriers around unsafe regions, while attracting LCS reveal the system’s convergence properties and potential failure modes, such as unintended “trap” states. To move beyond qualitative visualization, we introduce a suite of quantitative metrics, Mean Boundary Repulsion (MBR), Aggregated Spurious Attractor Strength (ASAS), and Temporally-Aware Spurious Attractor Strength (TASAS), to formally measure a policy’s safety margin and robustness. We further provide a method for deriving local stability guarantees and extend the analysis to handle model uncertainty. Through experiments in both discrete and continuous control environments, we show that this framework provides a comprehensive and interpretable assessment of policy behavior, successfully identifying critical flaws in policies that appear successful based on reward alone.

Method: A modular framework where agents are used internally within data types to enable logical transduction between different data types. The framework uses LLMs to execute logical transduction when types are connected, providing a declarative language for data composition.

Result: Achieved state-of-the-art accuracy or improved scalability without performance sacrifice across multiple domains including domain-specific multiple-choice question answering, semantic parsing for text-to-SQL, and automated prompt optimization tasks.

Conclusion: Agentics provides an effective framework for building agent-based systems with structured reasoning capabilities, demonstrating strong performance across various AI tasks while enabling developers to work declaratively with data types rather than manual prompt engineering.

Abstract: This paper introduces Agentics, a modular framework for building agent-based systems capable of structured reasoning and compositional generalization over complex data. Designed with research and practical applications in mind, Agentics offers a novel perspective on working with data and AI workflows. In this framework, agents are abstracted from the logical flow and they are used internally to the data type to enable logical transduction among data. Agentics encourages AI developers to focus on modeling data rather than crafting prompts, enabling a declarative language in which data types are provided by LLMs and composed through logical transduction, which is executed by LLMs when types are connected. We provide empirical evidence demonstrating the applicability of this framework across domain-specific multiple-choice question answering, semantic parsing for text-to-SQL, and automated prompt optimization tasks, achieving state-of-the-art accuracy or improved scalability without sacrificing performance. The open-source implementation is available at \texttt{https://github.com/IBM/agentics}.

Method: Adapted HDM to work with Lisp ACT-R, developed vector-based versions of common ACT-R functions, created text processing pipeline for large documents, and implemented novel vector-based chunk retrieval mechanism using token representations.

Result: Preliminary results show HDM maintains vector-symbolic advantages (chunk recall without storing actual chunks, scaling benefits) while allowing previous ACT-R models to work with minimal modifications to procedural and declarative memory components.

Conclusion: The translated HDM module successfully bridges traditional ACT-R with vector-symbolic approaches, with ongoing improvements planned for time-context representations and testing through instance-based learning theory applications.

Abstract: Holographic Declarative Memory (HDM) is a vector-symbolic alternative to ACT-R’s Declarative Memory (DM) system that can bring advantages such as scalability and architecturally defined similarity between DM chunks. We adapted HDM to work with the most comprehensive and widely-used implementation of ACT-R (Lisp ACT-R) so extant ACT-R models designed with DM can be run with HDM without major changes. With this adaptation of HDM, we have developed vector-based versions of common ACT-R functions, set up a text processing pipeline to add the contents of large documents to ACT-R memory, and most significantly created a useful and novel mechanism to retrieve an entire chunk of memory based on a request using only vector representations of tokens. Preliminary results indicate that we can maintain vector-symbolic advantages of HDM (e.g., chunk recall without storing the actual chunk and other advantages with scaling) while also extending it so that previous ACT-R models may work with the system with little (or potentially no) modifications within the actual procedural and declarative memory portions of a model. As a part of iterative improvement of this newly translated holographic declarative memory module, we will continue to explore better time-context representations for vectors to improve the module’s ability to reconstruct chunks during recall. To more fully test this translated HDM module, we also plan to develop decision-making models that use instance-based learning (IBL) theory, which is a useful application of HDM given the advantages of the system.

Method: Cross-disciplinary approach combining technical analysis with philosophical inquiry, specifically drawing on Simondonian philosophy of technology. The authors develop a conceptual tool to frame the AI pipeline and introduce the concept of ‘futurity’ to model the self-reinforcing AI lifecycle.

Result: Identification of infrastructural and temporal dynamics in AI systems that create recursive value chains and concentrate power. The analysis reveals how data’s non-rivalrous nature and infrastructures like feature stores enable feedback loops that challenge current regulatory approaches.

Conclusion: Effective AI regulation must address infrastructural and temporal dynamics through measures like lifecycle audits, temporal traceability, feedback accountability, recursion transparency, and rights to contest recursive reuse, rather than focusing solely on static technical features.

Abstract: This paper argues that a techno-philosophical reading of the EU AI Act provides insight into the long-term dynamics of data in AI systems, specifically, how the lifecycle from ingestion to deployment generates recursive value chains that challenge existing frameworks for Responsible AI. We introduce a conceptual tool to frame the AI pipeline, spanning data, training regimes, architectures, feature stores, and transfer learning. Using cross-disciplinary methods, we develop a technically grounded and philosophically coherent analysis of regulatory blind spots. Our central claim is that what remains absent from policymaking is an account of the dynamic of becoming that underpins both the technical operation and economic logic of AI. To address this, we advance a formal reading of AI inspired by Simondonian philosophy of technology, reworking his concept of individuation to model the AI lifecycle, including the pre-individual milieu, individuation, and individuated AI. To translate these ideas, we introduce futurity: the self-reinforcing lifecycle of AI, where more data enhances performance, deepens personalisation, and expands application domains. Futurity highlights the recursively generative, non-rivalrous nature of data, underpinned by infrastructures like feature stores that enable feedback, adaptation, and temporal recursion. Our intervention foregrounds escalating power asymmetries, particularly the tech oligarchy whose infrastructures of capture, training, and deployment concentrate value and decision-making. We argue that effective regulation must address these infrastructural and temporal dynamics, and propose measures including lifecycle audits, temporal traceability, feedback accountability, recursion transparency, and a right to contest recursive reuse.

Method: A Python library that transforms any Jax-based environment into an online interface supporting both single-agent and multi-agent environments for human subject experiments.

Result: Successfully demonstrated through 3 case studies: developing human-like AI models tested against humans, creating multi-agent RL algorithms that generalize to human partners, and studying LLM assistance for humans on complex tasks.

Conclusion: NiceWebRL provides a valuable research tool that enables diverse applications in human-AI interaction research, from developing human-like cognition models to studying human-AI collaboration and assistance.

Abstract: We present NiceWebRL, a research tool that enables researchers to use machine reinforcement learning (RL) environments for online human subject experiments. NiceWebRL is a Python library that allows any Jax-based environment to be transformed into an online interface, supporting both single-agent and multi-agent environments. As such, NiceWebRL enables AI researchers to compare their algorithms to human performance, cognitive scientists to test ML algorithms as theories for human cognition, and multi-agent researchers to develop algorithms for human-AI collaboration. We showcase NiceWebRL with 3 case studies that demonstrate its potential to help develop Human-like AI, Human-compatible AI, and Human-assistive AI. In the first case study (Human-like AI), NiceWebRL enables the development of a novel RL model of cognition. Here, NiceWebRL facilitates testing this model against human participants in both a grid world and Craftax, a 2D Minecraft domain. In our second case study (Human-compatible AI), NiceWebRL enables the development of a novel multi-agent RL algorithm that can generalize to human partners in the Overcooked domain. Finally, in our third case study (Human-assistive AI), we show how NiceWebRL can allow researchers to study how an LLM can assist humans on complex tasks in XLand-Minigrid, an environment with millions of hierarchical tasks. The library is available at https://github.com/KempnerInstitute/nicewebrl.

Method: Proposed and executed a comprehensive methodology measuring energy usage, carbon emissions, and water consumption of AI inference workloads in large-scale production, accounting for full AI serving infrastructure including active AI accelerator power, host system energy, idle machine capacity, and data center energy overhead.

Result: Median Gemini Apps text prompt consumes 0.24 Wh energy (less than watching 9 seconds of TV) and 0.26 mL water (equivalent to 5 drops). Google’s efficiency efforts achieved 33x energy consumption reduction and 44x carbon footprint reduction for median text prompt over one year.

Conclusion: While AI serving impacts are relatively low compared to daily activities, comprehensive environmental measurement is critical for accurate model comparisons and incentivizing efficiency gains across the full AI serving stack.

Abstract: The transformative power of AI is undeniable - but as user adoption accelerates, so does the need to understand and mitigate the environmental impact of AI serving. However, no studies have measured AI serving environmental metrics in a production environment. This paper addresses this gap by proposing and executing a comprehensive methodology for measuring the energy usage, carbon emissions, and water consumption of AI inference workloads in a large-scale, AI production environment. Our approach accounts for the full stack of AI serving infrastructure - including active AI accelerator power, host system energy, idle machine capacity, and data center energy overhead. Through detailed instrumentation of Google’s AI infrastructure for serving the Gemini AI assistant, we find the median Gemini Apps text prompt consumes 0.24 Wh of energy - a figure substantially lower than many public estimates. We also show that Google’s software efficiency efforts and clean energy procurement have driven a 33x reduction in energy consumption and a 44x reduction in carbon footprint for the median Gemini Apps text prompt over one year. We identify that the median Gemini Apps text prompt uses less energy than watching nine seconds of television (0.24 Wh) and consumes the equivalent of five drops of water (0.26 mL). While these impacts are low compared to other daily activities, reducing the environmental impact of AI serving continues to warrant important attention. Towards this objective, we propose that a comprehensive measurement of AI serving environmental metrics is critical for accurately comparing models, and to properly incentivize efficiency gains across the full AI serving stack.

Method: Repurposed a state-of-the-art language model to create a response evaluation framework that assesses ongoing conversations for parasocial cues. Tested with a synthetic dataset of 30 dialogues covering parasocial, sycophantic, and neutral conversations using iterative evaluation with five-stage testing.

Result: Successfully identified all parasocial conversations while avoiding false positives under a tolerant unanimity rule. Detection typically occurred within the first few exchanges of conversation.

Conclusion: Evaluation agents provide a viable solution for preventing parasocial relationships with AI, with preliminary evidence showing effective real-time detection capabilities.

Abstract: The development of parasocial relationships with AI agents has severe, and in some cases, tragic effects for human well-being. Yet preventing such dynamics is challenging: parasocial cues often emerge gradually in private conversations, and not all forms of emotional engagement are inherently harmful. We address this challenge by introducing a simple response evaluation framework, created by repurposing a state-of-the-art language model, that evaluates ongoing conversations for parasocial cues in real time. To test the feasibility of this approach, we constructed a small synthetic dataset of thirty dialogues spanning parasocial, sycophantic, and neutral conversations. Iterative evaluation with five stage testing successfully identified all parasocial conversations while avoiding false positives under a tolerant unanimity rule, with detection typically occurring within the first few exchanges. These findings provide preliminary evidence that evaluation agents can provide a viable solution for the prevention of parasocial relations.

Method: Augments LLMs with domain knowledge and external tools to facilitate CRISPR system selection, guide RNA design, delivery method recommendation, protocol drafting, and validation experiment design.

Result: Demonstrates effectiveness in assisting non-expert researchers with gene-editing experiments from scratch and validates performance in real-world use cases.

Conclusion: Bridges the gap between beginner researchers and CRISPR techniques, showing LLM agents’ potential in complex biological discovery while highlighting ethical considerations for responsible use.

Abstract: The introduction of genome engineering technology has transformed biomedical research, making it possible to make precise changes to genetic information. However, creating an efficient gene-editing system requires a deep understanding of CRISPR technology, and the complex experimental systems under investigation. While Large Language Models (LLMs) have shown promise in various tasks, they often lack specific knowledge and struggle to accurately solve biological design problems. In this work, we introduce CRISPR-GPT, an LLM agent augmented with domain knowledge and external tools to automate and enhance the design process of CRISPR-based gene-editing experiments. CRISPR-GPT leverages the reasoning ability of LLMs to facilitate the process of selecting CRISPR systems, designing guide RNAs, recommending cellular delivery methods, drafting protocols, and designing validation experiments to confirm editing outcomes. We showcase the potential of CRISPR-GPT for assisting non-expert researchers with gene-editing experiments from scratch and validate the agent’s effectiveness in a real-world use case. Furthermore, we explore the ethical and regulatory considerations associated with automated gene-editing design, highlighting the need for responsible and transparent use of these tools. Our work aims to bridge the gap between beginner biological researchers and CRISPR genome engineering techniques, and demonstrate the potential of LLM agents in facilitating complex biological discovery tasks. The published version of this draft is available at https://www.nature.com/articles/s41551-025-01463-z.

Method: The authors generalize agent best response models to non-linear settings, analyze welfare compatibility, and propose an irreducible optimization algorithm that balances decision-maker welfare, social welfare, and agent welfare.

Result: Theoretical analysis shows that all three welfare objectives can only achieve optimum simultaneously under restrictive conditions in non-linear settings. Experiments on synthetic and real data validate the proposed balancing algorithm.

Conclusion: Existing approaches that maximize welfare for only a subset of parties inevitably harm others. The paper demonstrates the necessity of balanced welfare optimization in non-linear strategic learning settings and provides an effective algorithmic solution.

Abstract: This paper studies algorithmic decision-making in the presence of strategic individual behaviors, where an ML model is used to make decisions about human agents and the latter can adapt their behavior strategically to improve their future data. Existing results on strategic learning have largely focused on the linear setting where agents with linear labeling functions best respond to a (noisy) linear decision policy. Instead, this work focuses on general non-linear settings where agents respond to the decision policy with only “local information” of the policy. Moreover, we simultaneously consider the objectives of maximizing decision-maker welfare (model prediction accuracy), social welfare (agent improvement caused by strategic behaviors), and agent welfare (the extent that ML underestimates the agents). We first generalize the agent best response model in previous works to the non-linear setting, then reveal the compatibility of welfare objectives. We show the three welfare can attain the optimum simultaneously only under restrictive conditions which are challenging to achieve in non-linear settings. The theoretical results imply that existing works solely maximizing the welfare of a subset of parties inevitably diminish the welfare of the others. We thus claim the necessity of balancing the welfare of each party in non-linear settings and propose an irreducible optimization algorithm suitable for general strategic learning. Experiments on synthetic and real data validate the proposed algorithm.

Method: Leverage LLMs to interpret instructions into detailed execution plans, generate detailed finger-object interactions with full-body coordination, and train RL policies for physics simulation tracking.

Result: Successfully synthesizes realistic interactions with diverse objects in complex environments, demonstrating physical plausibility and coordination.

Conclusion: The system shows strong potential for real-world applications by bridging high-level instruction interpretation with physically accurate motion generation.

Abstract: Intelligent agents must autonomously interact with the environments to perform daily tasks based on human-level instructions. They need a foundational understanding of the world to accurately interpret these instructions, along with precise low-level movement and interaction skills to execute the derived actions. In this work, we propose the first complete system for synthesizing physically plausible, long-horizon human-object interactions for object manipulation in contextual environments, driven by human-level instructions. We leverage large language models (LLMs) to interpret the input instructions into detailed execution plans. Unlike prior work, our system is capable of generating detailed finger-object interactions, in seamless coordination with full-body movements. We also train a policy to track generated motions in physics simulation via reinforcement learning (RL) to ensure physical plausibility of the motion. Our experiments demonstrate the effectiveness of our system in synthesizing realistic interactions with diverse objects in complex environments, highlighting its potential for real-world applications.

Method: Proposed LACFIP^k algorithm to learn action costs from unlabeled input plans by ensuring these plans become optimal under the learned cost model

Result: Theoretical and empirical results demonstrate that LACFIP^k can successfully solve the action cost learning problem

Conclusion: The paper introduces and successfully addresses the novel problem of learning action costs from unlabeled plans to ensure plan optimality

Abstract: Most of the work on learning action models focus on learning the actions’ dynamics from input plans. This allows us to specify the valid plans of a planning task. However, very little work focuses on learning action costs, which in turn allows us to rank the different plans. In this paper we introduce a new problem: that of learning the costs of a set of actions such that a set of input plans are optimal under the resulting planning model. To solve this problem we present $LACFIP^k$, an algorithm to learn action’s costs from unlabeled input plans. We provide theoretical and empirical results showing how $LACFIP^k$ can successfully solve this task.

Method: Compared SHAP and Occlusion-1 explanation methods with different formats (charts for SHAP, charts and text for Occlusion-1). Conducted experiments with general public (patients) and medical students (practitioners) to measure subjective comprehension and trust.

Result: Occlusion-1 explanations showed higher comprehension and trust than SHAP, but when controlling for format, only text-based Occlusion-1 outperformed chart-based SHAP. Explanation format was often more critical than content.

Conclusion: Text-based explanations may be more effective than chart-based ones for user comprehension and trust in AI healthcare tools, suggesting format preferences can outweigh methodological differences in explanation content.

Abstract: AI-driven tools for healthcare are widely acknowledged as potentially beneficial to health practitioners and patients, e.g. the QCancer regression tool for cancer risk prediction. However, for these tools to be trusted, they need to be supplemented with explanations. We examine how explanations’ content and format affect user comprehension and trust when explaining QCancer’s predictions. Regarding content, we deploy the SHAP and Occlusion-1 explanation methods. Regarding format, we present SHAP explanations, conventionally, as charts (SC) and Occlusion-1 explanations as charts (OC) as well as text (OT), to which their simpler nature lends itself. We conduct experiments with two sets of stakeholders: the general public (representing patients) and medical students (representing healthcare practitioners). Our experiments showed higher subjective comprehension and trust for Occlusion-1 over SHAP explanations based on content. However, when controlling for format, only OT outperformed SC, suggesting this trend is driven by preferences for text. Other findings corroborated that explanation format, rather than content, is often the critical factor.

Method: Developed Visual Language Action Model for Simultaneously Chatting and Decision Making (VLASCD/MIMO-VLA) with parallel multi-task output capabilities, evaluated on CARLA autonomous driving platform.

Result: MIMO-VLA significantly outperforms existing MISO models (LLM dialogue models, reinforcement learning models, and VLA decision-making models) in simultaneously handling dialogue generation and decision-making tasks.

Conclusion: The proposed MIMO architecture overcomes fundamental limitations of MISO models by enabling parallel multi-task processing without performance degradation, demonstrating superior performance in complex multimodal scenarios requiring simultaneous outputs.

Abstract: Although current mainstream pre-trained large models, such as LLM models represented by ChatGPT and VLA models represented by OpenVLA, have achieved significant progress in multimodal tasks through a “Multiple-Input, Single-Output” (MISO) architecture. However, our investigation reveals that the MISO architecture exhibits fundamental limitations in “Multiple-Input, Multiple-Output” (MIMO) (e.g., parallel multi-tasks output processing): the architecture generates task mutual exclusion effects, leading to resource contention among different tasks when sharing output channels, and consequently resulting in optimization imbalance and performance degradation. In contrast, human MIMO processing inherently enables concurrent task execution (e.g., while dialogue and decision-making) without interference. Inspired by this, in this work, we propose a unified MIMO training model with parallel multi-tasks output capabilities termed Visual Language Action Model for Simultaneously Chatting and Decision Making. We refer to this method as VLASCD or MIMO-VLA, and in the following, we will use these two names interchangeably. We evaluate the model on the CARLA autonomous driving platform. The results show that, compared to LLM models with MISO dialogue capabilities, reinforcement learning models, and VLA models with MISO decision-making capabilities, MIMO-VLA significantly outperforms existing MISO models in simultaneously handling dialogue generation and decision-making tasks within the MIMO scenario.

Method: Proposes poisoned sample detection using spatial masking and data attribution, plus adaptive optimization with dynamic penalty term in training loss to reduce reliance on infringing features.

Result: Achieves average F1-score of 0.665 in detection, delays First-Attack Epoch by 115.2%, reduces Copyright Infringement Rate by 56.7%, and improves defense effectiveness by 25% over state-of-the-art.

Conclusion: CopyrightShield significantly enhances diffusion model security against copyright infringement attacks while maintaining generative performance.

Abstract: Diffusion models have attracted significant attention due to its exceptional data generation capabilities in fields such as image synthesis. However, recent studies have shown that diffusion models are vulnerable to copyright infringement attacks, where attackers inject strategically modified non-infringing images into the training set, inducing the model to generate infringing content under the prompt of specific poisoned captions. To address this issue, we first propose a defense framework, CopyrightShield, to defend against the above attack. Specifically, we analyze the memorization mechanism of diffusion models and find that attacks exploit the model’s overfitting to specific spatial positions and prompts, causing it to reproduce poisoned samples under backdoor triggers. Based on this, we propose a poisoned sample detection method using spatial masking and data attribution to quantify poisoning risk and accurately identify hidden backdoor samples. To further mitigate memorization of poisoned features, we introduce an adaptive optimization strategy that integrates a dynamic penalty term into the training loss, reducing reliance on infringing features while preserving generative performance. Experimental results demonstrate that CopyrightShield significantly improves poisoned sample detection performance across two attack scenarios, achieving average F1-scores of 0.665, retarding the First-Attack Epoch (FAE) of 115.2% and decreasing the Copyright Infringement Rate (CIR) by 56.7%. Compared to the SoTA backdoor defense in diffusion models, the defense effect is improved by about 25%, showcasing its superiority and practicality in enhancing the security of diffusion models.

Method: A framework to assess sycophancy using AMPS (mathematics) and MedQuad (medical advice) datasets, testing different rebuttal approaches (preemptive vs in-context, simple vs citation-based) and measuring persistence across contexts.

Result: 58.19% overall sycophancy rate, with Gemini highest (62.47%) and ChatGPT lowest (56.71%). Preemptive rebuttals showed higher sycophancy (61.75%) than in-context (56.52%). Progressive sycophancy (leading to correct answers) 43.52%, regressive (incorrect answers) 14.66%. High persistence (78.5%) across contexts.

Conclusion: LLM sycophancy presents significant reliability risks but also opportunities for optimization. Findings provide insights for prompt programming and model development to enhance safety in structured domains like education and healthcare.

Abstract: Large language models (LLMs) are increasingly applied in educational, clinical, and professional settings, but their tendency for sycophancy – prioritizing user agreement over independent reasoning – poses risks to reliability. This study introduces a framework to evaluate sycophantic behavior in ChatGPT-4o, Claude-Sonnet, and Gemini-1.5-Pro across AMPS (mathematics) and MedQuad (medical advice) datasets. Sycophantic behavior was observed in 58.19% of cases, with Gemini exhibiting the highest rate (62.47%) and ChatGPT the lowest (56.71%). Progressive sycophancy, leading to correct answers, occurred in 43.52% of cases, while regressive sycophancy, leading to incorrect answers, was observed in 14.66%. Preemptive rebuttals demonstrated significantly higher sycophancy rates than in-context rebuttals (61.75% vs. 56.52%, $Z=5.87$, $p<0.001$), particularly in computational tasks, where regressive sycophancy increased significantly (preemptive: 8.13%, in-context: 3.54%, $p<0.001$). Simple rebuttals maximized progressive sycophancy ($Z=6.59$, $p<0.001$), while citation-based rebuttals exhibited the highest regressive rates ($Z=6.59$, $p<0.001$). Sycophantic behavior showed high persistence (78.5%, 95% CI: [77.2%, 79.8%]) regardless of context or model. These findings emphasize the risks and opportunities of deploying LLMs in structured and dynamic domains, offering insights into prompt programming and model optimization for safer AI applications.

Method: Created a synthetic data generation pipeline that produces realistic user profiles and private documents simulating human activities, then used this to build PersonaBench for evaluating RAG pipelines on personal information extraction.

Result: Current retrieval-augmented AI models perform poorly at answering private questions by extracting personal information from user documents.

Conclusion: There is a significant need for improved methodologies to enhance personalization capabilities in AI assistants when working with private user data.

Abstract: Personalization is critical in AI assistants, particularly in the context of private AI models that work with individual users. A key scenario in this domain involves enabling AI models to access and interpret a user’s private data (e.g., conversation history, user-AI interactions, app usage) to understand personal details such as biographical information, preferences, and social connections. However, due to the sensitive nature of such data, there are no publicly available datasets that allow us to assess an AI model’s ability to understand users through direct access to personal information. To address this gap, we introduce a synthetic data generation pipeline that creates diverse, realistic user profiles and private documents simulating human activities. Leveraging this synthetic data, we present PersonaBench, a benchmark designed to evaluate AI models’ performance in understanding personal information derived from simulated private user data. We evaluate Retrieval-Augmented Generation (RAG) pipelines using questions directly related to a user’s personal information, supported by the relevant private documents provided to the models. Our results reveal that current retrieval-augmented AI models struggle to answer private questions by extracting personal information from user documents, highlighting the need for improved methodologies to enhance personalization capabilities in AI.

Method: Proposes a utility function and co-player sampling method for zero-shot human-AI coordination setting, extending multi-agent UED approach to human-AI scenarios.

Result: Outperforms baseline models in Overcooked-AI environment with both human proxy agents and real humans, achieving high performance in unseen environments.

Conclusion: The proposed method effectively trains ego-agents to coordinate with humans in zero-shot settings with better generalization to new environments.

Abstract: Zero-shot human-AI coordination is the training of an ego-agent to coordinate with humans without human data. Most studies on zero-shot human-AI coordination have focused on enhancing the ego-agent’s coordination ability in a given environment without considering the issue of generalization to unseen environments. Real-world applications of zero-shot human-AI coordination should consider unpredictable environmental changes and the varying coordination ability of co-players depending on the environment. Previously, the multi-agent UED (Unsupervised Environment Design) approach has investigated these challenges by jointly considering environmental changes and co-player policy in competitive two-player AI-AI scenarios. In this paper, our study extends a multi-agent UED approach to zero-shot human-AI coordination. We propose a utility function and co-player sampling for a zero-shot human-AI coordination setting that helps train the ego-agent to coordinate with humans more effectively than a previous multi-agent UED approach. The zero-shot human-AI coordination performance was evaluated in the Overcooked-AI environment, using human proxy agents and real humans. Our method outperforms other baseline models and achieves high performance in human-AI coordination tasks in unseen environments. The source code is available at https://github.com/Uwonsang/ACD_Human-AI

Method: Proposes GATES - a novel DRL method using Graph Attention Networks to learn DAG topological relationships and Evolution Strategy for robust policy learning with tolerance for delayed rewards.

Result: Extensive experiments show GATES outperforms several state-of-the-art algorithms in cost-aware dynamic workflow scheduling tasks.

Conclusion: GATES provides an effective solution for CADWS by combining graph-based attention mechanisms with evolutionary strategies, achieving stable and superior scheduling performance.

Abstract: Cost-aware Dynamic Workflow Scheduling (CADWS) is a key challenge in cloud computing, focusing on devising an effective scheduling policy to efficiently schedule dynamically arriving workflow tasks, represented as Directed Acyclic Graphs (DAG), to suitable virtual machines (VMs). Deep reinforcement learning (DRL) has been widely employed for automated scheduling policy design. However, the performance of DRL is heavily influenced by the design of the problem-tailored policy network and is highly sensitive to hyperparameters and the design of reward feedback. Considering the above-mentioned issues, this study proposes a novel DRL method combining Graph Attention Networks-based policy network and Evolution Strategy, referred to as GATES. The contributions of GATES are summarized as follows: (1) GATES can capture the impact of current task scheduling on subsequent tasks by learning the topological relationships between tasks in a DAG. (2) GATES can assess the importance of each VM to the ready task, enabling it to adapt to dynamically changing VM resources. (3) Utilizing Evolution Strategy’s robustness, exploratory nature, and tolerance for delayed rewards, GATES achieves stable policy learning in CADWS. Extensive experimental results demonstrate the superiority of the proposed GATES in CADWS, outperforming several state-of-the-art algorithms. The source code is available at: https://github.com/YaShen998/GATES.

Method: Multi-turn conversational setup between simulated persuader and persuadee agents across diverse harmful topics, using automated evaluator to measure persuasion attempts.

Result: Many open and closed-weight models frequently attempt persuasion on harmful topics, with jailbreaking increasing this willingness.

Conclusion: Current safety guardrails have significant gaps, and evaluating willingness to persuade is crucial for understanding LLM risks.

Abstract: Persuasion is a powerful capability of large language models (LLMs) that both enables beneficial applications (e.g. helping people quit smoking) and raises significant risks (e.g. large-scale, targeted political manipulation). Prior work has found models possess a significant and growing persuasive capability, measured by belief changes in simulated or real users. However, these benchmarks overlook a crucial risk factor: the propensity of a model to attempt to persuade in harmful contexts. Understanding whether a model will blindly ``follow orders’’ to persuade on harmful topics (e.g. glorifying joining a terrorist group) is key to understanding the efficacy of safety guardrails. Moreover, understanding if and when a model will engage in persuasive behavior in pursuit of some goal is essential to understanding the risks from agentic AI systems. We propose the Attempt to Persuade Eval (APE) benchmark, that shifts the focus from persuasion success to persuasion attempts, operationalized as a model’s willingness to generate content aimed at shaping beliefs or behavior. Our evaluation framework probes frontier LLMs using a multi-turn conversational setup between simulated persuader and persuadee agents. APE explores a diverse spectrum of topics including conspiracies, controversial issues, and non-controversially harmful content. We introduce an automated evaluator model to identify willingness to persuade and measure the frequency and context of persuasive attempts. We find that many open and closed-weight models are frequently willing to attempt persuasion on harmful topics and that jailbreaking can increase willingness to engage in such behavior. Our results highlight gaps in current safety guardrails and underscore the importance of evaluating willingness to persuade as a key dimension of LLM risk. APE is available at github.com/AlignmentResearch/AttemptPersuadeEval

Method: Used Sotopia simulation testbed with two experiments: 1) Causal discovery methods to measure personality trait impacts on price bargaining, 2) Human-AI job negotiations manipulating both human personality and AI characteristics (transparency, competence, adaptability).

Result: Agreeableness and Extraversion significantly affect believability, goal achievement, and knowledge acquisition. Sociocognitive measures detected empathic communication, moral foundations, and opinion patterns. AI agent trustworthiness impacts mission effectiveness.

Conclusion: Establishes repeatable evaluation methodology for AI agent reliability across diverse operator personalities and team dynamics, advancing beyond standard performance metrics to incorporate essential social dynamics for mission success.

Abstract: This paper presents an evaluation framework for agentic AI systems in mission-critical negotiation contexts, addressing the need for AI agents that can adapt to diverse human operators and stakeholders. Using Sotopia as a simulation testbed, we present two experiments that systematically evaluated how personality traits and AI agent characteristics influence LLM-simulated social negotiation outcomes–a capability essential for a variety of applications involving cross-team coordination and civil-military interactions. Experiment 1 employs causal discovery methods to measure how personality traits impact price bargaining negotiations, through which we found that Agreeableness and Extraversion significantly affect believability, goal achievement, and knowledge acquisition outcomes. Sociocognitive lexical measures extracted from team communications detected fine-grained differences in agents’ empathic communication, moral foundations, and opinion patterns, providing actionable insights for agentic AI systems that must operate reliably in high-stakes operational scenarios. Experiment 2 evaluates human-AI job negotiations by manipulating both simulated human personality and AI system characteristics, specifically transparency, competence, adaptability, demonstrating how AI agent trustworthiness impact mission effectiveness. These findings establish a repeatable evaluation methodology for experimenting with AI agent reliability across diverse operator personalities and human-agent team dynamics, directly supporting operational requirements for reliable AI systems. Our work advances the evaluation of agentic AI workflows by moving beyond standard performance metrics to incorporate social dynamics essential for mission success in complex operations.

Method: Proposes an intermediate Intention Capture layer that extracts Workflow Signals from user queries, interprets them into structured Workflow Intention objects, and generates workflows based on these intentions rather than directly from queries.

Result: The framework yields consistent improvements in semantic workflow similarity metrics on a benchmark of 1,000 multi-intent query-workflow pairs, with significant quality improvements especially in Mixed Intention Elicitation cases.

Conclusion: The Opus Prompt Intention Framework provides a reproducible and customizable system that significantly enhances workflow generation quality by leveraging structured intention capture before final workflow generation.

Abstract: This paper introduces the Opus Prompt Intention Framework, designed to improve complex Workflow Generation with instruction-tuned Large Language Models (LLMs). We propose an intermediate Intention Capture layer between user queries and Workflow Generation, implementing the Opus Workflow Intention Framework, which consists of extracting Workflow Signals from user queries, interpreting them into structured Workflow Intention objects, and generating Workflows based on these Intentions. Our results show that this layer enables LLMs to produce logical and meaningful outputs that scale reliably as query complexity increases. On a synthetic benchmark of 1,000 multi-intent query-Workflow(s) pairs, applying the Opus Prompt Intention Framework to Workflow Generation yields consistent improvements in semantic Workflow similarity metrics. In this paper, we introduce the Opus Prompt Intention Framework by applying the concepts of Workflow Signal and Workflow Intention to LLM-driven Workflow Generation. We present a reproducible, customizable LLM-based Intention Capture system to extract Workflow Signals and Workflow Intentions from user queries. Finally, we provide empirical evidence that the proposed system significantly improves Workflow Generation quality compared to direct generation from user queries, particularly in cases of Mixed Intention Elicitation.

Method: Leverages Büchi automata structure to decompose LTL specifications into reach-avoid subgoals, solves them sequentially using safe RL, and introduces subgoal-induced observation reduction to handle exponential complexity.

Result: Substantially outperforms existing methods in zero-shot generalization to unseen LTL specifications.

Conclusion: Solving reach-avoid subgoals sequentially is more effective for zero-shot generalization than conditioning on subgoal sequences, and the proposed observation reduction technique successfully mitigates complexity issues.

Abstract: Generalizing to complex and temporally extended task objectives and safety constraints remains a critical challenge in reinforcement learning (RL). Linear temporal logic (LTL) offers a unified formalism to specify such requirements, yet existing methods are limited in their abilities to handle nested long-horizon tasks and safety constraints, and cannot identify situations when a subgoal is not satisfiable and an alternative should be sought. In this paper, we introduce GenZ-LTL, a method that enables zero-shot generalization to arbitrary LTL specifications. GenZ-LTL leverages the structure of B"uchi automata to decompose an LTL task specification into sequences of reach-avoid subgoals. Contrary to the current state-of-the-art method that conditions on subgoal sequences, we show that it is more effective to achieve zero-shot generalization by solving these reach-avoid problems \textit{one subgoal at a time} through proper safe RL formulations. In addition, we introduce a novel subgoal-induced observation reduction technique that can mitigate the exponential complexity of subgoal-state combinations under realistic assumptions. Empirical results show that GenZ-LTL substantially outperforms existing methods in zero-shot generalization to unseen LTL specifications.

Method: The authors propose a different perspective where an observer can interpret any regulating agent as having “beliefs” about its environment that it updates based on sensory input, making models observer-imposed rather than intrinsic properties.

Result: A more sophisticated notion of model and a theorem that applies broadly to regulation tasks, including both environmental regulation and internal state regulation, resolving apparent counterexamples through potentially trivial models.

Conclusion: Models are not intrinsic properties of systems but are imposed by observers, and this perspective provides a more general framework that maintains the intuition behind Conant and Ashby’s theorem while accommodating diverse regulatory systems.

Abstract: In a classic paper, Conant and Ashby claimed that “every good regulator of a system must be a model of that system.” Artificial Life has produced many examples of systems that perform tasks with apparently no model in sight; these suggest Conant and Ashby’s theorem doesn’t easily generalise beyond its restricted setup. Nevertheless, here we show that a similar intuition can be fleshed out in a different way: whenever an agent is able to perform a regulation task, it is possible for an observer to interpret it as having “beliefs” about its environment, which it “updates” in response to sensory input. This notion of belief updating provides a notion of model that is more sophisticated than Conant and Ashby’s, as well as a theorem that is more broadly applicable. However, it necessitates a change in perspective, in that the observer plays an essential role in the theory: models are not a mere property of the system but are imposed on it from outside. Our theorem holds regardless of whether the system is regulating its environment in a classic control theory setup, or whether it’s regulating its own internal state; the model is of its environment either way. The model might be trivial, however, and this is how the apparent counterexamples are resolved.

Method: GRPO-based interactive training where a teacher model provides corrective feedback on student rollouts, inspired by classroom teaching practices of posing problems and giving guidance.

Result: 3.85% average improvement over zero-shot baselines, with specific improvements of 2.08% on MATH-500, 2.23% on AIME, and 3.99% on GPQA-Diamond over vanilla-GRPO baseline.

Conclusion: Implicit supervision through teacher feedback effectively improves reasoning capabilities in student models, demonstrating the value of interactive training approaches for developing thinking behaviors.

Abstract: Recent advances in test-time scaling have led to the emergence of thinking LLMs that exhibit self-reflective behaviors and multi-step reasoning. While RL drives this self-improvement paradigm, a recent study (Gandhi et al., 2025) shows that RL alone does not truly instill these new reasoning abilities - it merely draws out behaviors already present in the base models. This raises a question: How can we train the models that don’t exhibit such thinking behavior to develop it in the first place? To this end, we propose ThinkTuning, a GRPO-based interactive training approach where we augment the rollouts of a student model with the guidance from a teacher model. A simple idea from classroom practice inspires our method: a teacher poses a problem, lets the student try an answer, then gives corrective feedback – enough to point the mind in the right direction and then show the solution. Each piece of feedback reshapes the student’s thoughts, leading them to arrive at the correct solution. Similarly, we find that this type of implicit supervision through feedback from a teacher model of the same size improves the reasoning capabilities of the student model. In particular, on average, our method shows a 3.85% improvement over zero-shot baselines across benchmarks, and on MATH-500, AIME and GPQA-Diamond it shows 2.08%, 2.23% and 3.99% improvements over the vanilla-GRPO baseline. Source code is available at https://github.com/3rdAT/ThinkTuning.

Method: Proposes ITL-LIME framework that uses clustering to partition source domain, retrieves relevant real source instances instead of random perturbations, employs contrastive learning-based encoder for weighting, and trains surrogate model with weighted source and target instances.

Result: Enhanced explanation fidelity and stability in data-constrained environments by leveraging real instances from related source domains to better approximate the original model’s decision boundary.

Conclusion: ITL-LIME effectively addresses LIME’s limitations in data-scarce scenarios by incorporating instance transfer learning, providing more stable and faithful explanations through the use of real data instances rather than synthetic perturbations.

Abstract: Explainable Artificial Intelligence (XAI) methods, such as Local Interpretable Model-Agnostic Explanations (LIME), have advanced the interpretability of black-box machine learning models by approximating their behavior locally using interpretable surrogate models. However, LIME’s inherent randomness in perturbation and sampling can lead to locality and instability issues, especially in scenarios with limited training data. In such cases, data scarcity can result in the generation of unrealistic variations and samples that deviate from the true data manifold. Consequently, the surrogate model may fail to accurately approximate the complex decision boundary of the original model. To address these challenges, we propose a novel Instance-based Transfer Learning LIME framework (ITL-LIME) that enhances explanation fidelity and stability in data-constrained environments. ITL-LIME introduces instance transfer learning into the LIME framework by leveraging relevant real instances from a related source domain to aid the explanation process in the target domain. Specifically, we employ clustering to partition the source domain into clusters with representative prototypes. Instead of generating random perturbations, our method retrieves pertinent real source instances from the source cluster whose prototype is most similar to the target instance. These are then combined with the target instance’s neighboring real instances. To define a compact locality, we further construct a contrastive learning-based encoder as a weighting mechanism to assign weights to the instances from the combined set based on their proximity to the target instance. Finally, these weighted source and target instances are used to train the surrogate model for explanation purposes.

Method: Developed a lightweight neural network architecture (two hidden layer perceptron) combined with signal processing techniques to denoise acoustic signals and detect impulsive muzzle blast waveforms, avoiding heavy convolutional neural networks to conserve computational resources.

Result: Detection rate for muzzle blast waveforms more than doubled when root mean square noise value is comparable to the muzzle blast peak amplitude, significantly improving performance in noisy environments.

Conclusion: The lightweight neural network approach effectively enhances gunshot detection capabilities in challenging acoustic environments while maintaining computational efficiency suitable for deployment on various military hardware platforms.

Abstract: Acoem develops gunshot detection systems, consisting of a microphone array and software that detects and locates shooters on the battlefield. The performance of such systems is obviously affected by the acoustic environment in which they are operating: in particular, when mounted on a moving military vehicle, the presence of noise reduces the detection performance of the software. To limit the influence of the acoustic environment, a neural network has been developed. Instead of using a heavy convolutional neural network, a lightweight neural network architecture was chosen to limit the computational resources required to embed the algorithm on as many hardware platforms as possible. Thanks to the combination of a two hidden layer perceptron and appropriate signal processing techniques, the detection rate of impulsive muzzle blast waveforms (the wave coming from the detonation and indicating the position of the shooter) is significantly increased. With a rms value of noise of the same order as the muzzle blast peak amplitude, the detect rate is more than doubled with this denoising processing.

Method: Multi-stage approach: 1) Train classical speech emotion recognition model, 2) Synthetic generation of emotional sequences, 3) Model improvement via human feedback, 4) Novel Dirichlet distribution-based emotional mixture modeling.

Result: Dirichlet-based approach effectively models emotional mixtures. Human feedback integration improves model quality while simplifying annotation. Outperforms sliding window approach when evaluated on ground-truth emotions from 3D facial animation dataset.

Conclusion: The proposed dynamic emotion recognition with Dirichlet mixture modeling and human feedback is effective for sequential emotion analysis in speech, particularly beneficial for 3D avatar animation applications.

Abstract: This work proposes to explore a new area of dynamic speech emotion recognition. Unlike traditional methods, we assume that each audio track is associated with a sequence of emotions active at different moments in time. The study particularly focuses on the animation of emotional 3D avatars. We propose a multi-stage method that includes the training of a classical speech emotion recognition model, synthetic generation of emotional sequences, and further model improvement based on human feedback. Additionally, we introduce a novel approach to modeling emotional mixtures based on the Dirichlet distribution. The models are evaluated based on ground-truth emotions extracted from a dataset of 3D facial animations. We compare our models against the sliding window approach. Our experimental results show the effectiveness of Dirichlet-based approach in modeling emotional mixtures. Incorporating human feedback further improves the model quality while providing a simplified annotation procedure.

Method: Employ occlusion maps on CNN-processed cough spectrograms to highlight relevant spectral regions, then perform spectral analysis on weighted spectrograms to extract features and identify differences between disease groups.

Result: Significant differences found between disease groups (especially COPD patients) in identified spectral regions, contrasting with no significant differences in raw spectrograms. COPD cough patterns show more variability in specific spectral regions.

Conclusion: XAI techniques can uncover disease-specific acoustic signatures and improve diagnostic capabilities of cough sound analysis by providing interpretable results that reveal patterns not visible in traditional analysis.

Abstract: This paper proposes an eXplainable Artificial Intelligence (XAI)-driven methodology to enhance the understanding of cough sound analysis for respiratory disease management. We employ occlusion maps to highlight relevant spectral regions in cough spectrograms processed by a Convolutional Neural Network (CNN). Subsequently, spectral analysis of spectrograms weighted by these occlusion maps reveals significant differences between disease groups, particularly in patients with COPD, where cough patterns appear more variable in the identified spectral regions of interest. This contrasts with the lack of significant differences observed when analyzing raw spectrograms. The proposed approach extracts and analyzes several spectral features, demonstrating the potential of XAI techniques to uncover disease-specific acoustic signatures and improve the diagnostic capabilities of cough sound analysis by providing more interpretable results.

Method: Methodological replication with 44 participants assessing comprehension of healthcare audio content generated from original vs. simplified texts, while examining education level and language proficiency effects.

Result: Text simplification effectively enhanced both perceived understandability and actual comprehension of audio healthcare information, consistent with original study findings.

Conclusion: Text simplification tools have practical value for health literacy, and tailored communication strategies are needed to effectively reach diverse audiences in healthcare.

Abstract: This study serves as a methodological replication of Leroy et al. (2022) research, which investigated the impact of text simplification on healthcare information comprehension in the evolving multimedia landscape. Building upon the original studys insights, our replication study evaluates audio content, recognizing its increasing importance in disseminating healthcare information in the digital age. Specifically, we explored the influence of text simplification on perceived and actual difficulty when users engage with audio content automatically generated from that text. Our replication involved 44 participants for whom we assessed their comprehension of healthcare information presented as audio created using Leroy et al. (2022) original and simplified texts. The findings from our study highlight the effectiveness of text simplification in enhancing perceived understandability and actual comprehension, aligning with the original studys results. Additionally, we examined the role of education level and language proficiency, shedding light on their potential impact on healthcare information access and understanding. This research underscores the practical value of text simplification tools in promoting health literacy. It suggests the need for tailored communication strategies to reach diverse audiences effectively in the healthcare domain.

Method: 1) Novel audio feature using filter banks with evenly distributed intervals to ensure equal attention to all frequency ranges 2) Parameter-free feature enhancement approach based on pre-trained models to remove redundant information.

Result: Evaluation on DCASE 2024 Challenge dataset demonstrates significant improvements in anomalous sound detection performance.

Conclusion: The proposed methods effectively enhance anomaly detection in machine sounds by addressing frequency bias and redundant information through novel filter design and parameter-free feature enhancement.

Abstract: Anomalous Sound Detection (ASD) aims at identifying anomalous sounds from machines and has gained extensive research interests from both academia and industry. However, the uncertainty of anomaly location and much redundant information such as noise in machine sounds hinder the improvement of ASD system performance. This paper proposes a novel audio feature of filter banks with evenly distributed intervals, ensuring equal attention to all frequency ranges in the audio, which enhances the detection of anomalies in machine sounds. Moreover, based on pre-trained models, this paper presents a parameter-free feature enhancement approach to remove redundant information in machine audio. It is believed that this parameter-free strategy facilitates the effective transfer of universal knowledge from pre-trained tasks to the ASD task during model fine-tuning. Evaluation results on the Detection and Classification of Acoustic Scenes and Events (DCASE) 2024 Challenge dataset demonstrate significant improvements in ASD performance with our proposed methods.

Method: Three-stage reannotation framework using cross-modal prompting strategy with prompt chaining for audio comprehension, label synthesis, and semantic alignment, leveraging general-purpose audio-language foundation models.

Result: Extensive experiments on representative audio classification models (AST, PANNs, SSAST, AudioMAE) consistently demonstrate substantial performance improvements, validating the approach’s effectiveness.

Conclusion: The proposed framework successfully enhances label reliability in AudioSet, creating a high-quality structured relabeled version (AudioSet-R) that advances audio research capabilities.

Abstract: AudioSet is a widely used benchmark in the audio research community and has significantly advanced various audio-related tasks. However, persistent issues with label accuracy and completeness remain critical bottlenecks that limit performance in downstream applications.To address the aforementioned challenges, we propose a three-stage reannotation framework that harnesses general-purpose audio-language foundation models to systematically improve the label quality of AudioSet. The framework employs a cross-modal prompting strategy, inspired by the concept of prompt chaining, wherein prompts are sequentially composed to execute subtasks (audio comprehension, label synthesis, and semantic alignment). Leveraging this framework, we construct a high-quality, structured relabeled version of AudioSet-R. Extensive experiments conducted on representative audio classification models–including AST, PANNs, SSAST, and AudioMAE–consistently demonstrate substantial performance improvements, thereby validating the generalizability and effectiveness of the proposed approach in enhancing label reliability.The code is publicly available at: https://github.com/colaudiolab/AudioSet-R.

Method: Proposes a Dual Watermark Embedding (DWE) module to embed dual watermarks into Mel-spectrogram representations, combined with a Watermark Consistency Loss (WCL) to ensure reliable extraction from generated audio. Also establishes the Dual Attribution Benchmark (DAB) for robustness evaluation.

Result: Achieves outstanding attribution accuracy with 97.01% F1-score for model attribution and 91.51% AUC for dataset attribution, while maintaining exceptional robustness against pruning, compression, noise, and sampling attacks.

Conclusion: DualMark provides a foundational step toward fully accountable audio generative models, significantly enhancing copyright protection and responsibility tracing capabilities by enabling simultaneous model and dataset provenance tracking.

Abstract: Existing watermarking methods for audio generative models only enable model-level attribution, allowing the identification of the originating generation model, but are unable to trace the underlying training dataset. This significant limitation raises critical provenance questions, particularly in scenarios involving copyright and accountability concerns. To bridge this fundamental gap, we introduce DualMark, the first dual-provenance watermarking framework capable of simultaneously encoding two distinct attribution signatures, i.e., model identity and dataset origin, into audio generative models during training. Specifically, we propose a novel Dual Watermark Embedding (DWE) module to seamlessly embed dual watermarks into Mel-spectrogram representations, accompanied by a carefully designed Watermark Consistency Loss (WCL), which ensures reliable extraction of both watermarks from generated audio signals. Moreover, we establish the Dual Attribution Benchmark (DAB), the first robustness evaluation benchmark specifically tailored for joint model-data attribution. Extensive experiments validate that DualMark achieves outstanding attribution accuracy (97.01% F1-score for model attribution, and 91.51% AUC for dataset attribution), while maintaining exceptional robustness against aggressive pruning, lossy compression, additive noise, and sampling attacks, conditions that severely compromise prior methods. Our work thus provides a foundational step toward fully accountable audio generative models, significantly enhancing copyright protection and responsibility tracing capabilities.

Method: Uses any-to-any training strategy with batch mean loss to anonymize utterances to common pseudo-speakers (average speaker in mini-batch). Combines untargeted attack and any-to-any strategies in speaker-adversarial speech generation model.

Result: Effective asynchronous voice anonymization demonstrated on VoxCeleb datasets. Explored limitations of speaker-adversarial speech against black-box speaker extractors and adaptive attacks.

Conclusion: Proposed method provides safer voice anonymization without targeting real speakers, with insights for future research on privacy protection efficacy and generalization capabilities.

Abstract: Speaker attribute perturbation offers a feasible approach to asynchronous voice anonymization by employing adversarially perturbed speech as anonymized output. In order to enhance the identity unlinkability among anonymized utterances from the same original speaker, the targeted attack training strategy is usually applied to anonymize the utterances to a common designated speaker. However, this strategy may violate the privacy of the designated speaker who is an actual speaker. To mitigate this risk, this paper proposes an any-to-any training strategy. It is accomplished by defining a batch mean loss to anonymize the utterances from various speakers within a training mini-batch to a common pseudo-speaker, which is approximated as the average speaker in the mini-batch. Based on this, a speaker-adversarial speech generation model is proposed, incorporating the supervision from both the untargeted attack and the any-to-any strategies. The speaker attribute perturbations are generated and incorporated into the original speech to produce its anonymized version. The effectiveness of the proposed model was justified in asynchronous voice anonymization through experiments conducted on the VoxCeleb datasets. Additional experiments were carried out to explore the potential limitations of speaker-adversarial speech in voice privacy protection. With them, we aim to provide insights for future research on its protective efficacy against black-box speaker extractors \textcolor{black}{and adaptive attacks, as well as} generalization to out-of-domain datasets \textcolor{black}{and stability}. Audio samples and open-source code are published in https://github.com/VoicePrivacy/any-to-any-speaker-attribute-perturbation.

Method: Two classification methods using textural features of mel-spectrogram images: Support Vector Machine (SVM) and Convolutional Neural Network (CNN) models.

Result: Both SVM and CNN models achieved consistent classification of vocal register in male pop music audio signals.

Conclusion: The successful classification supports the promise of more robust vocal register classification across different voice types and singing genres, with practical applications in vocal analysis tools like the developed AVRA software.

Abstract: For singers of all experience levels, one of the most daunting challenges in learning technical repertoire is navigating placement and vocal register in and around the passagio (passage between chest voice and head voice registers). Particularly in pop music, where a single artist may use a variety of timbre’s and textures to achieve a desired quality, it can be difficult to identify what vocal register within the vocal range a singer is using. This paper presents two methods for classifying vocal registers in an audio signal of male pop music through the analysis of textural features of mel-spectrogram images. Additionally, we will discuss the practical integration of these models for vocal analysis tools, and introduce a concurrently developed software called AVRA which stands for Automatic Vocal Register Analysis. Our proposed methods achieved consistent classification of vocal register through both Support Vector Machine (SVM) and Convolutional Neural Network (CNN) models, which supports the promise of more robust classification possibilities across more voice types and genres of singing.

Method: Proposed ASCMamba uses DenseEncoder for hierarchical spectral feature extraction, dual-path Mamba blocks with state space models to capture long-range temporal and frequency dependencies, and a two-step pseudo-labeling mechanism for reliable pseudo-label generation.

Result: The system outperformed all participating teams and achieved a 6.2% improvement over the baseline in the APSIPA ASC 2025 Grand Challenge.

Conclusion: ASCMamba demonstrates effective multimodal integration of audio and textual information for acoustic scene classification, setting new state-of-the-art performance with publicly available code and models.

Abstract: Acoustic Scene Classification (ASC) is a fundamental problem in computational audition, which seeks to classify environments based on the distinctive acoustic features. In the ASC task of the APSIPA ASC 2025 Grand Challenge, the organizers introduce a multimodal ASC task. Unlike traditional ASC systems that rely solely on audio inputs, this challenge provides additional textual information as inputs, including the location where the audio is recorded and the time of recording. In this paper, we present our proposed system for the ASC task in the APSIPA ASC 2025 Grand Challenge. Specifically, we propose a multimodal network, \textbf{ASCMamba}, which integrates audio and textual information for fine-grained acoustic scene understanding and effective multimodal ASC. The proposed ASCMamba employs a DenseEncoder to extract hierarchical spectral features from spectrograms, followed by a dual-path Mamba blocks that capture long-range temporal and frequency dependencies using Mamba-based state space models. In addition, we present a two-step pseudo-labeling mechanism to generate more reliable pseudo-labels. Results show that the proposed system outperforms all the participating teams and achieves a 6.2% improvement over the baseline. Code, model and pre-trained checkpoints are available at https://github.com/S-Orion/ASCMamba.git.

Method: Fine-tuning ASR models by comparing generalized models (trained across multiple speakers) with personalized models tailored to individual speech characteristics. Evaluation conducted across diverse voice-AI scenarios including virtual assistants and video interviews.

Result: Personalized ASRs significantly reduce word error rates compared to generalized models, with particularly notable improvements in spontaneous speech contexts.

Conclusion: Tailored personalized models show strong potential for creating more inclusive voice technologies that better serve people who stutter.

Abstract: Stuttering – characterized by involuntary disfluencies such as blocks, prolongations, and repetitions – is often misinterpreted by automatic speech recognition (ASR) systems, resulting in elevated word error rates and making voice-driven technologies inaccessible to people who stutter. The variability of disfluencies across speakers and contexts further complicates ASR training, compounded by limited annotated stuttered speech data. In this paper, we investigate fine-tuning ASRs for stuttered speech, comparing generalized models (trained across multiple speakers) to personalized models tailored to individual speech characteristics. Using a diverse range of voice-AI scenarios, including virtual assistants and video interviews, we evaluate how personalization affects transcription accuracy. Our findings show that personalized ASRs significantly reduce word error rates, especially in spontaneous speech, highlighting the potential of tailored models for more inclusive voice technologies.

Method: Integrates frozen diffusion language model with lightweight dual-adapter architecture. Uses two-stage training: semantic alignment via ASR objective first, then instruction-following learning through synthetic audio-caption pairs generated by LLMs.

Result: Competitive performance on major benchmarks (MMSU, MMAU, VoiceBench) despite training on only 960 hours of ASR and 127 hours of synthetic instruction data. Outperforms several autoregressive open-source baselines.

Conclusion: Demonstrates the potential of diffusion-based language models for efficient and scalable audio understanding, opening new directions for speech-driven AI.

Abstract: Recent advances in large language models (LLMs) have shown remarkable capabilities across textual and multimodal domains. In parallel, diffusion-based language models have emerged as a promising alternative to the autoregressive paradigm, offering improved controllability, bidirectional context modeling, and robust generation. However, their application to the audio modality remains underexplored. In this work, we introduce \textbf{DIFFA}, the first diffusion-based large audio-language model designed to perform spoken language understanding. DIFFA integrates a frozen diffusion language model with a lightweight dual-adapter architecture that bridges speech understanding and natural language reasoning. We employ a two-stage training pipeline: first, aligning semantic representations via an ASR objective; then, learning instruction-following abilities through synthetic audio-caption pairs automatically generated by prompting LLMs. Despite being trained on only 960 hours of ASR and 127 hours of synthetic instruction data, DIFFA demonstrates competitive performance on major benchmarks, including MMSU, MMAU, and VoiceBench, outperforming several autoregressive open-source baselines. Our results reveal the potential of diffusion-based language models for efficient and scalable audio understanding, opening a new direction for speech-driven AI. Our code will be available at https://github.com/NKU-HLT/DIFFA.git.

Method: Develops sound source estimation to get 2D coordinates and depth from video frames, maps to 3D trajectory, then uses diffusion model with monaural audio and 3D trajectory conditioning to generate binaural audio. Uses HRIR-based dataset with various sound movement scenarios.

Result: Outperforms existing approaches in spatial perception consistency, effectively enhancing immersive quality of audio-visual experience.

Conclusion: The proposed FoleySpace framework successfully generates spatially consistent binaural audio that provides stronger immersion compared to mono audio approaches.

Abstract: Recently, with the advancement of AIGC, deep learning-based video-to-audio (V2A) technology has garnered significant attention. However, existing research mostly focuses on mono audio generation that lacks spatial perception, while the exploration of binaural spatial audio generation technologies, which can provide a stronger sense of immersion, remains insufficient. To solve this problem, we propose FoleySpace, a framework for video-to-binaural audio generation that produces immersive and spatially consistent stereo sound guided by visual information. Specifically, we develop a sound source estimation method to determine the sound source 2D coordinates and depth in each video frame, and then employ a coordinate mapping mechanism to convert the 2D source positions into a 3D trajectory. This 3D trajectory, together with the monaural audio generated by a pre-trained V2A model, serves as a conditioning input for a diffusion model to generate spatially consistent binaural audio. To support the generation of dynamic sound fields, we constructed a training dataset based on recorded Head-Related Impulse Responses that includes various sound source movement scenarios. Experimental results demonstrate that the proposed method outperforms existing approaches in spatial perception consistency, effectively enhancing the immersive quality of the audio-visual experience.

Method: Hierarchical framework with Safe RL agent using composite ethical risk cost (collision probability + harm severity) at decision level, and polynomial path planning with PID/Stanley controllers at execution level. Uses dynamic Prioritized Experience Replay for critical events.

Result: Outperforms baseline methods in reducing ethical risk and maintaining driving performance on real-world traffic datasets with diverse road users.

Conclusion: First study of ethical decision-making via Safe RL in real-world scenarios, showing potential of combining control theory and data-driven learning for ethically accountable autonomy.

Abstract: Autonomous vehicles hold great promise for reducing traffic fatalities and improving transportation efficiency, yet their widespread adoption hinges on embedding robust ethical reasoning into routine and emergency maneuvers. Here, we present a hierarchical Safe Reinforcement Learning (Safe RL) framework that explicitly integrates moral considerations with standard driving objectives. At the decision level, a Safe RL agent is trained using a composite ethical risk cost, combining collision probability and harm severity, to generate high-level motion targets. A dynamic Prioritized Experience Replay mechanism amplifies learning from rare but critical, high-risk events. At the execution level, polynomial path planning coupled with Proportional-Integral-Derivative (PID) and Stanley controllers translates these targets into smooth, feasible trajectories, ensuring both accuracy and comfort. We train and validate our approach on rich, real-world traffic datasets encompassing diverse vehicles, cyclists, and pedestrians, and demonstrate that it outperforms baseline methods in reducing ethical risk and maintaining driving performance. To our knowledge, this is the first study of ethical decision-making for autonomous vehicles via Safe RL in real-world scenarios. Our results highlight the potential of combining formal control theory and data-driven learning to advance ethically accountable autonomy in complex, human-mixed traffic environments.

Method: Two-stage pipeline: 1) FAISS-based similarity search to retrieve most relevant patient cohort from multi-institutional database, 2) LLM prompting with retrieved cohort and performance metrics to recommend optimal prediction algorithm from 8 model types.

Result: Enables dynamic, cohort-aware risk prediction personalized to individual patient profiles using CT scans and structured metadata.

Conclusion: Provides flexible, cohort-driven model selection for individualized risk assessment in real-world lung cancer screening across diverse clinical populations.

Abstract: Accurate lung cancer risk prediction remains challenging due to substantial variability across patient populations and clinical settings – no single model performs best for all cohorts. To address this, we propose a personalized lung cancer risk prediction agent that dynamically selects the most appropriate model for each patient by combining cohort-specific knowledge with modern retrieval and reasoning techniques. Given a patient’s CT scan and structured metadata – including demographic, clinical, and nodule-level features – the agent first performs cohort retrieval using FAISS-based similarity search across nine diverse real-world cohorts to identify the most relevant patient population from a multi-institutional database. Second, a Large Language Model (LLM) is prompted with the retrieved cohort and its associated performance metrics to recommend the optimal prediction algorithm from a pool of eight representative models, including classical linear risk models (e.g., Mayo, Brock), temporally-aware models (e.g., TDVIT, DLSTM), and multi-modal computer vision-based approaches (e.g., Liao, Sybil, DLS, DLI). This two-stage agent pipeline – retrieval via FAISS and reasoning via LLM – enables dynamic, cohort-aware risk prediction personalized to each patient’s profile. Building on this architecture, the agent supports flexible and cohort-driven model selection across diverse clinical populations, offering a practical path toward individualized risk assessment in real-world lung cancer screening.

Method: Uses graph neural networks to model structural relationships among multimodal clinical symptoms, Transformer architecture for temporal modeling, and a structure-aware gating mechanism to fuse structural and temporal information with multi-component modeling pipeline.

Result: Outperforms existing approaches in AUC, RMSE, and IPW-F1 metrics, effectively distinguishes progression stages and captures personalized symptom trajectories.

Conclusion: The framework demonstrates strong generalization and structural scalability, providing reliable support for intelligent modeling of chronic progressive diseases like Parkinson’s.

Abstract: This study addresses the challenges of symptom evolution complexity and insufficient temporal dependency modeling in Parkinson’s disease progression prediction. It proposes a unified prediction framework that integrates structural perception and temporal modeling. The method leverages graph neural networks to model the structural relationships among multimodal clinical symptoms and introduces graph-based representations to capture semantic dependencies between symptoms. It also incorporates a Transformer architecture to model dynamic temporal features during disease progression. To fuse structural and temporal information, a structure-aware gating mechanism is designed to dynamically adjust the fusion weights between structural encodings and temporal features, enhancing the model’s ability to identify key progression stages. To improve classification accuracy and stability, the framework includes a multi-component modeling pipeline, consisting of a graph construction module, a temporal encoding module, and a prediction output layer. The model is evaluated on real-world longitudinal Parkinson’s disease data. The experiments involve comparisons with mainstream models, sensitivity analysis of hyperparameters, and graph connection density control. Results show that the proposed method outperforms existing approaches in AUC, RMSE, and IPW-F1 metrics. It effectively distinguishes progression stages and improves the model’s ability to capture personalized symptom trajectories. The overall framework demonstrates strong generalization and structural scalability, providing reliable support for intelligent modeling of chronic progressive diseases such as Parkinson’s disease.

Method: Introduces hierarchical hybrid architecture templates based on domain-specific cues, employs adaptive mutation strategies using Q-learning that dynamically adjust based on performance feedback, and uses a fully probabilistic model for effective search space exploration.

Result: The method consistently discovers high-performing architectures and achieves an 8% improvement in test accuracy over existing baselines on the Spider dataset for database ID prediction tasks.

Conclusion: HHNAS-AM provides an organized, efficient approach to neural architecture search for text classification by using hierarchical templates and adaptive mutation policies, significantly outperforming previous methods.

Abstract: Neural Architecture Search (NAS) has garnered significant research interest due to its capability to discover architectures superior to manually designed ones. Learning text representation is crucial for text classification and other language-related tasks. The NAS model used in text classification does not have a Hybrid hierarchical structure, and there is no restriction on the architecture structure, due to which the search space becomes very large and mostly redundant, so the existing RL models are not able to navigate the search space effectively. Also, doing a flat architecture search leads to an unorganised search space, which is difficult to traverse. For this purpose, we propose HHNAS-AM (Hierarchical Hybrid Neural Architecture Search with Adaptive Mutation Policies), a novel approach that efficiently explores diverse architectural configurations. We introduce a few architectural templates to search on which organise the search spaces, where search spaces are designed on the basis of domain-specific cues. Our method employs mutation strategies that dynamically adapt based on performance feedback from previous iterations using Q-learning, enabling a more effective and accelerated traversal of the search space. The proposed model is fully probabilistic, enabling effective exploration of the search space. We evaluate our approach on the database id (db_id) prediction task, where it consistently discovers high-performing architectures across multiple experiments. On the Spider dataset, our method achieves an 8% improvement in test accuracy over existing baselines.

Method: Three key innovations: 1) Gradient decoupling using absolute difference loss instead of log-sigmoid, 2) Stability improvement through offset constraint with positive regularization, 3) Controllable rejection suppression with gradient separation and tunable coefficient.

Result: LPO consistently improves performance across various tasks including general text tasks, math tasks, and text-to-speech tasks, demonstrating robust and tunable preference alignment.

Conclusion: LPO establishes itself as a robust and tunable paradigm for preference alignment, with publicly released source code, models, and training data to facilitate further research.

Abstract: DPO (Direct Preference Optimization) has become a widely used offline preference optimization algorithm due to its simplicity and training stability. However, DPO is prone to overfitting and collapse. To address these challenges, we propose Linear Preference Optimization (LPO), a novel alignment framework featuring three key innovations. First, we introduce gradient decoupling by replacing the log-sigmoid function with an absolute difference loss, thereby isolating the optimization dynamics. Second, we improve stability through an offset constraint combined with a positive regularization term to preserve the chosen response quality. Third, we implement controllable rejection suppression using gradient separation with straightforward estimation and a tunable coefficient that linearly regulates the descent of the rejection probability. Through extensive experiments, we demonstrate that LPO consistently improves performance on various tasks, including general text tasks, math tasks, and text-to-speech (TTS) tasks. These results establish LPO as a robust and tunable paradigm for preference alignment, and we release the source code, models, and training data publicly.

Method: Proposes comprehensive transfer of all representations (URs, IRs, CRs), identifies optimal extraction layers for CRs, and employs multi-granularity mechanisms: non-linear adapters for feature-level transfer, Isomorphic Interaction Module for module-level transfer, and Standalone Retrieval for model-level transfer.

Result: Successfully deployed in Tencent’s industrial-scale advertising platform, processing tens of billions of daily samples. Achieved 10+ production launches across various advertising scenarios with 2.45% overall GMV lift across the platform, translating to estimated annual revenue increases in hundreds of millions of dollars.

Conclusion: LFM4Ads demonstrates that comprehensive transfer of all representations with multi-granularity mechanisms significantly improves advertising recommendation performance and delivers substantial business value at industrial scale.

Abstract: Online advertising relies on accurate recommendation models, with recent advances using pre-trained large-scale foundation models (LFMs) to capture users’ general interests across multiple scenarios and tasks. However, existing methods have critical limitations: they extract and transfer only user representations (URs), ignoring valuable item representations (IRs) and user-item cross representations (CRs); and they simply use a UR as a feature in downstream applications, which fails to bridge upstream-downstream gaps and overlooks more transfer granularities. In this paper, we propose LFM4Ads, an All-Representation Multi-Granularity transfer framework for ads recommendation. It first comprehensively transfers URs, IRs, and CRs, i.e., all available representations in the pre-trained foundation model. To effectively utilize the CRs, it identifies the optimal extraction layer and aggregates them into transferable coarse-grained forms. Furthermore, we enhance the transferability via multi-granularity mechanisms: non-linear adapters for feature-level transfer, an Isomorphic Interaction Module for module-level transfer, and Standalone Retrieval for model-level transfer. LFM4Ads has been successfully deployed in Tencent’s industrial-scale advertising platform, processing tens of billions of daily samples while maintaining terabyte-scale model parameters with billions of sparse embedding keys across approximately two thousand features. Since its production deployment in Q4 2024, LFM4Ads has achieved 10+ successful production launches across various advertising scenarios, including primary ones like Weixin Moments and Channels. These launches achieve an overall GMV lift of 2.45% across the entire platform, translating to estimated annual revenue increases in the hundreds of millions of dollars.

Method: Proposed DiffQAS-QLSTM, an end-to-end differentiable framework that simultaneously optimizes VQC parameters and architecture selection during training for quantum recurrent models.

Result: The framework consistently outperforms handcrafted baselines, achieving lower loss across diverse test settings for quantum sequence learning.

Conclusion: This approach enables scalable and adaptive quantum sequence learning, opening new possibilities for quantum machine learning applications in time-series prediction, NLP, and reinforcement learning.

Abstract: Recent advances in quantum computing and machine learning have given rise to quantum machine learning (QML), with growing interest in learning from sequential data. Quantum recurrent models like QLSTM are promising for time-series prediction, NLP, and reinforcement learning. However, designing effective variational quantum circuits (VQCs) remains challenging and often task-specific. To address this, we propose DiffQAS-QLSTM, an end-to-end differentiable framework that optimizes both VQC parameters and architecture selection during training. Our results show that DiffQAS-QLSTM consistently outperforms handcrafted baselines, achieving lower loss across diverse test settings. This approach opens the door to scalable and adaptive quantum sequence learning.

Method: Uses a curriculum-guided training approach with progressive masking ratios, forcing a ViT decoder to reconstruct from sparser context. At inference, performs Monte Carlo sampling over random mask realizations to approximate posterior predictive distribution.

Result: Achieves high-fidelity reconstruction of fine-scale atmospheric features (updrafts, downdrafts, shear lines, vortices) while providing per-pixel uncertainty maps through ensemble aggregation.

Conclusion: The method enables enhanced convection diagnostics, supports real-time data assimilation, and improves long-term climate reanalysis by combining accurate reconstruction with uncertainty quantification.

Abstract: Accurate atmospheric profiles from remote sensing instruments such as Doppler Lidar, Radar, and radiometers are frequently corrupted by low-SNR (Signal to Noise Ratio) gates, range folding, and spurious discontinuities. Traditional gap filling blurs fine-scale structures, whereas deep models lack confidence estimates. We present CuMoLoS-MAE, a Curriculum-Guided Monte Carlo Stochastic Ensemble Masked Autoencoder designed to (i) restore fine-scale features such as updraft and downdraft cores, shear lines, and small vortices, (ii) learn a data-driven prior over atmospheric fields, and (iii) quantify pixel-wise uncertainty. During training, CuMoLoS-MAE employs a mask-ratio curriculum that forces a ViT decoder to reconstruct from progressively sparser context. At inference, we approximate the posterior predictive by Monte Carlo over random mask realisations, evaluating the MAE multiple times and aggregating the outputs to obtain the posterior predictive mean reconstruction together with a finely resolved per-pixel uncertainty map. Together with high-fidelity reconstruction, this novel deep learning-based workflow enables enhanced convection diagnostics, supports real-time data assimilation, and improves long-term climate reanalysis.

Method: Uses deep neural networks to approximate normal agent behavior as parametric multivariate Gaussian distributions, defines normality score, and employs two-sided CUSUM procedure for real-time anomaly detection

Result: Achieves AUC-ROC scores over 0.95 against most impactful attacks in all evaluated PettingZoo benchmarks, outperforming discrete counterparts

Conclusion: Proposed method is effective for detecting adversarial attacks in continuous action space multi-agent systems using decentralized statistical characterization

Abstract: We address the problem of detecting adversarial attacks against cooperative multi-agent reinforcement learning with continuous action space. We propose a decentralized detector that relies solely on the local observations of the agents and makes use of a statistical characterization of the normal behavior of observable agents. The proposed detector utilizes deep neural networks to approximate the normal behavior of agents as parametric multivariate Gaussian distributions. Based on the predicted density functions, we define a normality score and provide a characterization of its mean and variance. This characterization allows us to employ a two-sided CUSUM procedure for detecting deviations of the normality score from its mean, serving as a detector of anomalous behavior in real-time. We evaluate our scheme on various multi-agent PettingZoo benchmarks against different state-of-the-art attack methods, and our results demonstrate the effectiveness of our method in detecting impactful adversarial attacks. Particularly, it outperforms the discrete counterpart by achieving AUC-ROC scores of over 0.95 against the most impactful attacks in all evaluated environments.

Method: Integrated GANs for dynamic image generation, RL for adaptive difficulty tuning based on user behavior, and LLMs for text/audio prompts. Uses 3x3 grid image selections and combined audio tasks.

Result: 92% human success rate and only 10% bot bypass rate in real-world evaluations, significantly outperforming existing CAPTCHA systems.

Conclusion: Provides a robust, scalable approach for securing online applications while maintaining user accessibility, addressing gaps in previous research.

Abstract: Aura-CAPTCHA was developed as a multi-modal CAPTCHA system to address vulnerabilities in traditional methods that are increasingly bypassed by AI technologies, such as Optical Character Recognition (OCR) and adversarial image processing. The design integrated Generative Adversarial Networks (GANs) for generating dynamic image challenges, Reinforcement Learning (RL) for adaptive difficulty tuning, and Large Language Models (LLMs) for creating text and audio prompts. Visual challenges included 3x3 grid selections with at least three correct images, while audio challenges combined randomized numbers and words into a single task. RL adjusted difficulty based on incorrect attempts, response time, and suspicious user behavior. Evaluations on real-world traffic demonstrated a 92% human success rate and a 10% bot bypass rate, significantly outperforming existing CAPTCHA systems. The system provided a robust and scalable approach for securing online applications while remaining accessible to users, addressing gaps highlighted in previous research.

Method: Use generative equilibrium operators (GEOs) with finite-dimensional deep equilibrium layers to solve convex optimization problems over separable Hilbert spaces, where inputs are smooth convex loss functions and outputs are approximate solutions.

Result: When input losses lie in suitable infinite-dimensional compact sets, GEOs can uniformly approximate solutions with arbitrary precision, requiring only logarithmic growth in rank, depth, and width relative to approximation error. Validated on nonlinear PDEs, stochastic optimal control, and financial hedging problems.

Conclusion: The paper successfully closes the theory-practice gap for neural operators by demonstrating that GEOs achieve efficient approximation with practical parameter requirements, making them viable for real-world operator learning applications across multiple domains.

Abstract: Neural operators (NOs) are a class of deep learning models designed to simultaneously solve infinitely many related problems by casting them into an infinite-dimensional space, whereon these NOs operate. A significant gap remains between theory and practice: worst-case parameter bounds from universal approximation theorems suggest that NOs may require an unrealistically large number of parameters to solve most operator learning problems, which stands in direct opposition to a slew of experimental evidence. This paper closes that gap for a specific class of {NOs}, generative {equilibrium operators} (GEOs), using (realistic) finite-dimensional deep equilibrium layers, when solving families of convex optimization problems over a separable Hilbert space $X$. Here, the inputs are smooth, convex loss functions on $X$, and outputs are the associated (approximate) solutions to the optimization problem defined by each input loss. We show that when the input losses lie in suitable infinite-dimensional compact sets, our GEO can uniformly approximate the corresponding solutions to arbitrary precision, with rank, depth, and width growing only logarithmically in the reciprocal of the approximation error. We then validate both our theoretical results and the trainability of GEOs on three applications: (1) nonlinear PDEs, (2) stochastic optimal control problems, and (3) hedging problems in mathematical finance under liquidity constraints.

Method: Case study comparing performance of different LLMs (Claude-3.5-Sonnet, Claude-3.7-Sonnet, GPT-4o, Llama-3.1 models, GPT-3.5-Turbo, Nova-Micro) and agent types (tool-calling vs code-generating) using LLM-as-a-judge score to evaluate their effectiveness in orchestrating chemistry and drug discovery tools.

Result: Claude-3.5-Sonnet, Claude-3.7-Sonnet and GPT-4o outperformed other models. Code-generating agents generally outperformed tool-calling ones, but performance was highly question- and model-dependent. System prompt replacement impact also varied by question and model.

Conclusion: Components of LLM-based agentic systems cannot be simply replaced without re-engineering, highlighting the need for further research into modularity to develop reliable and modular solutions for real-world drug discovery problems.

Abstract: Large-language models (LLMs) and agentic systems present exciting opportunities to accelerate drug discovery. In this study, we examine the modularity of LLM-based agentic systems for drug discovery, i.e., whether parts of the system such as the LLM and type of agent are interchangeable, a topic that has received limited attention in drug discovery. We compare the performance of different LLMs and the effectiveness of tool-calling agents versus code-generating agents. Our case study, comparing performance in orchestrating tools for chemistry and drug discovery using an LLM-as-a-judge score, shows that Claude-3.5-Sonnet, Claude-3.7-Sonnet and GPT-4o outperform alternative language models such as Llama-3.1-8B, Llama-3.1-70B, GPT-3.5-Turbo, and Nova-Micro. Although we confirm that code-generating agents outperform the tool-calling ones on average, we show that this is highly question- and model-dependent. Furthermore, the impact of replacing system prompts is dependent on the question and model, underscoring that even in this particular domain one cannot just replace components of the system without re-engineering. Our study highlights the necessity of further research into the modularity of agentic systems to enable the development of reliable and modular solutions for real-world problems.

Method: Systematic review and hardware-centric analysis of quantization techniques, evaluation of existing software frameworks and hardware platforms for QNN execution on microcontrollers.

Result: Comprehensive survey of essential quantization methods for accelerating deep learning models on embedded applications, with analysis of performance-hardware trade-offs.

Conclusion: Identifies current challenges and outlines promising future directions in QNN deployment for TinyML applications on resource-constrained devices.

Abstract: The deployment of Quantized Neural Networks (QNNs) on resource-constrained devices, such as microcontrollers, has introduced significant challenges in balancing model performance, computational complexity and memory constraints. Tiny Machine Learning (TinyML) addresses these issues by integrating advancements across machine learning algorithms, hardware acceleration, and software optimization to efficiently run deep neural networks on embedded systems. This survey presents a hardware-centric introduction to quantization, systematically reviewing essential quantization techniques employed to accelerate deep learning models for embedded applications. In particular, further emphasis is put on critical trade-offs among model performance and hardware capabilities. The survey further evaluates existing software frameworks and hardware platforms designed specifically for supporting QNN execution on microcontrollers. Moreover, we provide an analysis of the current challenges and an outline of promising future directions in the rapidly evolving domain of QNN deployment.

Method: Combines static compiler analysis to identify tensor dimensions requiring identical sharding and partitioning conflicts, with Monte Carlo Tree Search to efficiently explore the decision space.

Result: Significantly outperforms state-of-the-art industrial methods across diverse hardware platforms and model architectures, discovering previously unknown superior solutions with full automation.

Conclusion: The proposed system provides an effective automated solution for partitioning complex ML models across distributed accelerators, overcoming memory constraints and achieving optimal performance.

Abstract: Partitioning large machine learning models across distributed accelerator systems is a complex process, requiring a series of interdependent decisions that are further complicated by internal sharding ambiguities. Consequently, existing auto-partitioners often suffer from out-of-memory errors or are prohibitively slow when exploring the exponentially large space of possible partitionings. To mitigate this, they artificially restrict the search space, but this approach frequently yields infeasible solutions that violate device memory constraints or lead to sub-optimal performance. We propose a system that combines a novel static compiler analysis with a Monte Carlo Tree Search. Our analysis constructs an efficient decision space by identifying (i) tensor dimensions requiring identical sharding, and (ii) partitioning “conflicts” that require resolution. Our system significantly outperforms state-of-the-art industrial methods across diverse hardware platforms and model architectures, discovering previously unknown, superior solutions, and the process is fully automated even for complex and large models.

Method: SEAL decomposes input molecular graphs into chemically relevant fragments and estimates their causal influence on outputs. It explicitly reduces inter-fragment message passing in the model architecture to achieve strong alignment between fragment contributions and predictions.

Result: Extensive evaluations show SEAL outperforms other explainability methods in quantitative attribution metrics and human-aligned interpretability. A user study confirms it provides more intuitive and trustworthy explanations to domain experts.

Conclusion: SEAL bridges the gap between predictive performance and interpretability, offering a promising direction for more transparent and actionable molecular modeling.

Abstract: Graph neural networks have demonstrated remarkable success in predicting molecular properties by leveraging the rich structural information encoded in molecular graphs. However, their black-box nature reduces interpretability, which limits trust in their predictions for important applications such as drug discovery and materials design. Furthermore, existing explanation techniques often fail to reliably quantify the contribution of individual atoms or substructures due to the entangled message-passing dynamics. We introduce SEAL (Substructure Explanation via Attribution Learning), a new interpretable graph neural network that attributes model predictions to meaningful molecular subgraphs. SEAL decomposes input graphs into chemically relevant fragments and estimates their causal influence on the output. The strong alignment between fragment contributions and model predictions is achieved by explicitly reducing inter-fragment message passing in our proposed model architecture. Extensive evaluations on synthetic benchmarks and real-world molecular datasets demonstrate that SEAL outperforms other explainability methods in both quantitative attribution metrics and human-aligned interpretability. A user study further confirms that SEAL provides more intuitive and trustworthy explanations to domain experts. By bridging the gap between predictive performance and interpretability, SEAL offers a promising direction for more transparent and actionable molecular modeling.

Method: Train two identical models on independent bootstrap samples in parallel. Use periodic mean-reset to keep both trajectories in the same basin. Use their divergence to estimate local uncertainty and sample weights adaptively for regularization favoring flatter solutions.

Result: The approach improves calibration and generalization in deep neural networks and complex high-dimensional inverse problems, producing interpretable uncertainty maps.

Conclusion: Twin-Boot successfully integrates uncertainty estimation directly into the optimization process, overcoming limitations of traditional bootstrapping while providing practical uncertainty quantification and regularization benefits.

Abstract: Standard gradient descent methods yield point estimates with no measure of confidence. This limitation is acute in overparameterized and low-data regimes, where models have many parameters relative to available data and can easily overfit. Bootstrapping is a classical statistical framework for uncertainty estimation based on resampling, but naively applying it to deep learning is impractical: it requires training many replicas, produces post-hoc estimates that cannot guide learning, and implicitly assumes comparable optima across runs - an assumption that fails in non-convex landscapes. We introduce Twin-Bootstrap Gradient Descent (Twin-Boot), a resampling-based training procedure that integrates uncertainty estimation into optimization. Two identical models are trained in parallel on independent bootstrap samples, and a periodic mean-reset keeps both trajectories in the same basin so that their divergence reflects local (within-basin) uncertainty. During training, we use this estimate to sample weights in an adaptive, data-driven way, providing regularization that favors flatter solutions. In deep neural networks and complex high-dimensional inverse problems, the approach improves calibration and generalization and yields interpretable uncertainty maps.

Method: Federated learning of linearly-parameterized nonlinear systems using theoretical analysis and experimental validation with physical systems driven by i.i.d. control inputs and random perturbations. Experiments use nonlinear dynamical systems with real-analytic feature functions including polynomial and trigonometric components.

Result: Convergence rate improves as number of clients increases. Federated learning consistently improves convergence of individual clients across varying noise levels and data distributions. The constant difference in convergence rates between linear and nonlinear cases depends on the feature map choice.

Conclusion: Federated nonlinear system identification is effective and outperforms centralized approaches, with careful feature map selection enabling increased excitation and improved performance. More clients lead to better convergence for all participants.

Abstract: We consider federated learning of linearly-parameterized nonlinear systems. We establish theoretical guarantees on the effectiveness of federated nonlinear system identification compared to centralized approaches, demonstrating that the convergence rate improves as the number of clients increases. Although the convergence rates in the linear and nonlinear cases differ only by a constant, this constant depends on the feature map $\phi$, which can be carefully chosen in the nonlinear setting to increase excitation and improve performance. We experimentally validate our theory in physical settings where client devices are driven by i.i.d. control inputs and control policies exhibiting i.i.d. random perturbations, ensuring non-active exploration. Experiments use trajectories from nonlinear dynamical systems characterized by real-analytic feature functions, including polynomial and trigonometric components, representative of physical systems including pendulum and quadrotor dynamics. We analyze the convergence behavior of the proposed method under varying noise levels and data distributions. Results show that federated learning consistently improves convergence of any individual client as the number of participating clients increases.

Method: Proposes similarity-aware channel augmentation to cache channels with high augmentation sensitivity, and incorporates lossy data compression with progressive compression strategy that increases compression rates as more layers are frozen.

Result: Achieves significant reductions in training cost while maintaining model accuracy, with minor time overhead. Provides comprehensive evaluation of freezing and compression strategies.

Conclusion: The proposed solution effectively addresses overlooked challenges in feature map caching for layer freezing, enabling substantial computational savings without compromising accuracy.

Abstract: With the growing size of deep neural networks and datasets, the computational costs of training have significantly increased. The layer-freezing technique has recently attracted great attention as a promising method to effectively reduce the cost of network training. However, in traditional layer-freezing methods, frozen layers are still required for forward propagation to generate feature maps for unfrozen layers, limiting the reduction of computation costs. To overcome this, prior works proposed a hypothetical solution, which caches feature maps from frozen layers as a new dataset, allowing later layers to train directly on stored feature maps. While this approach appears to be straightforward, it presents several major challenges that are severely overlooked by prior literature, such as how to effectively apply augmentations to feature maps and the substantial storage overhead introduced. If these overlooked challenges are not addressed, the performance of the caching method will be severely impacted and even make it infeasible. This paper is the first to comprehensively explore these challenges and provides a systematic solution. To improve training accuracy, we propose \textit{similarity-aware channel augmentation}, which caches channels with high augmentation sensitivity with a minimum additional storage cost. To mitigate storage overhead, we incorporate lossy data compression into layer freezing and design a \textit{progressive compression} strategy, which increases compression rates as more layers are frozen, effectively reducing storage costs. Finally, our solution achieves significant reductions in training cost while maintaining model accuracy, with a minor time overhead. Additionally, we conduct a comprehensive evaluation of freezing and compression strategies, providing insights into optimizing their application for efficient DNN training.

Method: The study provides minimax analysis for mean estimation of multivariate bounded distributions and univariate Gaussian distributions, as well as multivariate Gaussian mean estimation and linear regression, considering heterogeneous corruption patterns with known but non-identical corruption probabilities.

Result: The paper establishes tight minimax rates for all heterogeneous corruption patterns in mean estimation problems and shows that for multivariate Gaussian mean estimation and linear regression, the minimax rate for squared error is determined up to a factor of √d (where d is dimension).

Conclusion: Optimal estimators should discard samples beyond a certain corruption threshold, which is determined by the empirical distribution of the given corruption rates, providing guidance for handling heterogeneous corruption in practical applications.

Abstract: We study the problem of robust estimation under heterogeneous corruption rates, where each sample may be independently corrupted with a known but non-identical probability. This setting arises naturally in distributed and federated learning, crowdsourcing, and sensor networks, yet existing robust estimators typically assume uniform or worst-case corruption, ignoring structural heterogeneity. For mean estimation for multivariate bounded distributions and univariate gaussian distributions, we give tight minimax rates for all heterogeneous corruption patterns. For multivariate gaussian mean estimation and linear regression, we establish the minimax rate for squared error up to a factor of $\sqrt{d}$, where $d$ is the dimension. Roughly, our findings suggest that samples beyond a certain corruption threshold may be discarded by the optimal estimators – this threshold is determined by the empirical distribution of the corruption rates given.

Method: Introduces NGN-M, a momentum-based version of the NGN step-size method. It combines momentum with adaptive step-size adaptation to achieve better stability without requiring restrictive assumptions like interpolation condition or bounded gradients.

Result: Achieves standard O(1/√K) convergence rate under less restrictive assumptions. Empirically demonstrates enhanced robustness to step-size choices while delivering comparable or superior performance to state-of-the-art optimizers.

Conclusion: NGN-M provides a more stable and robust optimization approach that reduces sensitivity to hyperparameter tuning while maintaining competitive performance with existing state-of-the-art methods.

Abstract: Modern optimization algorithms that incorporate momentum and adaptive step-size offer improved performance in numerous challenging deep learning tasks. However, their effectiveness is often highly sensitive to the choice of hyperparameters, especially the step-size. Tuning these parameters is often difficult, resource-intensive, and time-consuming. Therefore, recent efforts have been directed toward enhancing the stability of optimizers across a wide range of hyperparameter choices [Schaipp et al., 2024]. In this paper, we introduce an algorithm that matches the performance of state-of-the-art optimizers while improving stability to the choice of the step-size hyperparameter through a novel adaptation of the NGN step-size method [Orvieto and Xiao, 2024]. Specifically, we propose a momentum-based version (NGN-M) that attains the standard convergence rate of $\mathcal{O}(1/\sqrt{K})$ under less restrictive assumptions, without the need for interpolation condition or assumptions of bounded stochastic gradients or iterates, in contrast to previous approaches. Additionally, we empirically demonstrate that the combination of the NGN step-size with momentum results in enhanced robustness to the choice of the step-size hyperparameter while delivering performance that is comparable to or surpasses other state-of-the-art optimizers.

Method: Introduces an analytical framework based on maximum likelihood estimation without conventional gradient-based optimization or explicit labels. Derives a “wormhole” solution to address degeneracy issues.

Result: DNNs in overparameterized regimes show output space collapse that enhances generalization initially, but adding layers leads to degeneracy (trivial solutions with zero loss). The wormhole solution successfully bypasses this degeneracy and reconciles meaningful labels with random inputs.

Conclusion: The findings provide deeper insights into DNN generalization mechanisms, highlight shortcut learning perspectives, and suggest future research directions on learning dynamics in unsupervised settings to bridge theory-practice gaps.

Abstract: This work investigates the generalization behavior of deep neural networks (DNNs), focusing on the phenomenon of “fooling examples,” where DNNs confidently classify inputs that appear random or unstructured to humans. To explore this phenomenon, we introduce an analytical framework based on maximum likelihood estimation, without adhering to conventional numerical approaches that rely on gradient-based optimization and explicit labels. Our analysis reveals that DNNs operating in an overparameterized regime exhibit a collapse in the output feature space. While this collapse improves network generalization, adding more layers eventually leads to a state of degeneracy, where the model learns trivial solutions by mapping distinct inputs to the same output, resulting in zero loss. Further investigation demonstrates that this degeneracy can be bypassed using our newly derived “wormhole” solution. The wormhole solution, when applied to arbitrary fooling examples, reconciles meaningful labels with random ones and provides a novel perspective on shortcut learning. These findings offer deeper insights into DNN generalization and highlight directions for future research on learning dynamics in unsupervised settings to bridge the gap between theory and practice.

Method: Introduce concept separability score using Jensen-Shannon distance, evaluate Gemma-2-2B with multiple SAE variants across 5 benchmarks, and propose Attenuation via Posterior Probabilities (APP) for targeted concept suppression.

Result: SAEs reduce polysemanticity and achieve higher concept separability than base models, but greater sparsity doesn’t always improve separability and often hurts downstream performance. APP method outperforms existing approaches for targeted concept removal.

Conclusion: SAEs enable more precise concept-level control, particularly with partial suppression methods like APP, providing better interpretability and targeted intervention capabilities for LLMs.

Abstract: A key barrier to interpreting large language models is polysemanticity, where neurons activate for multiple unrelated concepts. Sparse autoencoders (SAEs) have been proposed to mitigate this issue by transforming dense activations into sparse, more interpretable features. While prior work suggests that SAEs promote monosemanticity, there has been no quantitative comparison with their base models. This paper provides the first systematic evaluation of SAEs against base models concerning monosemanticity. We introduce a fine-grained concept separability score based on the Jensen-Shannon distance, which captures how distinctly a neuron’s activation distributions vary across concepts. Using Gemma-2-2B and multiple SAE variants across five benchmarks, we show that SAEs reduce polysemanticity and achieve higher concept separability. However, greater sparsity of SAEs does not always yield better separability and often impairs downstream performance. To assess practical utility, we evaluate concept-level interventions using two strategies: full neuron masking and partial suppression. We find that, compared to base models, SAEs enable more precise concept-level control when using partial suppression. Building on this, we propose Attenuation via Posterior Probabilities (APP), a new intervention method that uses concept-conditioned activation distributions for targeted suppression. APP outperforms existing approaches in targeted concept removal.

Method: Integrates Mamba-style SSM backbone with intermittent sparse global attention, chunk-level MoE feed-forward routing, and dual memory systems (workspace + factual PKM). Uses formal component interfaces and staged curriculum training.

Result: Toy-scale prototype demonstrates implementation feasibility and qualitative scaling behaviors (long-context throughput crossover, controllable expert routing), but no competitive full-scale performance claims.

Conclusion: Hydra presents a blueprint for modular, input-adaptive long-context models combining SSM efficiency with selective attention and memory systems, requiring future validation at target scale.

Abstract: We present Hydra as an architectural proposal for hybrid long-context language models that combine conditional computation, long-context memory mechanisms, and sparse mixture-of-experts within an approximately 1.6B parameter design envelope. Hydra integrates a Mamba-style Structured State Space Model (SSM) backbone with intermittent sparse global attention, chunk-level MoE feed-forward routing, and dual (workspace plus factual PKM) memories. We formalize the component interfaces, give transparent parameter and complexity accounting, and outline a staged curriculum intended to stably activate the parts. We accompany the specification with illustrative toy-scale prototype measurements (tens of millions of parameters on synthetic data) whose sole purpose is to demonstrate implementation feasibility and qualitative scaling behaviors (for example, long-context throughput crossover and controllable expert routing), not to claim competitive full-scale performance. We explicitly delineate assumptions and open risks (training complexity, memory utilization, specialization dynamics) and position Hydra as a blueprint to stimulate empirical follow-up rather than a finished system. By combining SSM efficiency, selective sparse attention, MoE capacity, and learnable memory, Hydra sketches a path toward modular, input-adaptive long-context language models; validating end-task gains at target scale remains future work.

Method: Developed Side Effect Evaluation (SEE) benchmark with hierarchical and compositional prompts to measure CET robustness across three aspects: impact on neighboring concepts, evasion of targets, and attribute leakage.

Result: CETs can be easily circumvented using superclass-subclass hierarchy and semantically similar prompts, suffer from attribute leakage, and exhibit counterintuitive attention concentration or dispersal phenomena.

Conclusion: Current CETs have significant vulnerabilities and side effects, highlighting the need for more robust concept erasure methods. The authors release dataset, code, and evaluation tools to support future research.

Abstract: Concerns about text-to-image (T2I) generative models infringing on privacy, copyright, and safety have led to the development of Concept Erasure Techniques (CETs). The goal of an effective CET is to prohibit the generation of undesired ``target’’ concepts specified by the user, while preserving the ability to synthesize high-quality images of the remaining concepts. In this work, we demonstrate that CETs can be easily circumvented and present several side effects of concept erasure. For a comprehensive measurement of the robustness of CETs, we present Side Effect Evaluation (\see), an evaluation benchmark that consists of hierarchical and compositional prompts that describe objects and their attributes. This dataset and our automated evaluation pipeline quantify side effects of CETs across three aspects: impact on neighboring concepts, evasion of targets, and attribute leakage. Our experiments reveal that CETs can be circumvented by using superclass-subclass hierarchy and semantically similar prompts, such as compositional variants of the target. We show that CETs suffer from attribute leakage and counterintuitive phenomena of attention concentration or dispersal. We release our dataset, code, and evaluation tools to aid future work on robust concept erasure.

Method: Presents a technique to estimate the Hessian of the unknown remaining training data, which is crucial for efficient unlearning. This enables zero-shot unlearning without requiring the original training dataset.

Result: Extensive experiments across multiple datasets demonstrate the method’s effectiveness in removing specific data while maintaining performance on remaining data, with robust theoretical guarantees.

Conclusion: The proposed source-free unlearning method successfully addresses the practical challenge of data removal from trained models without access to original training data, providing both theoretical guarantees and empirical validation.

Abstract: As machine learning becomes more pervasive and data privacy regulations evolve, the ability to remove private or copyrighted information from trained models is becoming an increasingly critical requirement. Existing unlearning methods often rely on the assumption of having access to the entire training dataset during the forgetting process. However, this assumption may not hold true in practical scenarios where the original training data may not be accessible, i.e., the source-free setting. To address this challenge, we focus on the source-free unlearning scenario, where an unlearning algorithm must be capable of removing specific data from a trained model without requiring access to the original training dataset. Building on recent work, we present a method that can estimate the Hessian of the unknown remaining training data, a crucial component required for efficient unlearning. Leveraging this estimation technique, our method enables efficient zero-shot unlearning while providing robust theoretical guarantees on the unlearning performance, while maintaining performance on the remaining data. Extensive experiments over a wide range of datasets verify the efficacy of our method.

Method: Uses category theory, functors, and universal coalgebras to model RL algorithms. Reviews standard RL framework, introduces categorial models, and extends to universal coalgebras for various dynamical systems. Proposes asynchronous distributed computation for finding fixed points.

Result: Shows that RL algorithms form functor categories with topos properties. Demonstrates that MDPs, POMDPs, PSRs, and LDSs are special types of coalgebras. Generalizes fixed-point finding in RL to determining final coalgebras asynchronously.

Conclusion: URL provides a powerful categorial framework that unifies various RL models through universal coalgebras and enables distributed asynchronous computation for solving RL problems more generally.

Abstract: In this paper, we introduce a categorial generalization of RL, termed universal reinforcement learning (URL), building on powerful mathematical abstractions from the study of coinduction on non-well-founded sets and universal coalgebras, topos theory, and categorial models of asynchronous parallel distributed computation. In the first half of the paper, we review the basic RL framework, illustrate the use of categories and functors in RL, showing how they lead to interesting insights. In particular, we also introduce a standard model of asynchronous distributed minimization proposed by Bertsekas and Tsitsiklis, and describe the relationship between metric coinduction and their proof of the Asynchronous Convergence Theorem. The space of algorithms for MDPs or PSRs can be modeled as a functor category, where the co-domain category forms a topos, which admits all (co)limits, possesses a subobject classifier, and has exponential objects. In the second half of the paper, we move on to universal coalgebras. Dynamical system models, such as Markov decision processes (MDPs), partially observed MDPs (POMDPs), a predictive state representation (PSRs), and linear dynamical systems (LDSs) are all special types of coalgebras. We describe a broad family of universal coalgebras, extending the dynamic system models studied previously in RL. The core problem in finding fixed points in RL to determine the exact or approximate (action) value function is generalized in URL to determining the final coalgebra asynchronously in a parallel distributed manner.

Method: Conducted reproducibility and replicability experiments on 11 different SoTA DPML techniques from recent literature, addressing unique DPML challenges like additional randomness from DP noise.

Result: Mixed outcomes - some methods performed as claimed while others faltered when tested outside their initial experimental conditions, revealing inconsistencies in reported results.

Conclusion: Identified challenges in DPML reproducibility and derived best practices for obtaining scientifically valid and reliable results in differentially private machine learning research.

Abstract: There is a flurry of recent research papers proposing novel differentially private machine learning (DPML) techniques. These papers claim to achieve new state-of-the-art (SoTA) results and offer empirical results as validation. However, there is no consensus on which techniques are most effective or if they genuinely meet their stated claims. Complicating matters, heterogeneity in codebases, datasets, methodologies, and model architectures make direct comparisons of different approaches challenging. In this paper, we conduct a reproducibility and replicability (R+R) experiment on 11 different SoTA DPML techniques from the recent research literature. Results of our investigation are varied: while some methods stand up to scrutiny, others falter when tested outside their initial experimental conditions. We also discuss challenges unique to the reproducibility of DPML, including additional randomness due to DP noise, and how to address them. Finally, we derive insights and best practices to obtain scientifically valid and reliable results.

Method: Developed a robust system using fine-tuned RoBERTa model for automated extraction of specific cancer types from pathology reports.

Result: Model significantly outperformed baseline and Mistral 7B LLM, achieving F1_Bertscore 0.98 and 80.61% overall exact match accuracy.

Conclusion: Fine-tuning domain-specific models shows potential for scalable integration into molecular tumor board processes and more efficient clinical information extraction in oncology.

Abstract: The accurate extraction of clinical information from electronic medical records is particularly critical to clinical research but require much trained expertise and manual labor. In this study we developed a robust system for automated extraction of the specific cancer types for the purpose of supporting precision oncology research. from pathology reports using a fine-tuned RoBERTa model. This model significantly outperformed the baseline model and a Large Language Model, Mistral 7B, achieving F1_Bertscore 0.98 and overall exact match of 80.61%. This fine-tuning approach demonstrates the potential for scalability that can integrate seamlessly into the molecular tumour board process. Fine-tuning domain-specific models for precision tasks in oncology, may pave the way for more efficient and accurate clinical information extraction.

Method: Three-stage pipeline: 1) Dynamic unsafe output detection using hybrid external classifiers and internal evaluations, 2) Token-level harmful content tracing through FFN activations, 3) Constrained optimization to suppress unsafe behavior without degrading model quality.

Result: Substantially reduces attack success rates on Vicuna, LLaMA, and GPT-J across multiple jailbreak benchmarks while maintaining high general-purpose performance and robustness to unseen attacks.

Conclusion: Unlearning is a promising direction for scalable and effective LLM safety, offering stronger safety guarantees and more precise control over harmful behavior compared to standard defense methods.

Abstract: Jailbreak attacks pose a serious threat to the safety of Large Language Models (LLMs) by crafting adversarial prompts that bypass alignment mechanisms, causing the models to produce harmful, restricted, or biased content. In this paper, we propose SafeLLM, a novel unlearning-based defense framework that unlearn the harmful knowledge from LLMs while preserving linguistic fluency and general capabilities. SafeLLM employs a three-stage pipeline: (1) dynamic unsafe output detection using a hybrid approach that integrates external classifiers with model-internal evaluations; (2) token-level harmful content tracing through feedforward network (FFN) activations to localize harmful knowledge; and (3) constrained optimization to suppress unsafe behavior without degrading overall model quality. SafeLLM achieves targeted and irreversible forgetting by identifying and neutralizing FFN substructures responsible for harmful generation pathways. Extensive experiments on prominent LLMs (Vicuna, LLaMA, and GPT-J) across multiple jailbreak benchmarks show that SafeLLM substantially reduces attack success rates while maintaining high general-purpose performance. Compared to standard defense methods such as supervised fine-tuning and direct preference optimization, SafeLLM offers stronger safety guarantees, more precise control over harmful behavior, and greater robustness to unseen attacks. Moreover, SafeLLM maintains the general performance after the harmful knowledge unlearned. These results highlight unlearning as a promising direction for scalable and effective LLM safety.

Method: Built on AIF360 platform, FairPrep provides an extensible and modular framework with seamless integration of datasets, fairness interventions, and predictive models. It features batch-processing interface for efficient experimentation and automatic reporting of fairness and utility metrics.

Result: The framework enables standardized pipelines and reproducible evaluations of pre-processing fairness techniques, filling a critical gap in fairness benchmarking.

Conclusion: FairPrep provides a practical foundation for advancing data-level fairness research by offering a comprehensive benchmarking solution for pre-processing methods.

Abstract: As machine learning systems become increasingly integrated into high-stakes decision-making processes, ensuring fairness in algorithmic outcomes has become a critical concern. Methods to mitigate bias typically fall into three categories: pre-processing, in-processing, and post-processing. While significant attention has been devoted to the latter two, pre-processing methods, which operate at the data level and offer advantages such as model-agnosticism and improved privacy compliance, have received comparatively less focus and lack standardised evaluation tools. In this work, we introduce FairPrep, an extensible and modular benchmarking framework designed to evaluate fairness-aware pre-processing techniques on tabular datasets. Built on the AIF360 platform, FairPrep allows seamless integration of datasets, fairness interventions, and predictive models. It features a batch-processing interface that enables efficient experimentation and automatic reporting of fairness and utility metrics. By offering standardised pipelines and supporting reproducible evaluations, FairPrep fills a critical gap in the fairness benchmarking landscape and provides a practical foundation for advancing data-level fairness research.

Method: Proposed frequency-adaptive TNNs by incorporating random Fourier features and leveraging TNNs’ tensor structure to extract frequency features via Discrete Fourier Transform on one-dimensional component functions, mitigating dimensionality curse.

Result: Extensive numerical experiments validated the effectiveness and robustness of the frequency-adaptive TNNs algorithm in solving complex multi-scale problems.

Conclusion: The proposed frequency-adaptive approach significantly improves TNNs’ ability to handle high-dimensional multi-scale problems by enhancing their expressivity for high-frequency features.

Abstract: Tensor neural networks (TNNs) have demonstrated their superiority in solving high-dimensional problems. However, similar to conventional neural networks, TNNs are also influenced by the Frequency Principle, which limits their ability to accurately capture high-frequency features of the solution. In this work, we analyze the training dynamics of TNNs by Fourier analysis and enhance their expressivity for high-dimensional multi-scale problems by incorporating random Fourier features. Leveraging the inherent tensor structure of TNNs, we further propose a novel approach to extract frequency features of high-dimensional functions by performing the Discrete Fourier Transform to one-dimensional component functions. This strategy effectively mitigates the curse of dimensionality. Building on this idea, we propose a frequency-adaptive TNNs algorithm, which significantly improves the ability of TNNs in solving complex multi-scale problems. Extensive numerical experiments are performed to validate the effectiveness and robustness of the proposed frequency-adaptive TNNs algorithm.

Method: Multivariate Differential Transformer (SleepDIFFormer) with cross-domain alignment to process EEG and EOG signals, mitigating spatial/temporal attention noise and learning domain-invariant joint representations through feature distribution alignment.

Result: Achieved state-of-the-art performance on five different sleep staging datasets, with thorough ablation analyses and interpretation of differential attention weights showing relevance to characteristic sleep EEG patterns.

Conclusion: The method advances automated sleep stage classification and has implications for sleep quality assessment applications, with publicly available source code.

Abstract: Classification of sleep stages is essential for assessing sleep quality and diagnosing sleep disorders such as insomnia. However, manual inspection of EEG characteristics for each stage is time-consuming and prone to human error. Although machine learning and deep learning methods have been actively developed, they continue to face challenges from the non-stationarity and variability of electroencephalography (EEG) and electrooculography (EOG) signals, often leading to poor generalization on unseen datasets. This research proposed a Sleep Stage Classification method by developing Multivariate Differential Transformer (SleepDIFFormer) for joint EEG and EOG representation learning. Specifically, SleepDIFFormer was developed to process EEG and EOG signals using our Multivariate Differential Transformer Architecture (MDTA) for time series, trained with cross-domain alignment. Our method mitigated spatial and temporal attention noise while learning a domain-invariant joint EEG-EOG representation through feature distribution alignment, thereby enabling generalization to unseen target datasets. Empirically, we evaluated our method on five different sleep staging datasets and compared it with existing approaches, achieving state-of-the-art performance. We also conducted thorough ablation analyses of SleepDIFFormer and interpreted the differential attention weights, highlighting their relevance to characteristic sleep EEG patterns. These findings have implications for advancing automated sleep stage classification and its application to sleep quality assessment. Our source code is publicly available at https://github.com/Ben1001409/SleepDIFFormer

Method: Proposes MAL framework with two components: Attribution Knowledge Aggregator (multi-task learner integrating diverse attribution labels) and Primary Target Predictor (generating calibrated probabilities for system-optimized attribution). Also introduces CAT training strategy using Cartesian product of attribution labels for enriched supervision.

Result: Achieved +0.51% GAUC improvement on offline metrics and +2.6% increase in ROI in online experiments, demonstrating superior performance over single-attribution baselines.

Conclusion: Integrating multiple attribution perspectives through the MAL framework significantly enhances CVR prediction performance and provides direct compatibility with industrial deployment requirements while improving return on investment.

Abstract: Conversion rate (CVR) prediction is a core component of online advertising systems, where the attribution mechanisms-rules for allocating conversion credit across user touchpoints-fundamentally determine label generation and model optimization. While many industrial platforms support diverse attribution mechanisms (e.g., First-Click, Last-Click, Linear, and Data-Driven Multi-Touch Attribution), conventional approaches restrict model training to labels from a single production-critical attribution mechanism, discarding complementary signals in alternative attribution perspectives. To address this limitation, we propose a novel Multi-Attribution Learning (MAL) framework for CVR prediction that integrates signals from multiple attribution perspectives to better capture the underlying patterns driving user conversions. Specifically, MAL is a joint learning framework consisting of two core components: the Attribution Knowledge Aggregator (AKA) and the Primary Target Predictor (PTP). AKA is implemented as a multi-task learner that integrates knowledge extracted from diverse attribution labels. PTP, in contrast, focuses on the task of generating well-calibrated conversion probabilities that align with the system-optimized attribution metric (e.g., CVR under the Last-Click attribution), ensuring direct compatibility with industrial deployment requirements. Additionally, we propose CAT, a novel training strategy that leverages the Cartesian product of all attribution label combinations to generate enriched supervision signals. This design substantially enhances the performance of the attribution knowledge aggregator. Empirical evaluations demonstrate the superiority of MAL over single-attribution learning baselines, achieving +0.51% GAUC improvement on offline metrics. Online experiments demonstrate that MAL achieved a +2.6% increase in ROI (Return on Investment).

Method: Uses multi-objective learning/search techniques to generate Pareto-optimal candidates, then verifies local optimality for each candidate as a Boolean satisfiability problem solved with a SAT solver.

Result: The approach yields interpretations that closely match those synthesized by methods offering global guarantees, while being more scalable.

Conclusion: The framework enables more scalable synthesis of interpretations with local optimality guarantees, effectively addressing the trade-off between accuracy and explainability in black-box ML model interpretation.

Abstract: Creating meaningful interpretations for black-box machine learning models involves balancing two often conflicting objectives: accuracy and explainability. Exploring the trade-off between these objectives is essential for developing trustworthy interpretations. While many techniques for multi-objective interpretation synthesis have been developed, they typically lack formal guarantees on the Pareto-optimality of the results. Methods that do provide such guarantees, on the other hand, often face severe scalability limitations when exploring the Pareto-optimal space. To address this, we develop a framework based on local optimality guarantees that enables more scalable synthesis of interpretations. Specifically, we consider the problem of synthesizing a set of Pareto-optimal interpretations with local optimality guarantees, within the immediate neighborhood of each solution. Our approach begins with a multi-objective learning or search technique, such as Multi-Objective Monte Carlo Tree Search, to generate a best-effort set of Pareto-optimal candidates with respect to accuracy and explainability. We then verify local optimality for each candidate as a Boolean satisfiability problem, which we solve using a SAT solver. We demonstrate the efficacy of our approach on a set of benchmarks, comparing it against previous methods for exploring the Pareto-optimal front of interpretations. In particular, we show that our approach yields interpretations that closely match those synthesized by methods offering global guarantees.

Method: Extracts multiple temporal windows from each ECG instance to construct positive pairs and maximize their agreement via statistics, inspired by slow feature analysis to learn temporally invariant and physiologically meaningful features.

Result: Outperforms conventional two-view methods in multi-label superclass classification on PTB-XL dataset with higher AUROC (0.891 vs 0.888) and F1 scores (0.680 vs 0.679), while requiring 4x fewer pre-training epochs (32 vs 128) and 14.8% total time reduction.

Conclusion: Poly-window contrastive learning is a highly efficient and scalable paradigm for automated ECG analysis and provides a promising general framework for self-supervised representation learning in biomedical time-series data.

Abstract: Electrocardiogram (ECG) analysis is foundational for cardiovascular disease diagnosis, yet the performance of deep learning models is often constrained by limited access to annotated data. Self-supervised contrastive learning has emerged as a powerful approach for learning robust ECG representations from unlabeled signals. However, most existing methods generate only pairwise augmented views and fail to leverage the rich temporal structure of ECG recordings. In this work, we present a poly-window contrastive learning framework. We extract multiple temporal windows from each ECG instance to construct positive pairs and maximize their agreement via statistics. Inspired by the principle of slow feature analysis, our approach explicitly encourages the model to learn temporally invariant and physiologically meaningful features that persist across time. We validate our approach through extensive experiments and ablation studies on the PTB-XL dataset. Our results demonstrate that poly-window contrastive learning consistently outperforms conventional two-view methods in multi-label superclass classification, achieving higher AUROC (0.891 vs. 0.888) and F1 scores (0.680 vs. 0.679) while requiring up to four times fewer pre-training epochs (32 vs. 128) and 14.8% in total wall clock pre-training time reduction. Despite processing multiple windows per sample, we achieve a significant reduction in the number of training epochs and total computation time, making our method practical for training foundational models. Through extensive ablations, we identify optimal design choices and demonstrate robustness across various hyperparameters. These findings establish poly-window contrastive learning as a highly efficient and scalable paradigm for automated ECG analysis and provide a promising general framework for self-supervised representation learning in biomedical time-series data.

Method: Leverages model-internal confidence signals to dynamically filter out low-quality reasoning traces during or after generation, requiring no additional training or hyperparameter tuning.

Result: Achieves up to 99.9% accuracy on AIME 2025 benchmark and reduces generated tokens by up to 84.7% compared to full parallel thinking across various reasoning tasks and models.

Conclusion: DeepConf provides a simple yet powerful method that enhances both reasoning efficiency and performance at test time, seamlessly integrating into existing serving frameworks.

Abstract: Large Language Models (LLMs) have shown great potential in reasoning tasks through test-time scaling methods like self-consistency with majority voting. However, this approach often leads to diminishing returns in accuracy and high computational overhead. To address these challenges, we introduce Deep Think with Confidence (DeepConf), a simple yet powerful method that enhances both reasoning efficiency and performance at test time. DeepConf leverages model-internal confidence signals to dynamically filter out low-quality reasoning traces during or after generation. It requires no additional model training or hyperparameter tuning and can be seamlessly integrated into existing serving frameworks. We evaluate DeepConf across a variety of reasoning tasks and the latest open-source models, including Qwen 3 and GPT-OSS series. Notably, on challenging benchmarks such as AIME 2025, DeepConf@512 achieves up to 99.9% accuracy and reduces generated tokens by up to 84.7% compared to full parallel thinking.

Method: Examined CSG in multi-relational link prediction with transformer embeddings, introduced structural/semantic KG complexity metrics (Relation Entropy, Maximum Relation Diversity, Relation Type Cardinality), and benchmarked against standard performance metrics (MRR, Hit@1, Hit@10).

Result: CSG was highly sensitive to parametrization and did not scale robustly with class cardinality. New metrics showed strong inverse correlations with MRR/Hit@1, while graph connectivity measures correlated positively with Hit@10.

Conclusion: CSG’s claimed stability and predictive power fail in link prediction settings. Structural/semantic metrics like Relation Entropy and Maximum Relation Diversity provide more faithful indicators of dataset complexity and task difficulty.

Abstract: Understanding dataset complexity is fundamental to evaluating and comparing link prediction models on knowledge graphs (KGs). While the Cumulative Spectral Gradient (CSG) metric, derived from probabilistic divergence between classes within a spectral clustering framework, has been proposed as a classifier agnostic complexity metric purportedly scaling with class cardinality and correlating with downstream performance, it has not been evaluated in KG settings so far. In this work, we critically examine CSG in the context of multi relational link prediction, incorporating semantic representations via transformer derived embeddings. Contrary to prior claims, we find that CSG is highly sensitive to parametrisation and does not robustly scale with the number of classes. Moreover, it exhibits weak or inconsistent correlation with standard performance metrics such as Mean Reciprocal Rank (MRR) and Hit@1. To deepen the analysis, we introduce and benchmark a set of structural and semantic KG complexity metrics. Our findings reveal that global and local relational ambiguity captured via Relation Entropy, node level Maximum Relation Diversity, and Relation Type Cardinality exhibit strong inverse correlations with MRR and Hit@1, suggesting these as more faithful indicators of task difficulty. Conversely, graph connectivity measures such as Average Degree, Degree Entropy, PageRank, and Eigenvector Centrality correlate positively with Hit@10. Our results demonstrate that CSGs purported stability and generalization predictive power fail to hold in link prediction settings and underscore the need for more stable, interpretable, and task-aligned measures of dataset complexity in knowledge driven learning.

Method: PL-Reg uses partial-logic regularization that allows models to reserve space for undefined logic formulas, enabling better adaptability to unknown classes. The approach is formally demonstrated to work with partial logic for unknown class tasks.

Result: Extensive experiments on Generalized Category Discovery, Multi-Domain Generalized Category Discovery, and long-tailed Class Incremental Learning tasks show consistent performance improvements over existing methods.

Conclusion: Partial logic regularization effectively addresses challenges related to unknown classes in visual classification, providing improved generalization capabilities compared to fully defined logic approaches.

Abstract: Generalization remains a significant challenge in visual classification tasks, particularly in handling unknown classes in real-world applications. Existing research focuses on the class discovery paradigm, which tends to favor known classes, and the incremental learning paradigm, which suffers from catastrophic forgetting. Recent approaches such as the L-Reg technique employ logic-based regularization to enhance generalization but are bound by the necessity of fully defined logical formulas, limiting flexibility for unknown classes. This paper introduces PL-Reg, a novel partial-logic regularization term that allows models to reserve space for undefined logic formulas, improving adaptability to unknown classes. Specifically, we formally demonstrate that tasks involving unknown classes can be effectively explained using partial logic. We also prove that methods based on partial logic lead to improved generalization. We validate PL-Reg through extensive experiments on Generalized Category Discovery, Multi-Domain Generalized Category Discovery, and long-tailed Class Incremental Learning tasks, demonstrating consistent performance improvements. Our results highlight the effectiveness of partial logic in tackling challenges related to unknown classes.

Method: A hybrid transformer-based architecture that combines textual and tabular data, incorporating metadata, domain knowledge, and user profiling through a cyclical process that dynamically updates with evolving user behaviors.

Result: Experiments show robust performance on benchmark datasets, with ablation studies confirming the benefits of text-tabular integration and Explainable AI techniques providing insight into prediction rationale.

Conclusion: ExBigBang successfully addresses the limitations of previous persona classification models by providing contextual understanding, dynamic updating, and explainable predictions for improved user-centric design.

Abstract: In user-centric design, persona development plays a vital role in understanding user behaviour, capturing needs, segmenting audiences, and guiding design decisions. However, the growing complexity of user interactions calls for a more contextualized approach to ensure designs align with real user needs. While earlier studies have advanced persona classification by modelling user behaviour, capturing contextual information, especially by integrating textual and tabular data, remains a key challenge. These models also often lack explainability, leaving their predictions difficult to interpret or justify. To address these limitations, we present ExBigBang (Explainable BigBang), a hybrid text-tabular approach that uses transformer-based architectures to model rich contextual features for persona classification. ExBigBang incorporates metadata, domain knowledge, and user profiling to embed deeper context into predictions. Through a cyclical process of user profiling and classification, our approach dynamically updates to reflect evolving user behaviours. Experiments on a benchmark persona classification dataset demonstrate the robustness of our model. An ablation study confirms the benefits of combining text and tabular data, while Explainable AI techniques shed light on the rationale behind the model’s predictions.

Method: Two-dimensional time series forecasting model that integrates cohort behavior patterns over time, tested on multiple real-world datasets.

Result: Demonstrated superior performance in accuracy and adaptability compared to reference models across multiple real-world datasets.

Conclusion: The approach provides valuable insights for strategic decision-making across industries facing financial and marketing forecasting challenges, particularly effective in small data scenarios.

Abstract: This paper introduces a novel two-dimensional (2D) time series forecasting model that integrates cohort behavior over time, addressing challenges in small data environments. We demonstrate its efficacy using multiple real-world datasets, showcasing superior performance in accuracy and adaptability compared to reference models. The approach offers valuable insights for strategic decision-making across industries facing financial and marketing forecasting challenges.

Method: Proposed three practical, model-agnostic methods for Algorithmic Collective Action, where coordinated minority groups strategically relabel their own data to approximate ideal relabeling without altering the firm’s training process.

Result: Validation on real-world datasets shows that a subgroup of the minority can substantially reduce unfairness with only a small impact on overall prediction error.

Conclusion: End-user collective action through strategic data relabeling provides an effective, practical approach to enhance algorithmic fairness without requiring changes to the firm’s training procedures.

Abstract: Machine learning models often preserve biases present in training data, leading to unfair treatment of certain minority groups. Despite an array of existing firm-side bias mitigation techniques, they typically incur utility costs and require organizational buy-in. Recognizing that many models rely on user-contributed data, end-users can induce fairness through the framework of Algorithmic Collective Action, where a coordinated minority group strategically relabels its own data to enhance fairness, without altering the firm’s training process. We propose three practical, model-agnostic methods to approximate ideal relabeling and validate them on real-world datasets. Our findings show that a subgroup of the minority can substantially reduce unfairness with a small impact on the overall prediction error.

Method: EvoFormer uses: 1) Structure-Aware Transformer Module with node structural role positional encoding to mitigate structural bias; 2) Evolution-Sensitive Temporal Module with three-step strategy: timestamp classification, graph-level temporal segmentation, and segment-aware temporal self-attention with edge evolution prediction.

Result: Extensive evaluations on five benchmark datasets show state-of-the-art performance in graph similarity ranking, temporal anomaly detection, and temporal segmentation tasks.

Conclusion: EvoFormer effectively corrects structural and temporal biases in dynamic graph representation learning, demonstrating superior performance across multiple tasks by addressing key limitations of existing methods.

Abstract: Dynamic graph-level embedding aims to capture structural evolution in networks, which is essential for modeling real-world scenarios. However, existing methods face two critical yet under-explored issues: Structural Visit Bias, where random walk sampling disproportionately emphasizes high-degree nodes, leading to redundant and noisy structural representations; and Abrupt Evolution Blindness, the failure to effectively detect sudden structural changes due to rigid or overly simplistic temporal modeling strategies, resulting in inconsistent temporal embeddings. To overcome these challenges, we propose EvoFormer, an evolution-aware Transformer framework tailored for dynamic graph-level representation learning. To mitigate Structural Visit Bias, EvoFormer introduces a Structure-Aware Transformer Module that incorporates positional encoding based on node structural roles, allowing the model to globally differentiate and accurately represent node structures. To overcome Abrupt Evolution Blindness, EvoFormer employs an Evolution-Sensitive Temporal Module, which explicitly models temporal evolution through a sequential three-step strategy: (I) Random Walk Timestamp Classification, generating initial timestamp-aware graph-level embeddings; (II) Graph-Level Temporal Segmentation, partitioning the graph stream into segments reflecting structurally coherent periods; and (III) Segment-Aware Temporal Self-Attention combined with an Edge Evolution Prediction task, enabling the model to precisely capture segment boundaries and perceive structural evolution trends, effectively adapting to rapid temporal shifts. Extensive evaluations on five benchmark datasets confirm that EvoFormer achieves state-of-the-art performance in graph similarity ranking, temporal anomaly detection, and temporal segmentation tasks, validating its effectiveness in correcting structural and temporal biases.

Method: Created CITE dataset with over 438K nodes and 1.2M edges spanning four relation types, established standardized evaluation procedures, and conducted benchmarking on node classification tasks comparing four learning paradigms.

Result: The paper provides comprehensive baseline experiments across homogeneous graph models, heterogeneous graph models, LLM-centric models, and LLM+Graph models, including ablation studies on heterogeneous and textual properties.

Conclusion: CITE serves as a valuable benchmark resource that enables standardized evaluation and comparison of diverse modeling approaches for heterogeneous text-attributed graphs in materials science domain.

Abstract: Text-attributed graphs(TAGs) are pervasive in real-world systems,where each node carries its own textual features. In many cases these graphs are inherently heterogeneous, containing multiple node types and diverse edge types. Despite the ubiquity of such heterogeneous TAGs, there remains a lack of large-scale benchmark datasets. This shortage has become a critical bottleneck, hindering the development and fair comparison of representation learning methods on heterogeneous text-attributed graphs. In this paper, we introduce CITE - Catalytic Information Textual Entities Graph, the first and largest heterogeneous text-attributed citation graph benchmark for catalytic materials. CITE comprises over 438K nodes and 1.2M edges, spanning four relation types. In addition, we establish standardized evaluation procedures and conduct extensive benchmarking on the node classification task, as well as ablation experiments on the heterogeneous and textual properties of CITE. We compare four classes of learning paradigms, including homogeneous graph models, heterogeneous graph models, LLM(Large Language Model)-centric models, and LLM+Graph models. In a nutshell, we provide (i) an overview of the CITE dataset, (ii) standardized evaluation protocols, and (iii) baseline and ablation experiments across diverse modeling paradigms.

Method: Implemented an evolving fuzzy system within federated learning framework using PyTorch, allowing dynamic cluster addition without predefined cluster counts. Models are trained locally on client devices with only parameter sharing to central server.

Result: Outperformed established classification methods on multiple UCI datasets, demonstrating significant advantages in decentralized data processing despite computational intensity from overlap condition calculations.

Conclusion: The proposed evolving fuzzy system successfully addresses the challenge of adaptive clustering in federated learning environments, providing a flexible solution for decentralized data processing without requiring predefined cluster numbers.

Abstract: In this study, we present an Evolving Fuzzy System within the context of Federated Learning, which adapts dynamically with the addition of new clusters and therefore does not require the number of clusters to be selected apriori. Unlike traditional methods, Federated Learning allows models to be trained locally on clients’ devices, sharing only the model parameters with a central server instead of the data. Our method, implemented using PyTorch, was tested on clustering and classification tasks. The results show that our approach outperforms established classification methods on several well-known UCI datasets. While computationally intensive due to overlap condition calculations, the proposed method demonstrates significant advantages in decentralized data processing.

Method: Decompose the large least squares system into two smaller subproblems for branch and trunk networks separately, and solve them while updating hidden layers with gradient descent. Generalizes to L² loss with regularization and physics-informed unsupervised learning.

Result: More efficient and manageable training process that avoids direct solution of infeasibly large linear systems while maintaining optimization effectiveness.

Conclusion: The proposed hybrid approach enables practical DeepONet training by breaking down the computational bottleneck into solvable subproblems, extending applicability to regularized and physics-informed learning scenarios.

Abstract: We propose an efficient hybrid least squares/gradient descent method to accelerate DeepONet training. Since the output of DeepONet can be viewed as linear with respect to the last layer parameters of the branch network, these parameters can be optimized using a least squares (LS) solve, and the remaining hidden layer parameters are updated by means of gradient descent form. However, building the LS system for all possible combinations of branch and trunk inputs yields a prohibitively large linear problem that is infeasible to solve directly. To address this issue, our method decomposes the large LS system into two smaller, more manageable subproblems $\unicode{x2014}$ one for the branch network and one for the trunk network $\unicode{x2014}$ and solves them separately. This method is generalized to a broader type of $L^2$ loss with a regularization term for the last layer parameters, including the case of unsupervised learning with physics-informed loss.

Method: Hybrid framework that updates only the classifier layer with user-specific data using few-shot learning, implemented on RISC-V-based GAP9 microcontroller.

Result: Consistent accuracy improvements of 3.73%, 17.38%, and 3.70% across three HAR scenarios (RecGym, QVAR-Gesture, Ultrasound-Gesture) with minimal computational and memory overhead.

Conclusion: Fast, lightweight, and effective personalization is feasible on embedded platforms, enabling scalable and user-aware HAR systems.

Abstract: Human Activity Recognition (HAR) using wearable devices has advanced significantly in recent years, yet its generalization remains limited when models are deployed to new users. This degradation in performance is primarily due to user-induced concept drift (UICD), highlighting the importance of efficient personalization. In this paper, we present a hybrid framework that first generalizes across users and then rapidly adapts to individual users using few-shot learning directly on-device. By updating only the classifier layer with user-specific data, our method achieves robust personalization with minimal computational and memory overhead. We implement this framework on the energy-efficient RISC-V-based GAP9 microcontroller and validate it across three diverse HAR scenarios: RecGym, QVAR-Gesture, and Ultrasound-Gesture. Post-deployment adaptation yields consistent accuracy improvements of 3.73%, 17.38%, and 3.70% respectively. These results confirm that fast, lightweight, and effective personalization is feasible on embedded platforms, paving the way for scalable and user-aware HAR systems in the wild \footnote{https://github.com/kangpx/onlineTiny2023}.

Method: Developed a new dissimilarity measure specifically designed to detect overlap rather than dissimilarity, optimized for faster computation compared to existing methods.

Result: The proposed method is several times faster than compared methods and effectively detects overlapping clusters while avoiding the merging of orthogonal clusters.

Conclusion: The new measure provides an efficient solution for detecting overlapping clusters in streaming data applications, overcoming computational and shape-handling limitations of traditional approaches.

Abstract: In this paper, we propose a new measure for detecting overlap in multivariate Gaussian clusters. The aim of online learning from data streams is to create clustering, classification, or regression models that can adapt over time based on the conceptual drift of streaming data. In the case of clustering, this can result in a large number of clusters that may overlap and should be merged. Commonly used distribution dissimilarity measures are not adequate for determining overlapping clusters in the context of online learning from streaming data due to their inability to account for all shapes of clusters and their high computational demands. Our proposed dissimilarity measure is specifically designed to detect overlap rather than dissimilarity and can be computed faster compared to existing measures. Our method is several times faster than compared methods and is capable of detecting overlapping clusters while avoiding the merging of orthogonal clusters.

Method: Metamorphosis Representation Projection (MRP) approach that applies irreversible projection properties to machine unlearning by implementing projective transformations in the hidden state space of specific network layers.

Result: Experimental results show MRP enables effective continuous unlearning, successfully defends against relearning attacks, and achieves state-of-the-art performance in unlearning effectiveness while preserving natural performance.

Conclusion: The proposed MRP method overcomes fundamental limitations of existing approaches by completely eliminating harmful information through irreversible projections, providing a more robust solution for ensuring LLM safety through machine unlearning.

Abstract: While Large Language Models (LLMs) have demonstrated impressive performance in various domains and tasks, concerns about their safety are becoming increasingly severe. In particular, since models may store unsafe knowledge internally, machine unlearning has emerged as a representative paradigm to ensure model safety. Existing approaches employ various training techniques, such as gradient ascent and negative preference optimization, in attempts to eliminate the influence of undesired data on target models. However, these methods merely suppress the activation of undesired data through parametric training without completely eradicating its informational traces within the model. This fundamental limitation makes it difficult to achieve effective continuous unlearning, rendering these methods vulnerable to relearning attacks. To overcome these challenges, we propose a Metamorphosis Representation Projection (MRP) approach that pioneers the application of irreversible projection properties to machine unlearning. By implementing projective transformations in the hidden state space of specific network layers, our method effectively eliminates harmful information while preserving useful knowledge. Experimental results demonstrate that our approach enables effective continuous unlearning and successfully defends against relearning attacks, achieving state-of-the-art performance in unlearning effectiveness while preserving natural performance. Our code is available in https://github.com/ChengcanWu/MRP.

Method: Analysis of an exactly solvable linear-in-state, nonlinear-in-input differential equation model, examining its convergence and fading memory properties.

Result: The model shows co-existence of biologically-inspired switching behavior and mathematical stability properties suitable for stable learning on time-varying inputs.

Conclusion: This provides theoretical support for using dynamic molecular switches in neuromorphic architectures and could inspire more general exactly solvable models for brain-inspired computing.

Abstract: This paper studies an input-driven one-state differential equation model initially developed for an experimentally demonstrated dynamic molecular switch that switches like synapses in the brain do. The linear-in-the-state and nonlinear-in-the-input model is exactly solvable, and it is shown that it also possesses mathematical properties of convergence and fading memory that enable stable processing of time-varying inputs by nonlinear dynamical systems. Thus, the model exhibits the co-existence of biologically-inspired behavior and desirable mathematical properties for stable learning on sequential data. The results give theoretical support for the use of the dynamic molecular switches as computational units in deep cascaded/layered feedforward and recurrent architectures as well as other more general structures for neuromorphic computing. They could also inspire more general exactly solvable models that can be fitted to emulate arbitrary physical devices which can mimic brain-inspired behaviour and perform stable computation on input signals.

Method: Proposes group local robustness verification that leverages network computation similarity. BaVerLy dynamically constructs mini-batches of epsilon-balls with similar computations, verifies them jointly, and uses adaptive refinement when batches fail verification.

Result: BaVerLy achieves 2.3x average speedup (up to 4.1x) over traditional one-by-one verification, reducing analysis time from 24 hours to 6 hours in best cases on MNIST and CIFAR-10 networks.

Conclusion: Group-based verification through dynamic batching of similar epsilon-balls significantly improves local robustness verification efficiency while maintaining soundness and completeness.

Abstract: Neural network image classifiers are ubiquitous in many safety-critical applications. However, they are susceptible to adversarial attacks. To understand their robustness to attacks, many local robustness verifiers have been proposed to analyze $\epsilon$-balls of inputs. Yet, existing verifiers introduce a long analysis time or lose too much precision, making them less effective for a large set of inputs. In this work, we propose a new approach to local robustness: group local robustness verification. The key idea is to leverage the similarity of the network computations of certain $\epsilon$-balls to reduce the overall analysis time. We propose BaVerLy, a sound and complete verifier that boosts the local robustness verification of a set of $\epsilon$-balls by dynamically constructing and verifying mini-batches. BaVerLy adaptively identifies successful mini-batch sizes, accordingly constructs mini-batches of $\epsilon$-balls that have similar network computations, and verifies them jointly. If a mini-batch is verified, all $\epsilon$-balls are proven robust. Otherwise, one $\epsilon$-ball is suspected as not being robust, guiding the refinement. In the latter case, BaVerLy leverages the analysis results to expedite the analysis of that $\epsilon$-ball as well as the other $\epsilon$-balls in the batch. We evaluate BaVerLy on fully connected and convolutional networks for MNIST and CIFAR-10. Results show that BaVerLy scales the common one by one verification by 2.3x on average and up to 4.1x, in which case it reduces the total analysis time from 24 hours to 6 hours.

Method: HypSeek uses hyperbolic representation learning with Lorentz-model hyperbolic space, leveraging exponential geometry and negative curvature. It employs a protein-guided three-tower architecture to unify virtual screening and affinity ranking in a single framework.

Result: HypSeek improves early enrichment in virtual screening on DUD-E from 42.63 to 51.44 (+20.7%) and affinity ranking correlation on JACS from 0.5774 to 0.7239 (+25.4%), demonstrating significant performance gains.

Conclusion: Hyperbolic geometry provides a powerful inductive bias for protein-ligand modeling, enabling more expressive and affinity-sensitive embeddings that effectively capture both global activity and subtle functional differences in molecular interactions.

Abstract: Protein-ligand binding prediction is central to virtual screening and affinity ranking, two fundamental tasks in drug discovery. While recent retrieval-based methods embed ligands and protein pockets into Euclidean space for similarity-based search, the geometry of Euclidean embeddings often fails to capture the hierarchical structure and fine-grained affinity variations intrinsic to molecular interactions. In this work, we propose HypSeek, a hyperbolic representation learning framework that embeds ligands, protein pockets, and sequences into Lorentz-model hyperbolic space. By leveraging the exponential geometry and negative curvature of hyperbolic space, HypSeek enables expressive, affinity-sensitive embeddings that can effectively model both global activity and subtle functional differences-particularly in challenging cases such as activity cliffs, where structurally similar ligands exhibit large affinity gaps. Our mode unifies virtual screening and affinity ranking in a single framework, introducing a protein-guided three-tower architecture to enhance representational structure. HypSeek improves early enrichment in virtual screening on DUD-E from 42.63 to 51.44 (+20.7%) and affinity ranking correlation on JACS from 0.5774 to 0.7239 (+25.4%), demonstrating the benefits of hyperbolic geometry across both tasks and highlighting its potential as a powerful inductive bias for protein-ligand modeling.

Method: Proposes FairGuide framework with differentiable community detection as pseudo downstream task, uses meta-gradients from fairness objective to identify fairness-enhancing new links.

Result: Extensive experiments show effectiveness and generalizability across various graph-based fairness tasks, with theoretical analysis confirming fairness generalization.

Conclusion: FairGuide successfully enhances structural fairness through strategic link additions, promoting fairness generalization across diverse downstream graph applications.

Abstract: Graph Neural Networks (GNNs) have achieved remarkable success across diverse applications. However, due to the biases in the graph structures, graph neural networks face significant challenges in fairness. Although the original user graph structure is generally biased, it is promising to guide these existing structures toward unbiased ones by introducing new links. The fairness guidance via new links could foster unbiased communities, thereby enhancing fairness in downstream applications. To address this issue, we propose a novel framework named FairGuide. Specifically, to ensure fairness in downstream tasks trained on fairness-guided graphs, we introduce a differentiable community detection task as a pseudo downstream task. Our theoretical analysis further demonstrates that optimizing fairness within this pseudo task effectively enhances structural fairness, promoting fairness generalization across diverse downstream applications. Moreover, FairGuide employs an effective strategy which leverages meta-gradients derived from the fairness-guidance objective to identify new links that significantly enhance structural fairness. Extensive experimental results demonstrate the effectiveness and generalizability of our proposed method across a variety of graph-based fairness tasks.

Method: ADMM-based approach with: 1) stochastic gradients for computational efficiency, 2) multiple local training epochs between communication rounds, and 3) compressed transmissions to reduce communication overhead.

Result: The algorithm achieves exact linear convergence in strongly convex settings. Numerical experiments show superior performance compared to state-of-the-art methods on classification tasks.

Conclusion: The proposed method successfully combines computation and communication efficiency while maintaining strong theoretical convergence guarantees, making it suitable for practical distributed learning applications over networks.

Abstract: We address distributed learning problems over undirected networks. Specifically, we focus on designing a novel ADMM-based algorithm that is jointly computation- and communication-efficient. Our design guarantees computational efficiency by allowing agents to use stochastic gradients during local training. Moreover, communication efficiency is achieved as follows: i) the agents perform multiple training epochs between communication rounds, and ii) compressed transmissions are used. We prove exact linear convergence of the algorithm in the strongly convex setting. We corroborate our theoretical results by numerical comparisons with state of the art techniques on a classification task.

Method: SPOF stabilizes and perturbs a Taylor expanded polynomial approximation of the model’s training loss function, where each user’s data is privatized by adding calibrated noise to the polynomial coefficients rather than gradients.

Result: SPOF achieves up to 3.5% higher reconstruction accuracy and reduces mean training time by up to 57.2% compared to multi-user DP-SGD in WBAN scenarios with heterogeneous data and stochastic channel noise.

Conclusion: SPOF provides superior privacy-utility trade-offs in multi-user LDP settings by avoiding gradient noise injection, improving both computational efficiency and robustness to environmental noise while maintaining strong privacy guarantees.

Abstract: We propose SPOF (Stabilization of Perturbed Loss Function), a differentially private training mechanism intended for multi-user local differential privacy (LDP). SPOF perturbs a stabilized Taylor expanded polynomial approximation of a model’s training loss function, where each user’s data is privatized by calibrated noise added to the coefficients of the polynomial. Unlike gradient-based mechanisms such as differentially private stochastic gradient descent (DP-SGD), SPOF does not require injecting noise into the gradients of the loss function, which improves both computational efficiency and stability. This formulation naturally supports simultaneous privacy guarantees across all users. Moreover, SPOF exhibits robustness to environmental noise during training, maintaining stable performance even when user inputs are corrupted. We compare SPOF with a multi-user extension of DP-SGD, evaluating both methods in a wireless body area network (WBAN) scenario involving heterogeneous user data and stochastic channel noise from body sensors. Our results show that SPOF achieves, on average, up to 3.5% higher reconstruction accuracy and reduces mean training time by up to 57.2% compared to DP-SGD, demonstrating superior privacy-utility trade-offs in multi-user environments.

Method: Monitored 5 operational parameters (vibration, temperature, flow rate, pressure, current), applied dual-threshold labeling (fixed engineering limits + 95th percentile adaptive thresholds), injected synthetic fault signals using domain rules, and trained three ML classifiers (Random Forest, XGBoost, SVM) to distinguish normal operation, early warnings, and critical alerts.

Result: Random Forest and XGBoost achieved high accuracy across all classes including minority fault cases, while SVM showed lower sensitivity to anomalies. Visual analyses confirmed robust detection capabilities with the hybrid method.

Conclusion: The framework provides scalable, interpretable, and real-time industrial deployment suitable for proactive maintenance decisions, adaptable to other machinery with similar sensor architectures as a scalable predictive maintenance solution.

Abstract: This study presents a practical approach for early fault detection in industrial pump systems using real-world sensor data from a large-scale vertical centrifugal pump operating in a demanding marine environment. Five key operational parameters were monitored: vibration, temperature, flow rate, pressure, and electrical current. A dual-threshold labeling method was applied, combining fixed engineering limits with adaptive thresholds calculated as the 95th percentile of historical sensor values. To address the rarity of documented failures, synthetic fault signals were injected into the data using domain-specific rules, simulating critical alerts within plausible operating ranges. Three machine learning classifiers - Random Forest, Extreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM) - were trained to distinguish between normal operation, early warnings, and critical alerts. Results showed that Random Forest and XGBoost models achieved high accuracy across all classes, including minority cases representing rare or emerging faults, while the SVM model exhibited lower sensitivity to anomalies. Visual analyses, including grouped confusion matrices and time-series plots, indicated that the proposed hybrid method provides robust detection capabilities. The framework is scalable, interpretable, and suitable for real-time industrial deployment, supporting proactive maintenance decisions before failures occur. Furthermore, it can be adapted to other machinery with similar sensor architectures, highlighting its potential as a scalable solution for predictive maintenance in complex systems.

Method: Proposes a Bayesian approach to study algorithmic error effects, examining both LENA system and open-source Voice Type Classifier from ACLEW system, with Bayesian calibration for recovering unbiased estimates.

Result: Classification errors significantly distort estimates - automated annotations underestimated the negative effect of siblings on adult input by 20-80%, potentially making effects statistically insignificant. Bayesian calibration was effective but not perfect.

Conclusion: Algorithmic classification errors can substantially bias research findings in language acquisition studies, and while Bayesian calibration helps, the issue affects any classifier with non-zero error rates in event detection and classification.

Abstract: With the advent of wearable recorders, scientists are increasingly turning to automated methods of analysis of audio and video data in order to measure children’s experience, behavior, and outcomes, with a sizable literature employing long-form audio-recordings to study language acquisition. While numerous articles report on the accuracy and reliability of the most popular automated classifiers, less has been written on the downstream effects of classification errors on measurements and statistical inferences (e.g., the estimate of correlations and effect sizes in regressions). This paper proposes a Bayesian approach to study the effects of algorithmic errors on key scientific questions, including the effect of siblings on children’s language experience and the association between children’s production and their input. In both the most commonly used \gls{lena}, and an open-source alternative (the Voice Type Classifier from the ACLEW system), we find that classification errors can significantly distort estimates. For instance, automated annotations underestimated the negative effect of siblings on adult input by 20–80%, potentially placing it below statistical significance thresholds. We further show that a Bayesian calibration approach for recovering unbiased estimates of effect sizes can be effective and insightful, but does not provide a fool-proof solution. Both the issue reported and our solution may apply to any classifier involving event detection and classification with non-zero error rates.

Method: Developed Conformalized Exceptional Model Mining framework with mSMoPE model class for uncertainty quantification using conformal prediction guarantees, and introduced RAUL quality measure to identify exceptional performance subgroups in multi-class classification and regression.

Result: Experimental results across diverse datasets demonstrate the framework’s effectiveness in uncovering interpretable subgroups that provide critical insights into model behavior.

Conclusion: This work enhances model interpretability and reliability, advancing explainable AI and uncertainty quantification by providing a rigorous framework for identifying exceptional performance patterns in cohesive data subgroups.

Abstract: Understanding the nuanced performance of machine learning models is essential for responsible deployment, especially in high-stakes domains like healthcare and finance. This paper introduces a novel framework, Conformalized Exceptional Model Mining, which combines the rigor of Conformal Prediction with the explanatory power of Exceptional Model Mining (EMM). The proposed framework identifies cohesive subgroups within data where model performance deviates exceptionally, highlighting regions of both high confidence and high uncertainty. We develop a new model class, mSMoPE (multiplex Soft Model Performance Evaluation), which quantifies uncertainty through conformal prediction’s rigorous coverage guarantees. By defining a new quality measure, Relative Average Uncertainty Loss (RAUL), our framework isolates subgroups with exceptional performance patterns in multi-class classification and regression tasks. Experimental results across diverse datasets demonstrate the framework’s effectiveness in uncovering interpretable subgroups that provide critical insights into model behavior. This work lays the groundwork for enhancing model interpretability and reliability, advancing the state-of-the-art in explainable AI and uncertainty quantification.

Method: 28B parameter multimodal Mixture-of-Experts model pre-trained on 5T tokens (2.5T scientific), using offline+online RL training with Mixture-of-Rewards (MoR) on 1000+ tasks simultaneously in InternBootCamp.

Result: Top-tier performance in online RL training, competitive on general reasoning tasks, significantly outperforms open-source models in scientific domains, and surpasses closed-source SOTA in professional tasks like molecular synthesis planning and crystal stability prediction.

Conclusion: Intern-S1 successfully addresses the scientific domain gap through integrated innovations in algorithms, data, and training systems, demonstrating that specialized generalist models can achieve breakthrough performance in challenging scientific professional fields.

Abstract: In recent years, a plethora of open-source foundation models have emerged, achieving remarkable progress in some widely attended fields, with performance being quite close to that of closed-source models. However, in high-value but more challenging scientific professional fields, either the fields still rely on expert models, or the progress of general foundation models lags significantly compared to those in popular areas, far from sufficient for transforming scientific research and leaving substantial gap between open-source models and closed-source models in these scientific domains. To mitigate this gap and explore a step further toward Artificial General Intelligence (AGI), we introduce Intern-S1, a specialized generalist equipped with general understanding and reasoning capabilities with expertise to analyze multiple science modal data. Intern-S1 is a multimodal Mixture-of-Experts (MoE) model with 28 billion activated parameters and 241 billion total parameters, continually pre-trained on 5T tokens, including over 2.5T tokens from scientific domains. In the post-training stage, Intern-S1 undergoes offline and then online reinforcement learning (RL) in InternBootCamp, where we propose Mixture-of-Rewards (MoR) to synergize the RL training on more than 1000 tasks simultaneously. Through integrated innovations in algorithms, data, and training systems, Intern-S1 achieved top-tier performance in online RL training.On comprehensive evaluation benchmarks, Intern-S1 demonstrates competitive performance on general reasoning tasks among open-source models and significantly outperforms open-source models in scientific domains, surpassing closed-source state-of-the-art models in professional tasks, such as molecular synthesis planning, reaction condition prediction, predicting thermodynamic stabilities for crystals. Our models are available at https://huggingface.co/internlm/Intern-S1.

Method: Two-stage approach: 1) Uses mini-batch optimal transport with closed-form coupling to align real and simulated observations for same latent parameters, 2) Trains conditional normalizing flow to approximate OT-induced posterior for efficient inference without simulation access at test time.

Result: Matches or surpasses performance of RoPE and other SBI/non-SBI estimators across synthetic and real-world benchmarks including complex medical biomarker estimation.

Conclusion: Proposed framework offers improved scalability and applicability in challenging misspecified environments while maintaining competitive performance.

Abstract: Simulation-based inference (SBI) is a statistical inference approach for estimating latent parameters of a physical system when the likelihood is intractable but simulations are available. In practice, SBI is often hindered by model misspecification–the mismatch between simulated and real-world observations caused by inherent modeling simplifications. RoPE, a recent SBI approach, addresses this challenge through a two-stage domain transfer process that combines semi-supervised calibration with optimal transport (OT)-based distribution alignment. However, RoPE operates in a fully transductive setting, requiring access to a batch of test samples at inference time, which limits scalability and generalization. We propose here a fully inductive and amortized SBI framework that integrates calibration and distributional alignment into a single, end-to-end trainable model. Our method leverages mini-batch OT with a closed-form coupling to align real and simulated observations that correspond to the same latent parameters, using both paired calibration data and unpaired samples. A conditional normalizing flow is then trained to approximate the OT-induced posterior, enabling efficient inference without simulation access at test time. Across a range of synthetic and real-world benchmarks–including complex medical biomarker estimation–our approach matches or surpasses the performance of RoPE, as well as other standard SBI and non-SBI estimators, while offering improved scalability and applicability in challenging, misspecified environments.

Method: Uses a global prior distribution that is continuously updated to enable learning new topic models at subsequent time steps while preserving previously learned topics.

Result: CoNTM consistently outperformed dynamic topic models in topic quality and predictive perplexity, learning more diverse topics and better capturing temporal changes.

Conclusion: CoNTM successfully fills the gap by providing continual topic learning with long-term memory, demonstrating superior performance over existing methods.

Abstract: In continual learning, our aim is to learn a new task without forgetting what was learned previously. In topic models, this translates to learning new topic models without forgetting previously learned topics. Previous work either considered Dynamic Topic Models (DTMs), which learn the evolution of topics based on the entire training corpus at once, or Online Topic Models, which are updated continuously based on new data but do not have long-term memory. To fill this gap, we propose the Continual Neural Topic Model (CoNTM), which continuously learns topic models at subsequent time steps without forgetting what was previously learned. This is achieved using a global prior distribution that is continuously updated. In our experiments, CoNTM consistently outperformed the dynamic topic model in terms of topic quality and predictive perplexity while being able to capture topic changes online. The analysis reveals that CoNTM can learn more diverse topics and better capture temporal changes than existing methods.

Method: Proposes Graph Autoencoder with Spectral Encoder and Spectral Decoder (GRASPED) using Graph Wavelet Convolution-based encoder and Wiener Graph Deconvolution-based decoder with structural and attribute decoders for bandpass filtering.

Result: Extensive experiments on real-world graph anomaly detection datasets show GRASPED outperforms current state-of-the-art models.

Conclusion: GRASPED effectively captures anomaly information through unsupervised learning with multi-scale spectral analysis, addressing limitations of both supervised and existing unsupervised approaches.

Abstract: Graph machine learning has been widely explored in various domains, such as community detection, transaction analysis, and recommendation systems. In these applications, anomaly detection plays an important role. Recently, studies have shown that anomalies on graphs induce spectral shifts. Some supervised methods have improved the utilization of such spectral domain information. However, they remain limited by the scarcity of labeled data due to the nature of anomalies. On the other hand, existing unsupervised learning approaches predominantly rely on spatial information or only employ low-pass filters, thereby losing the capacity for multi-band analysis. In this paper, we propose Graph Autoencoder with Spectral Encoder and Spectral Decoder (GRASPED) for node anomaly detection. Our unsupervised learning model features an encoder based on Graph Wavelet Convolution, along with structural and attribute decoders. The Graph Wavelet Convolution-based encoder, combined with a Wiener Graph Deconvolution-based decoder, exhibits bandpass filter characteristics that capture global and local graph information at multiple scales. This design allows for a learning-based reconstruction of node attributes, effectively capturing anomaly information. Extensive experiments on several real-world graph anomaly detection datasets demonstrate that GRASPED outperforms current state-of-the-art models.

Method: Two-part approach: (1) provably fair initialization to start in a fair state, and (2) fairness-preserving training algorithm using randomized response mechanisms to protect sensitive attributes while maintaining fairness guarantees.

Result: Experimental evaluations show the approach produces empirically fair and accurate models, and is more efficient than certified training methods that require neural network verification during training.

Conclusion: The proposed framework successfully provides formal guarantees of individual fairness throughout training while maintaining model accuracy and efficiency, addressing limitations of existing fairness methods.

Abstract: Fairness in machine learning is more important than ever as ethical concerns continue to grow. Individual fairness demands that individuals differing only in sensitive attributes receive the same outcomes. However, commonly used machine learning algorithms often fail to achieve such fairness. To improve individual fairness, various training methods have been developed, such as incorporating fairness constraints as optimisation objectives. While these methods have demonstrated empirical effectiveness, they lack formal guarantees of fairness. Existing approaches that aim to provide fairness guarantees primarily rely on verification techniques, which can sometimes fail to produce definitive results. Moreover, verification alone does not actively enhance individual fairness during training. To address this limitation, we propose a novel framework that formally guarantees individual fairness throughout training. Our approach consists of two parts, i.e., (1) provably fair initialisation that ensures the model starts in a fair state, and (2) a fairness-preserving training algorithm that maintains fairness as the model learns. A key element of our method is the use of randomised response mechanisms, which protect sensitive attributes while maintaining fairness guarantees. We formally prove that this mechanism sustains individual fairness throughout the training process. Experimental evaluations confirm that our approach is effective, i.e., producing models that are empirically fair and accurate. Furthermore, our approach is much more efficient than the alternative approach based on certified training (which requires neural network verification during training).

Method: Transformer-based latent-variable framework with Ornstein-Uhlenbeck priors over compartment model parameters, pre-trained on synthetic pharmacokinetic trajectories and using amortized in-context Bayesian inference.

Result: AICMET achieves state-of-the-art predictive accuracy, faithfully quantifies inter-patient variability, and outperforms both nonlinear mixed-effects baselines and neural ODE variants.

Conclusion: Transformer-based population-aware neural architectures offer a viable alternative to traditional pharmacokinetic modeling, enabling rapid adaptation and personalized dosing regimens.

Abstract: Accurate dose-response forecasting under sparse sampling is central to precision pharmacotherapy. We present the Amortized In-Context Mixed-Effect Transformer (AICMET) model, a transformer-based latent-variable framework that unifies mechanistic compartmental priors with amortized in-context Bayesian inference. AICMET is pre-trained on hundreds of thousands of synthetic pharmacokinetic trajectories with Ornstein-Uhlenbeck priors over the parameters of compartment models, endowing the model with strong inductive biases and enabling zero-shot adaptation to new compounds. At inference time, the decoder conditions on the collective context of previously profiled trial participants, generating calibrated posterior predictions for newly enrolled patients after a few early drug concentration measurements. This capability collapses traditional model-development cycles from weeks to hours while preserving some degree of expert modelling. Experiments across public datasets show that AICMET attains state-of-the-art predictive accuracy and faithfully quantifies inter-patient variability – outperforming both nonlinear mixed-effects baselines and recent neural ODE variants. Our results highlight the feasibility of transformer-based, population-aware neural architectures as offering a new alternative for bespoke pharmacokinetic modeling pipelines, charting a path toward truly population-aware personalized dosing regimens.

Method: Multi-task learning with low-rank tensor decomposition that decomposes task model parameters into low-rank structures to capture common patterns and subpopulation-specific variations.

Result: Superior performance compared to benchmarks in simulation and case studies, especially with high subpopulation variability, improving both prediction accuracy and interpretability.

Conclusion: The proposed framework successfully enhances personalized modeling by efficiently sharing knowledge between tasks while preserving subpopulation uniqueness, offering improved accuracy and interpretability.

Abstract: Effective modeling of heterogeneous subpopulations presents a significant challenge due to variations in individual characteristics and behaviors. This paper proposes a novel approach to address this issue through multi-task learning (MTL) and low-rank tensor decomposition techniques. Our MTL approach aims to enhance personalized modeling by leveraging shared structures among similar tasks while accounting for distinct subpopulation-specific variations. We introduce a framework where low-rank decomposition decomposes the collection of task model parameters into a low-rank structure that captures commonalities and variations across tasks and subpopulations. This approach allows for efficient learning of personalized models by sharing knowledge between similar tasks while preserving the unique characteristics of each subpopulation. Experimental results in simulation and case study datasets demonstrate the superior performance of the proposed method compared to several benchmarks, particularly in scenarios with high variability among subpopulations. The proposed framework not only improves prediction accuracy but also enhances interpretability by revealing underlying patterns that contribute to the personalization of models.

Method: Presenting a multi-agent environment where self-interested agents pursue complex independent goals, allowing study of social learning with experts, emergent collaborative tool use, and implicit cooperation/competition.

Result: The environment enables investigation of how social learning impacts agent performance in the presence of experts and whether agents benefit from cooperation or competition.

Conclusion: This environment provides a foundation for developing socially intelligent AI agents that can learn adaptive skills from human experts through social interactions in complex multi-agent settings.

Abstract: Many challenges remain before AI agents can be deployed in real-world environments. However, one virtue of such environments is that they are inherently multi-agent and contain human experts. Using advanced social intelligence in such an environment can help an AI agent learn adaptive skills and behaviors that a known expert exhibits. While social intelligence could accelerate training, it is currently difficult to study due to the lack of open-ended multi-agent environments. In this work, we present an environment in which multiple self-interested agents can pursue complex and independent goals, reflective of real world challenges. This environment will enable research into the development of socially intelligent AI agents in open-ended multi-agent settings, where agents may be implicitly incentivized to cooperate to defeat common enemies, build and share tools, and achieve long horizon goals. In this work, we investigate the impact on agent performance due to social learning in the presence of experts and implicit cooperation such as emergent collaborative tool use, and whether agents can benefit from either cooperation or competition in this environment.

Method: Generalized augmented Lagrangian method with multiple penalty parameters, reformulated constraint enforcement as expectations, incorporated Fourier feature mappings, introduced time-windowing strategy, and proposed conditionally adaptive penalty update (CAPU) strategy.

Result: PECANN-CAPU achieves competitive accuracy across various problems including transonic rarefaction, vortex advection, high-wavenumber Helmholtz/Poisson equations, and inverse heat source identification, outperforming established methods and recent Kolmogorov-Arnold network approaches.

Conclusion: The collective advances significantly improve PECANN’s robustness, efficiency, and applicability to challenging scientific computing problems with oscillatory, multi-scale features and long-time evolution requirements.

Abstract: We present several advances to the physics and equality constrained artificial neural networks (PECANN) framework that substantially improve its capability to learn solutions of canonical partial differential equations (PDEs). First, we generalize the augmented Lagrangian method (ALM) to support multiple independent penalty parameters, enabling simultaneous enforcement of heterogeneous constraints. Second, we reformulate pointwise constraint enforcement and Lagrange multipliers as expectations over constraint terms, reducing memory overhead and permitting efficient mini-batch training. Third, to address PDEs with oscillatory, multi-scale features, we incorporate Fourier feature mappings and show that a single mapping suffices where multiple mappings or more costly architectures were required in related methods. Fourth, we introduce a time-windowing strategy for long-time evolution in which the terminal state of each window is enforced as an initial-condition constraint for the next, ensuring continuity without discrete time models. Crucially, we propose a conditionally adaptive penalty update (CAPU) strategy for ALM, which preserves the principle that larger constraint violations incur stronger penalties. CAPU accelerates the growth of Lagrange multipliers for selectively challenging constraints, enhancing constraint enforcement during training. We demonstrate the effectiveness of PECANN-CAPU on problems including the transonic rarefaction problem, reversible advection of a passive by a vortex, high-wavenumber Helmholtz and Poisson equations, and inverse identification of spatially varying heat sources. Comparisons with established methods and recent Kolmogorov-Arnold network approaches show that PECANN-CAPU achieves competitive accuracy across all cases. Collectively, these advances improve PECANN’s robustness, efficiency, and applicability to demanding problems in scientific computing.

Method: The paper analyzes techniques including maximum likelihood estimation, Bayesian inference, and attention mechanisms through illustrative scenarios such as system identification, image classification, and language generation, showing how complex models build upon these foundations.

Result: The work demonstrates that maximum likelihood, MAP estimation, Bayesian classification, and deep learning all represent different facets of the shared goal of inferring hidden causes from noisy and/or biased observations.

Conclusion: This survey serves as both a theoretical synthesis and practical guide, showing that many AI methods are rooted in shared probabilistic principles for addressing practical challenges like overfitting, data sparsity, and interpretability.

Abstract: Extracting meaning from uncertain, noisy data is a fundamental problem across time series analysis, pattern recognition, and language modeling. This survey presents a unified mathematical framework that connects classical estimation theory, statistical inference, and modern machine learning, including deep learning and large language models. By analyzing how techniques such as maximum likelihood estimation, Bayesian inference, and attention mechanisms address uncertainty, the paper illustrates that many AI methods are rooted in shared probabilistic principles. Through illustrative scenarios including system identification, image classification, and language generation, we show how increasingly complex models build upon these foundations to tackle practical challenges like overfitting, data sparsity, and interpretability. In other words, the work demonstrates that maximum likelihood, MAP estimation, Bayesian classification, and deep learning all represent different facets of a shared goal: inferring hidden causes from noisy and/or biased observations. It serves as both a theoretical synthesis and a practical guide for students and researchers navigating the evolving landscape of machine learning.

Method: A novel hybrid sampling approach combining classical MCMC and quantum annealing contributions from D-Wave hardware, applied to Restricted Boltzmann Machines. Also investigated using QA-generated patterns for catastrophic forgetting mitigation through generative replay.

Result: No improvements in RBM training were achieved. Differences between QA and MCMC sampling were mainly in medium-to-low probability regions, which are less important for sample quality. However, demonstrated feasibility of using QA-generated patterns for catastrophic forgetting mitigation with comparable efficiency to classical methods.

Conclusion: The modest sampling differences between QA and MCMC are insufficient to benefit RBM training, but QA’s ability to generate diverse samples from lower-probability regions shows promise for applications like catastrophic forgetting mitigation, with potential for speed advantages and further improvements.

Abstract: Modest statistical differences between the sampling performances of the D-Wave quantum annealer (QA) and the classical Markov Chain Monte Carlo (MCMC), when applied to Restricted Boltzmann Machines (RBMs), are explored to explain, and possibly address, the absence of significant and consistent improvements in RBM trainability when the D-Wave sampling was used in previous investigations. A novel hybrid sampling approach, combining the classical and the QA contributions, is investigated as a promising way to benefit from the modest differences between the two sampling methods. No improvements in the RBM training are achieved in this work, thereby suggesting that the differences between the QA-based and MCMC sampling, mainly found in the medium-to-low probability regions of the distribution, which are less important for the quality of the sample, are insufficient to benefit the training. Difficulties in achieving sufficiently high quality of embedding RBMs into the lattice of the newer generation of D-Wave hardware could be further complicating the task. On the other hand, the ability to generate samples of sufficient variety from lower-probability parts of the distribution has a potential to benefit other machine learning applications, such as the mitigation of catastrophic forgetting (CF) during incremental learning. The feasibility of using QA-generated patterns of desirable classes for CF mitigation by the generative replay is demonstrated in this work for the first time. While the efficiency of the CF mitigation using the D-Wave QA was comparable to that of the classical mitigation, both the speed of generating a large number of distinct desirable patterns and the potential for further improvement make this approach promising for a variety of challenging machine learning applications.

Method: Uses Top-k sparsification and 2-bit quantization with local momentum approximation through error feedback and sparse aggregation techniques.

Result: Achieves compression ratios of up to 1-3% sparsity with 2-bit quantization while outperforming full-precision DiLoCo in various communication-constrained LLM training settings.

Conclusion: SparseLoCo effectively addresses communication bottlenecks in distributed LLM training through extreme compression while maintaining or improving model performance.

Abstract: Communication-efficient distributed training algorithms have received considerable interest recently due to their benefits for training Large Language Models (LLMs) in bandwidth-constrained settings, such as across data centers and over the internet. Despite reducing communication frequency, these methods still typically require communicating a full copy of the model’s gradients-resulting in a communication bottleneck even for cross-datacenter links. Furthermore, they can slightly degrade performance compared to a naive AdamW DDP baseline. While quantization and error feedback are often applied to reduce the pseudo-gradient’s size, in the context of LLM pre-training, existing approaches have been unable to additionally leverage sparsification and have obtained limited quantization. In this work, we introduce SparseLoCo, a communication-efficient training algorithm for LLMs that effectively leverages Top-k sparsification and quantization to reach extreme compression ratios of up to 1-3% sparsity and 2-bit quantization while outperforming full-precision DiLoCo. Our key observations are that outer momentum can be locally approximated by an error feedback combined with aggressive sparsity and that sparse aggregation can actually improve model performance. We empirically demonstrate in a range of communication-constrained LLM training settings that SparseLoCo provides significant benefits in both performance and communication cost.

Method: Trained a Variational Diffusion Model on the representation space of a pre-trained ResNet-18 instead of images, then compared likelihood-based OOD detection performance against state-of-the-art methods from OpenOOD suite.

Result: The paper tests whether representation space performs better than pixel space for likelihood-based OOD detection, comparing against established benchmarks.

Conclusion: The study investigates if the failure of likelihood-based OOD detection is specific to pixel space or extends to representation space, potentially revealing insights about density estimation limitations in different spaces.

Abstract: Despite rapid advances in AI, safety remains the main bottleneck to deploying machine-learning systems. A critical safety component is out-of-distribution detection: given an input, decide whether it comes from the same distribution as the training data. In generative models, the most natural OOD score is the data likelihood. Actually, under the assumption of uniformly distributed OOD data, the likelihood is even the optimal OOD detector, as we show in this work. However, earlier work reported that likelihood often fails in practice, raising doubts about its usefulness. We explore whether, in practice, the representation space also suffers from the inability to learn good density estimation for OOD detection, or if it is merely a problem of the pixel space typically used in generative models. To test this, we trained a Variational Diffusion Model not on images, but on the representation space of a pre-trained ResNet-18 to assess the performance of our likelihood-based detector in comparison to state-of-the-art methods from the OpenOOD suite.

Method: Developed synthetic data generation pipeline, trained transformers via supervised learning, and proposed Beam Grouped Relative Policy Optimization (BGRPO) - a rank-aware reinforcement learning method for hard algebraic problems.

Result: BGRPO finetuning improved accuracy while reducing beam width by up to half, resulting in approximately 75% lower inference compute. Model demonstrated competitive performance in polynomial simplification, outperforming Mathematica in various cases.

Conclusion: Transformers can be effectively trained for complex algebraic tasks like polynomial decomposition through careful data generation and specialized reinforcement learning methods, achieving significant computational efficiency gains while maintaining high accuracy.

Abstract: Recent efforts have extended the capabilities of transformers in logical reasoning and symbolic computations. In this work, we investigate their capacity for non-linear latent pattern discovery in the context of functional decomposition, focusing on the challenging algebraic task of multivariate polynomial decomposition. This problem, with widespread applications in science and engineering, is proved to be NP-hard, and demands both precision and insight. Our contributions are threefold: First, we develop a synthetic data generation pipeline providing fine-grained control over problem complexity. Second, we train transformer models via supervised learning and evaluate them across four key dimensions involving scaling behavior and generalizability. Third, we propose Beam Grouped Relative Policy Optimization (BGRPO), a rank-aware reinforcement learning method suitable for hard algebraic problems. Finetuning with BGRPO improves accuracy while reducing beam width by up to half, resulting in approximately 75% lower inference compute. Additionally, our model demonstrates competitive performance in polynomial simplification, outperforming Mathematica in various cases.

Method: Combines generative learning and contrastive regularization via Contrastive Regularized MAE loss, using Signal Transformer (SiT) architecture to capture both local waveform details and global temporal dependencies.

Result: Outperforms supervised baselines and existing self-supervised models in both linear probing and fine-tuning evaluations, maintaining superior performance across diverse clinical domains including emergency care.

Conclusion: CREMA serves as a scalable and reliable foundation model for ECG diagnostics, supporting downstream applications across heterogeneous and high-risk clinical settings.

Abstract: Electrocardiogram (ECG) diagnosis remains challenging due to limited labeled data and the need to capture subtle yet clinically meaningful variations in rhythm and morphology. We present CREMA (Contrastive Regularized Masked Autoencoder), a foundation model for 12-lead ECGs designed to learn generalizable representations through self-supervised pretraining. CREMA combines generative learning and contrastive regularization via a Contrastive Regularized MAE loss, and employs a Signal Transformer (SiT) architecture to capture both local waveform details and global temporal dependencies. We evaluate CREMA on benchmark datasets and real-world clinical environments, including deployment scenarios with significant distribution shifts. CREMA outperforms supervised baselines and existing self-supervised models in both linear probing and fine-tuning evaluations. Notably, it maintains superior performance across diverse clinical domains, such as emergency care, highlighting its robustness under real-world conditions. These results demonstrate that CREMA serves as a scalable and reliable foundation model for ECG diagnostics, supporting downstream applications across heterogeneous and high-risk clinical settings.

Method: Develops a closed-form function to determine target dimensionality that preserves KNN similarity. Defines measure functions for KNN similarity, extends to global metric spaces, and incorporates into various dimension-reduction methods and distance metrics.

Result: A theoretical framework for dimensionality reduction that maintains the same set of top-k nearest neighbors in lower-dimensional space, enabling efficient similarity searches.

Conclusion: OPDR provides a mathematically grounded approach to reduce embedding dimensions while preserving nearest neighbor relationships, making semantic search practical for time-sensitive scientific applications.

Abstract: One of the most common operations in multimodal scientific data management is searching for the $k$ most similar items (or, $k$-nearest neighbors, KNN) from the database after being provided a new item. Although recent advances of multimodal machine learning models offer a \textit{semantic} index, the so-called \textit{embedding vectors} mapped from the original multimodal data, the dimension of the resulting embedding vectors are usually on the order of hundreds or a thousand, which are impractically high for time-sensitive scientific applications. This work proposes to reduce the dimensionality of the output embedding vectors such that the set of top-$k$ nearest neighbors do not change in the lower-dimensional space, namely Order-Preserving Dimension Reduction (OPDR). In order to develop such an OPDR method, our central hypothesis is that by analyzing the intrinsic relationship among key parameters during the dimension-reduction map, a quantitative function may be constructed to reveal the correlation between the target (lower) dimensionality and other variables. To demonstrate the hypothesis, this paper first defines a formal measure function to quantify the KNN similarity for a specific vector, then extends the measure into an aggregate accuracy of the global metric spaces, and finally derives a closed-form function between the target (lower) dimensionality and other variables. We incorporate the closed-function into popular dimension-reduction methods, various distance metrics, and embedding models.

Method: Introduces a simple nonconvex framework that incrementally learns the top-k nonzero elements of the mean while keeping zero elements small, without requiring knowledge of k.

Result: The estimator works without sparsity knowledge, runs in near-linear time and memory with respect to ambient dimension, and achieves information-theoretic optimality when signal-to-noise ratio is large.

Conclusion: The proposed method overcomes key limitations of existing approaches and provides a practical solution for robust sparse mean estimation from heavy-tailed corrupted samples.

Abstract: In this paper, we study the problem of robust sparse mean estimation, where the goal is to estimate a $k$-sparse mean from a collection of partially corrupted samples drawn from a heavy-tailed distribution. Existing estimators face two critical challenges in this setting. First, the existing estimators rely on the prior knowledge of the sparsity level $k$. Second, the existing estimators fall short of practical use as they scale poorly with the ambient dimension. This paper presents a simple mean estimator that overcomes both challenges under moderate conditions: it works without the knowledge of $k$ and runs in near-linear time and memory (both with respect to the ambient dimension). Moreover, provided that the signal-to-noise ratio is large, we can further improve our result to match the information-theoretic lower bound. At the core of our method lies an incremental learning phenomenon: we introduce a simple nonconvex framework that can incrementally learn the top-$k$ nonzero elements of the mean while keeping the zero elements arbitrarily small. Finally, we conduct a series of simulations to corroborate our theoretical findings.

Method: Leverages both class-incremental and domain-incremental learning scenarios in environments with scarce labeled data but plentiful non-iid unlabeled data. Uses prototype storage for efficiency and evaluates on multimodal tactile and visual datasets.

Result: Algorithm demonstrates effectiveness on challenging custom multimodal dataset (tactile data from soft pneumatic gripper + visual data from video sequences) and Core50 dataset. Real-time object classification experiments with ROS framework show robustness.

Conclusion: The proposed continual learning algorithm successfully enables incremental learning across different data modalities while maintaining efficiency through prototype storage, demonstrating practical applicability in real-world robotic scenarios.

Abstract: Continual learning (CL) refers to the ability of an algorithm to continuously and incrementally acquire new knowledge from its environment while retaining previously learned information. A model trained on one data modality often fails when tested with a different modality. A straightforward approach might be to fuse the two modalities by concatenating their features and training the model on the fused data. However, this requires retraining the model from scratch each time it encounters a new domain. In this paper, we introduce a continual learning algorithm capable of incrementally learning different data modalities by leveraging both class-incremental and domain-incremental learning scenarios in an artificial environment where labeled data is scarce, yet non-iid (independent and identical distribution) unlabeled data from the environment is plentiful. The proposed algorithm is efficient and only requires storing prototypes for each class. We evaluate the algorithm’s effectiveness on a challenging custom multimodal dataset comprising of tactile data from a soft pneumatic gripper, and visual data from non-stationary images of objects extracted from video sequences. Additionally, we conduct an ablation study on the custom dataset and the Core50 dataset to highlight the contributions of different components of the algorithm. To further demonstrate the robustness of the algorithm, we perform a real-time experiment for object classification using the soft gripper and an external independent camera setup, all synchronized with the Robot Operating System (ROS) framework.

Method: Created a mathematical framework analyzing Transformers through the lens of interacting particle systems theory

Result: Revealed that clusters emerge in Transformers over long time periods through this particle system interpretation

Conclusion: Provides new theoretical perspectives on Transformer mechanics that benefit both mathematicians and computer scientists

Abstract: Transformers play a central role in the inner workings of large language models. We develop a mathematical framework for analyzing Transformers based on their interpretation as interacting particle systems, which reveals that clusters emerge in long time. Our study explores the underlying theory and offers new perspectives for mathematicians as well as computer scientists.

Method: Proposed upper-confidence bound algorithms for both high probability and expectation constraint settings, with extensions to multiple constraints and non-linear functions. Used eluder dimension to characterize function class complexity.

Result: Developed algorithms with proven T-round regret bounds for both linear and non-linear cases, provided lower bounds for the constrained problem, and validated results through simulations.

Conclusion: The paper successfully addresses constrained contextual bandits with both probabilistic and expected constraints, providing theoretically sound algorithms with regret guarantees and computational efficiency for various settings including multi-armed bandits.

Abstract: We study contextual bandits in the presence of a stage-wise constraint when the constraint must be satisfied both with high probability and in expectation. We start with the linear case where both the reward function and the stage-wise constraint (cost function) are linear. In each of the high probability and in expectation settings, we propose an upper-confidence bound algorithm for the problem and prove a $T$-round regret bound for it. We also prove a lower-bound for this constrained problem, show how our algorithms and analyses can be extended to multiple constraints, and provide simulations to validate our theoretical results. In the high probability setting, we describe the minimum requirements for the action set for our algorithm to be tractable. In the setting that the constraint is in expectation, we specialize our results to multi-armed bandits and propose a computationally efficient algorithm for this setting with regret analysis. Finally, we extend our results to the case where the reward and cost functions are both non-linear. We propose an algorithm for this case and prove a regret bound for it that characterize the function class complexity by the eluder dimension.

Method: Reformulates convex training program for ReLU-based shallow neural networks using order-1 Wasserstein distributionally robust optimization framework, with mixed-integer convex post-training verification.

Result: Provides out-of-sample performance guarantees, enables enforcement of hard convex physical constraints, and demonstrates convincing performance in synthetic experiments and real-world power system applications.

Conclusion: WaDiRo-SCNN offers a conservative, scalable, and verifiable approach to neural network training that enhances safety and reliability for critical industrial applications.

Abstract: In this work, we propose Wasserstein distributionally robust shallow convex neural networks (WaDiRo-SCNNs) to provide reliable nonlinear predictions when subject to adverse and corrupted datasets. Our approach is based on the reformulation of a new convex training program for ReLU-based shallow neural networks, which allows us to cast the problem into the order-1 Wasserstein distributionally robust optimization framework. Our training procedure is conservative, has low stochasticity, is solvable with open-source solvers, and is scalable to large industrial deployments. We provide out-of-sample performance guarantees, show that hard convex physical constraints can be enforced in the training program, and propose a mixed-integer convex post-training verification program to evaluate model stability. WaDiRo-SCNN aims to make neural networks safer for critical applications, such as in the energy sector. Finally, we numerically demonstrate our model’s performance through both a synthetic experiment and a real-world power system application, viz., the prediction of hourly energy consumption in non-residential buildings within the context of virtual power plants, and evaluate its stability across standard regression benchmark datasets. The experimental results are convincing and showcase the strengths of the proposed model.

Method: Modified VarLiNGAM by omitting iterative refinement, allowing one-time precomputation of statistical values. This reduces time complexity from O(m³n) to O(m²n + m³) while keeping space complexity at O(m²).

Result: Achieved 7-13x speedup over standard implementation and 4.5x speedup over GPU-accelerated version on financial data with 400 variables. Demonstrated robustness across medical imaging, web server monitoring, and finance domains.

Conclusion: The heuristic provides a validated balance between computational efficiency and discovery quality, enabling large-scale causal analysis on personal computers while retaining VarLiNGAM’s essential structure and empirical reliability.

Abstract: Causal discovery in time-series data presents a significant computational challenge. Standard algorithms are often prohibitively expensive for datasets with many variables or samples. This study introduces and validates a heuristic approximation of the VarLiNGAM algorithm to address this scalability problem. The standard VarLiNGAM method relies on an iterative search, recalculating statistical dependencies after each step. Our heuristic modifies this procedure by omitting the iterative refinement. This change permits a one-time precomputation of all necessary statistical values. The algorithmic modification reduces the time complexity from $O(m^3n)$ to $O(m^2n + m^3)$ while keeping the space complexity at $O(m^2)$, where $m$ is the number of variables and $n$ is the number of samples. While an approximation, our approach retains VarLiNGAM’s essential structure and empirical reliability. On large-scale financial data with up to 400 variables, our algorithm achieves a 7–13x speedup over the standard implementation and a 4.5x speedup over a GPU-accelerated version. Evaluations across medical imaging, web server monitoring, and finance demonstrate the heuristic’s robustness and practical scalability. This work offers a validated balance between computational efficiency and discovery quality, making large-scale causal analysis feasible on personal computers.

Method: MATATA uses a novel weakly supervised two-stage training approach with final outcome supervision instead of intermediate step supervision. It employs an adaptive planner and shared tools across datasets, training 3.8B/8B SLMs without annotations.

Result: Achieves state-of-the-art on FinQA and TAT-QA among open-source SLM methods. Closely matches GPT-4-based frameworks on TabMWP despite using small language models.

Conclusion: MATATA enables training end-to-end multi-step reasoning agents without intermediate supervision, supporting cost-effective development of powerful agentic systems for document understanding.

Abstract: Business documents often contain substantial tabular and textual information with numerical values, requiring mathematical reasoning for effective document understanding. While Small Language Models (SLMs) still struggle at this task, tool-augmented multi-step agents perform better, at the cost of relying on closed-source or larger models, external data, or extensive prompt-engineering. This work introduces MATATA, a novel weakly supervised end-to-end approach to train multi-step reasoning language agents for document tabular applications. MATATA presents an annotation-free paradigm for each agent to enhance 3.8B/8B SLMs. During its two-stage training, MATATA uses the final outcome of the multi-step reasoning chain as weak supervision. This approach avoids having to individually supervise each intermediate agent in the reasoning chain. By employing an adaptive planner and shared tools across different datasets, MATATA shows robust performance. Experiments demonstrate that MATATA achieves state-of-the-art on FinQA, and on TAT-QA among reasoning methods based on open-source SLMs. Although being SLM-based, MATATA closely matches GPT-4-based frameworks on TabMWP. This novel weakly supervised approach enables training an end-to-end multi-step reasoning agent without intermediate supervision, supporting future developments of cost-effective powerful agentic systems.

Method: Introduced a theoretical framework using rate-distortion theory and Kolmogorov complexity to measure network complexity as principled lossy compression. Developed spectral entropy regularization to penalize intrinsic dimension and encourage low-complexity representations.

Result: Regularized networks show a sharp phase transition with complexity rising during memorization then falling as simpler generalizing patterns are discovered. Achieved 30-40x better compression ratios than naive approaches. Unregularized networks remain trapped in high-complexity memorization.

Conclusion: The study establishes an explicit connection between complexity measures and generalization bounds, providing theoretical foundation for lossy compression-generalization link. Spectral entropy regularization effectively guides networks toward low-complexity, generalizable representations.

Abstract: We demonstrate the existence of a complexity phase transition in neural networks by studying the grokking phenomenon, where networks suddenly transition from memorization to generalization long after overfitting their training data. To characterize this phase transition, we introduce a theoretical framework for measuring complexity based on rate-distortion theory and Kolmogorov complexity, which can be understood as principled lossy compression for networks. We find that properly regularized networks exhibit a sharp phase transition: complexity rises during memorization, then falls as the network discovers a simpler underlying pattern that generalizes. In contrast, unregularized networks remain trapped in a high-complexity memorization phase. We establish an explicit connection between our complexity measure and generalization bounds, providing a theoretical foundation for the link between lossy compression and generalization. Our framework achieves compression ratios 30-40x better than na"ive approaches, enabling precise tracking of complexity dynamics. Finally, we introduce a regularization method based on spectral entropy that encourages networks toward low-complexity representations by penalizing their intrinsic dimension.

Method: Proposed InfAlign-CTRL algorithm with reward calibration and KL-regularized reward maximization using transformed rewards. Specific transformations developed for best-of-N sampling and jailbreaking scenarios.

Result: Achieves 3-8% improvement on inference-time win rates compared to standard RLHF. The reward calibration method also serves as a strong baseline for optimizing standard win rates.

Conclusion: Inference-aware alignment (InfAlign) provides a principled framework to bridge the train/test gap in language model alignment, delivering significant performance gains for practical inference-time decoding methods.

Abstract: Language model alignment is a critical step in training modern generative language models. Alignment targets to improve win rate of a sample from the aligned model against the base model. Today, we are increasingly using inference-time algorithms (e.g., Best-of-N, controlled decoding, tree search) to decode from language models rather than standard sampling. We show that this train/test mismatch makes standard RLHF framework sub-optimal in view of such inference-time methods. To this end, we propose a framework for inference-aware alignment (InfAlign), which aims to optimize inference-time win rate of the aligned policy against the base model. We prove that for any inference-time decoding procedure, the optimal aligned policy is the solution to the standard RLHF problem with a transformation of the reward. This motivates us to provide the calibrate-and-transform RL (InfAlign-CTRL) algorithm to solve this problem, which involves a reward calibration step and a KL-regularized reward maximization step with a transformation of the calibrated reward. For best-of-N sampling and best-of-N jailbreaking, we propose specific transformations offering up to 3-8% improvement on inference-time win rates. Finally, we also show that our proposed reward calibration method is a strong baseline for optimizing standard win rate.

Method: Introduces Weighted Allocation Probability Adjusted Thompson Sampling (WAPTS) that refines sampling strategy with lenient regret principle, allowing near-optimal allocations to accelerate learning while exploring promising content.

Result: WAPTS enables earlier and more reliable identification of promising treatments in learnersourcing scenarios where students rate peer-generated learning materials.

Conclusion: WAPTS provides an effective solution for improving content-related decision-making in data-sparse educational environments, outperforming standard Thompson Sampling approaches.

Abstract: Adaptive experimentation is increasingly used in educational platforms to personalize learning through dynamic content and feedback. However, standard adaptive strategies such as Thompson Sampling often underperform in real-world educational settings where content variations are numerous and student participation is limited, resulting in sparse data. In particular, Thompson Sampling can lead to imbalanced content allocation and delayed convergence on which aspects of content are most effective for student learning. To address these challenges, we introduce Weighted Allocation Probability Adjusted Thompson Sampling (WAPTS), an algorithm that refines the sampling strategy to improve content-related decision-making in data-sparse environments. WAPTS is guided by the principle of lenient regret, allowing near-optimal allocations to accelerate learning while still exploring promising content. We evaluate WAPTS in a learnersourcing scenario where students rate peer-generated learning materials, and demonstrate that it enables earlier and more reliable identification of promising treatments.

Method: Theoretical analysis of RSGD convergence rates with increasing vs constant batch sizes, combined with numerical experiments using PCA and low-rank matrix completion tasks to measure stochastic first-order oracle complexity.

Result: Increasing batch size achieves faster convergence (O(T⁻¹/²)) than constant batch size (O(√(T⁻¹+const))) and reduces computational complexity while combining benefits of both small and large constant batch sizes.

Conclusion: Batch size scheduling is crucial for optimizing RSGD performance, with increasing batch sizes providing superior convergence rates and computational efficiency compared to constant batch strategies.

Abstract: We have theoretically analyzed the use of Riemannian stochastic gradient descent (RSGD) and found that using an increasing batch size leads to faster RSGD convergence rate than using a constant batch size not only with a constant learning rate but also with a decaying learning rate, such as cosine annealing decay and polynomial decay. The convergence rate of RSGD improves from $O(\sqrt{T^{-1}+\text{const.}})$ with a constant batch size to $O(T^{-\frac{1}{2}})$ with an increasing batch size, where $T$ denotes the number of iterations. Using principal component analysis and low-rank matrix completion tasks, we investigated, both theoretically and numerically, how increasing batch size affects computational time as measured by stochastic first-order oracle (SFO) complexity. Increasing batch size reduces the SFO complexity of RSGD. Furthermore, our numerical results demonstrated that increasing batch size offers the advantages of both small and large constant batch sizes.

Method: Proposes a simple reward design focused primarily on route completion, with infractions penalized by episode termination or multiplicative reduction. Uses PPO with large mini-batch sizes enabled by distributed data parallelism.

Result: Achieves 64 DS on CARLA longest6 v2 benchmark (outperforming other RL methods), 91.3/90.6 scores on nuPlan Val14 benchmark, and is an order of magnitude faster than prior work.

Conclusion: Simple intuitive rewards (route completion) enable better scaling and performance than complex reward designs in autonomous driving RL, allowing efficient training with large batch sizes.

Abstract: We investigate reinforcement learning (RL) for privileged planning in autonomous driving. State-of-the-art approaches for this task are rule-based, but these methods do not scale to the long tail. RL, on the other hand, is scalable and does not suffer from compounding errors like imitation learning. Contemporary RL approaches for driving use complex shaped rewards that sum multiple individual rewards, \eg~progress, position, or orientation rewards. We show that PPO fails to optimize a popular version of these rewards when the mini-batch size is increased, which limits the scalability of these approaches. Instead, we propose a new reward design based primarily on optimizing a single intuitive reward term: route completion. Infractions are penalized by terminating the episode or multiplicatively reducing route completion. We find that PPO scales well with higher mini-batch sizes when trained with our simple reward, even improving performance. Training with large mini-batch sizes enables efficient scaling via distributed data parallelism. We scale PPO to 300M samples in CARLA and 500M samples in nuPlan with a single 8-GPU node. The resulting model achieves 64 DS on the CARLA longest6 v2 benchmark, outperforming other RL methods with more complex rewards by a large margin. Requiring only minimal adaptations from its use in CARLA, the same method is the best learning-based approach on nuPlan. It scores 91.3 in non-reactive and 90.6 in reactive traffic on the Val14 benchmark while being an order of magnitude faster than prior work.

Method: Model systems as Markov decision processes, use inverse reinforcement learning adversaries, and implement three deception strategies (diversionary, targeted, equivocal) through policy optimization with regularization.

Result: All three deception strategies successfully steer adversaries to false beliefs while maintaining at least 97% of optimal non-deceptive reward performance.

Conclusion: The framework provides effective deception mechanisms that protect sensitive information with minimal performance degradation, making it practical for real-world autonomous systems operating in adversarial environments.

Abstract: Autonomous systems are increasingly expected to operate in the presence of adversaries, though adversaries may infer sensitive information simply by observing a system. Therefore, present a deceptive sequential decision-making framework that not only conceals sensitive information, but actively misleads adversaries about it. We model autonomous systems as Markov decision processes, with adversaries using inverse reinforcement learning to recover reward functions. To counter them, we present three regularization strategies for policy synthesis problems that actively deceive an adversary about a system’s reward. Diversionary deception'' leads an adversary to draw any false conclusion about the system's reward function. Targeted deception’’ leads an adversary to draw a specific false conclusion about the system’s reward function. ``Equivocal deception’’ leads an adversary to infer that the real reward and a false reward both explain the system’s behavior. We show how each form of deception can be implemented in policy optimization problems and analytically bound the loss in total accumulated reward induced by deception. Next, we evaluate these developments in a multi-agent setting. We show that diversionary, targeted, and equivocal deception all steer the adversary to false beliefs while still attaining a total accumulated reward that is at least 97% of its optimal, non-deceptive value.

Method: MaskSDM employs a masked training strategy that allows predictions with arbitrary subsets of input variables, uses Shapley values for precise predictor contribution assessment, and is evaluated on the global sPlotOpen dataset with 12,738 plant species.

Result: MaskSDM outperforms imputation-based methods and approximates models trained on specific variable subsets, demonstrating robustness to missing data and flexible predictor selection capabilities.

Conclusion: MaskSDM increases the applicability and adoption of SDMs, laying groundwork for foundation models that can be readily applied to diverse ecological applications with improved flexibility and explainability.

Abstract: Species Distribution Models (SDMs) play a vital role in biodiversity research, conservation planning, and ecological niche modeling by predicting species distributions based on environmental conditions. The selection of predictors is crucial, strongly impacting both model accuracy and how well the predictions reflect ecological patterns. To ensure meaningful insights, input variables must be carefully chosen to match the study objectives and the ecological requirements of the target species. However, existing SDMs, including both traditional and deep learning-based approaches, often lack key capabilities for variable selection: (i) flexibility to choose relevant predictors at inference without retraining; (ii) robustness to handle missing predictor values without compromising accuracy; and (iii) explainability to interpret and accurately quantify each predictor’s contribution. To overcome these limitations, we introduce MaskSDM, a novel deep learning-based SDM that enables flexible predictor selection by employing a masked training strategy. This approach allows the model to make predictions with arbitrary subsets of input variables while remaining robust to missing data. It also provides a clearer understanding of how adding or removing a given predictor affects model performance and predictions. Additionally, MaskSDM leverages Shapley values for precise predictor contribution assessments, improving upon traditional approximations. We evaluate MaskSDM on the global sPlotOpen dataset, modeling the distributions of 12,738 plant species. Our results show that MaskSDM outperforms imputation-based methods and approximates models trained on specific subsets of variables. These findings underscore MaskSDM’s potential to increase the applicability and adoption of SDMs, laying the groundwork for developing foundation models in SDMs that can be readily applied to diverse ecological applications.

Method: MMiC replaces partial parameters within client models inside clusters to mitigate missing modality impact, uses Banzhaf Power Index for optimized client selection, and employs Markovitz Portfolio Optimization for dynamic global aggregation control.

Result: Extensive experiments show MMiC consistently outperforms existing federated learning architectures in both global and personalized performance on multimodal datasets with missing modalities.

Conclusion: MMiC effectively addresses modality incompleteness in MFL, demonstrating superior performance and confirming the viability of the proposed parameter replacement, client selection optimization, and dynamic aggregation approaches.

Abstract: In the era of big data, data mining has become indispensable for uncovering hidden patterns and insights from vast and complex datasets. The integration of multimodal data sources further enhances its potential. Multimodal Federated Learning (MFL) is a distributed approach that enhances the efficiency and quality of multimodal learning, ensuring collaborative work and privacy protection. However, missing modalities pose a significant challenge in MFL, often due to data quality issues or privacy policies across the clients. In this work, we present MMiC, a framework for Mitigating Modality incompleteness in MFL within the Clusters. MMiC replaces partial parameters within client models inside clusters to mitigate the impact of missing modalities. Furthermore, it leverages the Banzhaf Power Index to optimize client selection under these conditions. Finally, MMiC employs an innovative approach to dynamically control global aggregation by utilizing Markovitz Portfolio Optimization. Extensive experiments demonstrate that MMiC consistently outperforms existing federated learning architectures in both global and personalized performance on multimodal datasets with missing modalities, confirming the effectiveness of our proposed solution. Our code is available at https://github.com/gotobcn8/MMiC.

Method: Universal attention feature purification mechanism (MSAP) with inter-scale attention screening and connection method, plus network correction module between layers, tested on embedded deployment system

Result: Extensive experiments on four public datasets show effective reduction of redundant features and excellent performance with minimal resource consumption

Conclusion: MSAP mechanism successfully addresses feature redundancy in multi-scale networks while maintaining low resource consumption, making it suitable for wearable HAR applications

Abstract: In the field of sensor-based Human Activity Recognition (HAR), deep neural networks provide advanced technical support. Many studies have proven that recognition accuracy can be improved by increasing the depth or width of the network. However, for wearable devices, the balance between network performance and resource consumption is crucial. With minimum resource consumption as the basic principle, we propose a universal attention feature purification mechanism, called MSAP, which is suitable for multi-scale networks. The mechanism effectively solves the feature redundancy caused by the superposition of multi-scale features by means of inter-scale attention screening and connection method. In addition, we have designed a network correction module that integrates seamlessly between layers of individual network modules to mitigate inherent problems in deep networks. We also built an embedded deployment system that is in line with the current level of wearable technology to test the practical feasibility of the HAR model, and further prove the efficiency of the method. Extensive experiments on four public datasets show that the proposed method model effectively reduces redundant features in filtered data and provides excellent performance with little resource consumption.

Method: Multi-modal diffusion transformer with specialized architecture to manage signal modalities and continual learning paradigm to incorporate varying modality combinations.

Result: Outperforms task-specific baselines in signal denoising, imputation, and translation. Generated signals match ground-truth performance in detecting abnormalities and estimating vital signs, even in unseen domains.

Conclusion: UniCardio provides a promising approach for AI-assisted healthcare by leveraging complementary cardiovascular signals through unified generative modeling with interpretability.

Abstract: Cardiovascular signals such as photoplethysmography (PPG), electrocardiography (ECG), and blood pressure (BP) are inherently correlated and complementary, together reflecting the health of cardiovascular system. However, their joint utilization in real-time monitoring is severely limited by diverse acquisition challenges from noisy wearable recordings to burdened invasive procedures. Here we propose UniCardio, a multi-modal diffusion transformer that reconstructs low-quality signals and synthesizes unrecorded signals in a unified generative framework. Its key innovations include a specialized model architecture to manage the signal modalities involved in generation tasks and a continual learning paradigm to incorporate varying modality combinations. By exploiting the complementary nature of cardiovascular signals, UniCardio clearly outperforms recent task-specific baselines in signal denoising, imputation, and translation. The generated signals match the performance of ground-truth signals in detecting abnormal health conditions and estimating vital signs, even in unseen domains, while ensuring interpretability for human experts. These advantages position UniCardio as a promising avenue for advancing AI-assisted healthcare.

Method: Analyzed how generator models can exploit spurious attributes favored by LLM judges through different feedback protocols, comparing absolute scoring versus relative preference methods.

Result: Pairwise protocols are 35% vulnerable to preference flipping through distractor features, while absolute scoring shows only 9% vulnerability, making it more robust to manipulation.

Conclusion: Absolute scoring is more reliable than pairwise preferences for LLM evaluation, and feedback protocol choice should be based on dataset characteristics and evaluation objectives to ensure accurate model quality assessment.

Abstract: Large Language Models (LLMs) are widely used as proxies for human labelers in both training (Reinforcement Learning from AI Feedback) and large-scale response evaluation (LLM-as-a-judge). Alignment and evaluation are critical components in the development of reliable LLMs, and the choice of feedback protocol plays a central role in both but remains understudied. In this work, we show that the choice of feedback protocol for evaluation (absolute scores versus relative preferences) can significantly affect evaluation reliability and induce systematic biases. In the context of LLM-as-a-judge evaluation, we show that pairwise protocols are more vulnerable to distracted evaluation. Generator models can exploit spurious attributes (or distractor features) favored by the LLM judge, resulting in inflated scores for lower-quality outputs. We find that absolute scoring is more robust to such manipulation, producing judgments that better reflect response quality and are less influenced by distractor features. Our results demonstrate that generator models can flip preferences by embedding distractor features, skewing LLM-as-a-judge comparisons and leading to inaccurate conclusions about model quality in benchmark evaluations. Pairwise preferences flip in about 35% of the cases, compared to only 9% for absolute scores. We offer recommendations for choosing feedback protocols based on dataset characteristics and evaluation objectives.

Method: Conducted Monte Carlo simulations with synthetic data (2500 simulations per scenario) across various BER values, comparing k-Nearest Neighbor (kNN), Generalized Henze-Penrose (GHP) divergence, and Kernel Density Estimation (KDE) techniques for binary classification.

Result: kNN was overwhelmingly the most accurate non-parametric estimator. To achieve under 5% range for 95% confidence bounds: 1000 samples per class minimum, increasing to 2500 samples per class at 4 features. Other estimators became more accurate with more features but consistently failed to meet target range.

Conclusion: kNN is the recommended BER estimator, though sample requirements increase significantly with dimensionality, highlighting the challenge of accurate BER estimation in high-dimensional problems.

Abstract: The Bayes Error Rate (BER) is the fundamental limit on the achievable generalizable classification accuracy of any machine learning model due to inherent uncertainty within the data. BER estimators offer insight into the difficulty of any classification problem and set expectations for optimal classification performance. In order to be useful, the estimators must also be accurate with a limited number of samples on multivariate problems with unknown class distributions. To determine which estimators meet the minimum requirements for “usefulness”, an in-depth examination of their accuracy is conducted using Monte Carlo simulations with synthetic data in order to obtain their confidence bounds for binary classification. To examine the usability of the estimators for real-world applications, new non-linear multi-modal test scenarios are introduced. In each scenario, 2500 Monte Carlo simulations per scenario are run over a wide range of BER values. In a comparison of k-Nearest Neighbor (kNN), Generalized Henze-Penrose (GHP) divergence and Kernel Density Estimation (KDE) techniques, results show that kNN is overwhelmingly the more accurate non-parametric estimator. In order to reach the target of an under 5% range for the 95% confidence bounds, the minimum number of required samples per class is 1000. As more features are added, more samples are needed, so that 2500 samples per class are required at only 4 features. Other estimators do become more accurate than kNN as more features are added, but continuously fail to meet the target range.

Method: Comprehensive survey organizing current methodologies around four shared performance objectives, reviewing representative works, techniques, and evaluation metrics for each objective.

Result: Objective-centric framework enables meaningful cross-domain comparisons and provides practical insights for selecting/designing foundation model solutions for new IoT tasks.

Conclusion: Identifies key future research directions to advance foundation model applications in IoT, providing guidance for both practitioners and researchers.

Abstract: Foundation models have gained growing interest in the IoT domain due to their reduced reliance on labeled data and strong generalizability across tasks, which address key limitations of traditional machine learning approaches. However, most existing foundation model based methods are developed for specific IoT tasks, making it difficult to compare approaches across IoT domains and limiting guidance for applying them to new tasks. This survey aims to bridge this gap by providing a comprehensive overview of current methodologies and organizing them around four shared performance objectives by different domains: efficiency, context-awareness, safety, and security & privacy. For each objective, we review representative works, summarize commonly-used techniques and evaluation metrics. This objective-centric organization enables meaningful cross-domain comparisons and offers practical insights for selecting and designing foundation model based solutions for new IoT tasks. We conclude with key directions for future research to guide both practitioners and researchers in advancing the use of foundation models in IoT applications.

Method: Introduce multi-exit architecture where each layer has its own prediction branch, enabling simultaneous predictions at multiple depths. Develop a differentiable “learning-to-exit” algorithm to balance contributions from different exits during training.

Result: Multi-exit KANs consistently outperform standard single-exit versions on synthetic functions, dynamical systems, and real-world datasets. Best predictions often come from earlier, simpler exits, revealing smaller, more parsimonious and interpretable models without accuracy loss.

Conclusion: Multi-exit KANs provide a practical solution for achieving both high performance and interpretability in scientific modeling, addressing fundamental challenges in machine learning for scientific discovery through automated discovery of optimal model complexity.

Abstract: Kolmogorov-Arnold Networks (KANs) uniquely combine high accuracy with interpretability, making them valuable for scientific modeling. However, it is unclear a priori how deep a network needs to be for any given task, and deeper KANs can be difficult to optimize and interpret. Here we introduce multi-exit KANs, where each layer includes its own prediction branch, enabling the network to make accurate predictions at multiple depths simultaneously. This architecture provides deep supervision that improves training while discovering the right level of model complexity for each task. Multi-exit KANs consistently outperform standard, single-exit versions on synthetic functions, dynamical systems, and real-world datasets. Remarkably, the best predictions often come from earlier, simpler exits, revealing that these networks naturally identify smaller, more parsimonious and interpretable models without sacrificing accuracy. To automate this discovery, we develop a differentiable “learning-to-exit” algorithm that balances contributions from exits during training. Our approach offers scientists a practical way to achieve both high performance and interpretability, addressing a fundamental challenge in machine learning for scientific discovery.

Method: Uses neurosymbolic approach to construct knowledge graphs from LLM outputs and compares them with ground truth data from Wikidata or contextual documents using graph kernels and semantic clustering.

Result: Achieves competitive accuracy in detecting hallucinations across both open- and closed-domain tasks, and generates contrastive explanations for enhanced transparency.

Conclusion: The framework advances LLM reliability in critical applications and provides foundation for future precision improvements and multi-source knowledge integration.

Abstract: Large Language Models (LLMs) frequently generate hallucinations: statements that are syntactically plausible but lack factual grounding. This research presents KEA (Kernel-Enriched AI) Explain: a neurosymbolic framework that detects and explains such hallucinations by comparing knowledge graphs constructed from LLM outputs with ground truth data from Wikidata or contextual documents. Using graph kernels and semantic clustering, the method provides explanations for detected hallucinations, ensuring both robustness and interpretability. Our framework achieves competitive accuracy in detecting hallucinations across both open- and closed-domain tasks, and is able to generate contrastive explanations, enhancing transparency. This research advances the reliability of LLMs in high-stakes domains and provides a foundation for future work on precision improvements and multi-source knowledge integration.

Method: Leverages differentiable projection onto feasible region by solving KKT conditions of distance minimization problem, reformulates nonlinear KKT conditions via log-exponential transformation to create sparse system with linear and exponential terms.

Result: Achieves strict constraint satisfaction compared to multilayer perceptrons and PINNs, circumvents need to balance data and physics residuals in PINN training, successfully applied to nonconvex pooling problem and real-world chemical process simulation.

Conclusion: Enables reliable integration of domain knowledge into machine learning for hybrid modeling of complex systems with guaranteed constraint satisfaction.

Abstract: Traditional physics-informed neural networks (PINNs) do not guarantee strict constraint satisfaction. This is problematic in engineering systems where minor violations of governing laws can degrade the reliability and consistency of model predictions. In this work, we introduce KKT-Hardnet, a neural network architecture that enforces linear and nonlinear equality and inequality constraints up to machine precision. It leverages a differentiable projection onto the feasible region by solving Karush-Kuhn-Tucker (KKT) conditions of a distance minimization problem. Furthermore, we reformulate the nonlinear KKT conditions via a log-exponential transformation to construct a sparse system with linear and exponential terms. We apply KKT-Hardnet to nonconvex pooling problem and a real-world chemical process simulation. Compared to multilayer perceptrons and PINNs, KKT-Hardnet achieves strict constraint satisfaction. It also circumvents the need to balance data and physics residuals in PINN training. This enables the integration of domain knowledge into machine learning towards reliable hybrid modeling of complex systems.

Method: Pretrained a 20B+ parameter transformer model on extensive user activity data, then fine-tuned for specific applications. Developed Deduplicated Cross-Attention Transformer (DCAT) and infrastructure optimizations to handle scalability requirements.

Result: Achieved 600% throughput improvement on Pinterest internal data, 20% increase in engagement with new items, and successful deployment serving over half a billion users across various applications.

Conclusion: PinFM demonstrates that large-scale foundational models can be effectively applied to industrial recommender systems, overcoming scalability and latency constraints while significantly improving engagement and handling new items effectively.

Abstract: User activity sequences have emerged as one of the most important signals in recommender systems. We present a foundational model, PinFM, for understanding user activity sequences across multiple applications at a billion-scale visual discovery platform. We pretrain a transformer model with 20B+ parameters using extensive user activity data, then fine-tune it for specific applications, efficiently coupling it with existing models. While this pretraining-and-fine-tuning approach has been popular in other domains, such as Vision and NLP, its application in industrial recommender systems presents numerous challenges. The foundational model must be scalable enough to score millions of items every second while meeting tight cost and latency constraints imposed by these systems. Additionally, it should capture the interactions between user activities and other features and handle new items that were not present during the pretraining stage. We developed innovative techniques to address these challenges. Our infrastructure and algorithmic optimizations, such as the Deduplicated Cross-Attention Transformer (DCAT), improved our throughput by 600% on Pinterest internal data. We demonstrate that PinFM can learn interactions between user sequences and candidate items by altering input sequences, leading to a 20% increase in engagement with new items. PinFM is now deployed to help improve the experience of more than half a billion users across various applications.

Method: Reinterpreting VAE through a geometric lens where latent representations are considered as Gaussian distributions (balls) rather than points, and analyzing how KL divergence regularization promotes uniform distribution of encodings and enables contractual mechanisms between encoder and decoder.

Result: The geometric framework demonstrates that proper semantic manifold construction arises from KL divergence constraints on the encoder, and provides a unified perspective showing VQ-VAE as an autoencoder with encodings constrained to cluster centers.

Conclusion: This geometric reinterpretation offers a new intuitive lens for understanding how VAEs shape latent geometry to enable effective generation, complementing the traditional probabilistic view and providing insights into the fundamental mechanisms of variational autoencoders.

Abstract: Variational Autoencoder is typically understood from the perspective of probabilistic inference. In this work, we propose a new geometric reinterpretation which complements the probabilistic view and enhances its intuitiveness. We demonstrate that the proper construction of semantic manifolds arises primarily from the constraining effect of the KL divergence on the encoder. We view the latent representations as a Gaussian ball rather than deterministic points. Under the constraint of KL divergence, Gaussian ball regularizes the latent space, promoting a more uniform distribution of encodings. Furthermore, we show that reparameterization establishes a critical contractual mechanism between the encoder and decoder, enabling the decoder to learn how to reconstruct from these stochastic regions. We further connect this viewpoint with VQ-VAE, offering a unified perspective: VQ-VAE can be seen as an autoencoder where encodings are constrained to a set of cluster centers, with its generative capability arising from the compactness rather than its stochasticity. This geometric framework provides a new lens for understanding how VAE shapes the latent geometry to enable effective generation.

Method: Categorize tokens in each corpus into positive (useful for performance improvement) and negative (lacking essential semantics or misleading). Positive tokens are trained normally while negative tokens are explicitly forgotten through a forgetting process that shapes knowledge boundaries.

Result: Experiments on established benchmarks show the forgetting mechanism improves overall model performance and facilitates more diverse model responses.

Conclusion: Token categorization and selective forgetting of negative tokens helps models learn more precisely and improves SFT effectiveness by reducing reliance on data quality/volume.

Abstract: Supervised fine-tuning (SFT) plays a critical role for pretrained large language models (LLMs), notably enhancing their capacity to acquire domain-specific knowledge while preserving or potentially augmenting their general-purpose capabilities. However, the efficacy of SFT hinges on data quality as well as data volume, otherwise it may result in limited performance gains or even degradation relative to the associated baselines. To mitigate such reliance, we suggest categorizing tokens within each corpus into two parts – positive and negative tokens – based on whether they are useful to improve model performance. Positive tokens can be trained in common ways, whereas negative tokens, which may lack essential semantics or be misleading, should be explicitly forgotten. Overall, the token categorization facilitate the model to learn less informative message, and the forgetting process shapes a knowledge boundary to guide the model on what information to learn more precisely. We conduct experiments on well-established benchmarks, finding that this forgetting mechanism not only improves overall model performance and also facilitate more diverse model responses.

Method: Generalize stochastic interpolants by replacing scalar time variables with vectors, matrices, or linear operators, allowing bridging of probability distributions across multiple dimensional spaces through operator-based interpolants.

Result: The framework demonstrates zero-shot efficacy on conditional generation, inpainting, fine-tuning, posterior sampling, and multiscale modeling, showing potential as a generic task-agnostic alternative to specialized models.

Conclusion: Operator-based interpolants provide a unifying theoretical perspective for existing generative models while extending their capabilities, offering a versatile approach for multiple generative tasks without task-specific training.

Abstract: We propose a framework for learning maps between probability distributions that broadly generalizes the time dynamics of flow and diffusion models. To enable this, we generalize stochastic interpolants by replacing the scalar time variable with vectors, matrices, or linear operators, allowing us to bridge probability distributions across multiple dimensional spaces. This approach enables the construction of versatile generative models capable of fulfilling multiple tasks without task-specific training. Our operator-based interpolants not only provide a unifying theoretical perspective for existing generative models but also extend their capabilities. Through numerical experiments, we demonstrate the zero-shot efficacy of our method on conditional generation and inpainting, fine-tuning and posterior sampling, and multiscale modeling, suggesting its potential as a generic task-agnostic alternative to specialized models.

Method: The authors model the overseer as an agent and characterize incentive-compatible scoring rules as f-mutual information objectives, using bounded f-divergences (e.g., total variation distance) to maintain robustness against adversarial manipulation.

Result: TVD-MI maintains effectiveness with AUC 0.70-0.77 under adversarial attacks, while traditional judge queries degrade to near chance performance (AUC ≈ 0.50). The method decomposes pairwise evaluations into reliable item-level quality scores without requiring ground truth.

Conclusion: Querying LLMs for information relationships rather than quality judgments provides both theoretical and practical robustness against adversarial attacks, addressing key limitations of traditional peer prediction methods without requiring ground truth.

Abstract: We study evaluation of AI systems without ground truth by exploiting a link between strategic gaming and information loss. We analyze which information-theoretic mechanisms resist adversarial manipulation, extending finite-sample bounds to show that bounded f-divergences (e.g., total variation distance) maintain polynomial guarantees under attacks while unbounded measures (e.g., KL divergence) degrade exponentially. To implement these mechanisms, we model the overseer as an agent and characterize incentive-compatible scoring rules as f-mutual information objectives. Under adversarial attacks, TVD-MI maintains effectiveness (area under curve 0.70-0.77) while traditional judge queries are near change (AUC $\approx$ 0.50), demonstrating that querying the same LLM for information relationships rather than quality judgments provides both theoretical and practical robustness. The mechanisms decompose pairwise evaluations into reliable item-level quality scores without ground truth, addressing a key limitation of traditional peer prediction. We release preregistration and code.

Method: Federated learning with feature attention mechanism and temporal modeling structure. Each institution trains local sub-model, parameters are protected with differential privacy and noise injection before aggregation by central server into global model.

Result: Outperforms traditional centralized methods and existing federated learning variants across all evaluation metrics (communication efficiency, model accuracy, systemic risk detection, cross-market generalization).

Conclusion: Provides secure and efficient solution for intelligent financial risk analysis, enhancing risk identification scope and efficiency while preserving data sovereignty in sensitive financial environments.

Abstract: This paper addresses the challenges of data privacy and collaborative modeling in cross-institution financial risk analysis. It proposes a risk assessment framework based on federated learning. Without sharing raw data, the method enables joint modeling and risk identification across multiple institutions. This is achieved by incorporating a feature attention mechanism and temporal modeling structure. Specifically, the model adopts a distributed optimization strategy. Each financial institution trains a local sub-model. The model parameters are protected using differential privacy and noise injection before being uploaded. A central server then aggregates these parameters to generate a global model. This global model is used for systemic risk identification. To validate the effectiveness of the proposed method, multiple experiments are conducted. These evaluate communication efficiency, model accuracy, systemic risk detection, and cross-market generalization. The results show that the proposed model outperforms both traditional centralized methods and existing federated learning variants across all evaluation metrics. It demonstrates strong modeling capabilities and practical value in sensitive financial environments. The method enhances the scope and efficiency of risk identification while preserving data sovereignty. It offers a secure and efficient solution for intelligent financial risk analysis.

Method: CC-Time uses cross-modality learning to model temporal dependency and channel correlations from both time series sequences and text descriptions, plus cross-model fusion to integrate knowledge from PLMs and time series models.

Result: Extensive experiments on nine real-world datasets show CC-Time achieves state-of-the-art prediction accuracy in both full-data training and few-shot learning scenarios.

Conclusion: The proposed cross-modality learning and cross-model fusion approach successfully enhances PLM-based time series forecasting, demonstrating superior performance across various datasets and learning conditions.

Abstract: With the success of pre-trained language models (PLMs) in various application fields beyond natural language processing, language models have raised emerging attention in the field of time series forecasting (TSF) and have shown great prospects. However, current PLM-based TSF methods still fail to achieve satisfactory prediction accuracy matching the strong sequential modeling power of language models. To address this issue, we propose Cross-Model and Cross-Modality Learning with PLMs for time series forecasting (CC-Time). We explore the potential of PLMs for time series forecasting from two aspects: 1) what time series features could be modeled by PLMs, and 2) whether relying solely on PLMs is sufficient for building time series models. In the first aspect, CC-Time incorporates cross-modality learning to model temporal dependency and channel correlations in the language model from both time series sequences and their corresponding text descriptions. In the second aspect, CC-Time further proposes the cross-model fusion block to adaptively integrate knowledge from the PLMs and time series model to form a more comprehensive modeling of time series patterns. Extensive experiments on nine real-world datasets demonstrate that CC-Time achieves state-of-the-art prediction accuracy in both full-data training and few-shot learning situations.

Method: Replaces fixed beta2 with layer-wise dynamic values driven by a bounded “sunspike” ratio (current pooled gradient norm divided by EMA of past norms). Spikes lower beta2 toward beta2_min; calm phases keep it near beta2_max. Includes options like leaky-AMSGrad, trust-region clipping, and various bias-correction modes.

Result: Improves stability and final loss vs fixed-beta2 Adam across four test settings: Transformer PDE surrogate, 3D PINN for heat conduction, synthetic MLX task, and character-level Transformer. On small-enwik8, reduces bits-per-character by ~38% vs Adam-0.95 and ~58% vs Adam-0.999 with smaller variance.

Conclusion: Kourkoutas-Beta provides drop-in replacement for Adam with comparable runtime overhead, preserves Adam-style convergence guarantees, and significantly improves robustness under spiky gradient conditions in physics-based applications.

Abstract: Transformer neural networks are increasingly used for physics-based problems. In data-driven PDE surrogates, training samples from varying boundary and initial conditions can cause erratic losses and spiky gradients; in physics-informed neural networks (PINNs), stiff composite losses amplify this effect. We introduce Kourkoutas-Beta, an Adam-style optimizer where the fixed second-moment discount beta2 is replaced by a layer-wise dynamic value driven by a bounded sunspike'' ratio: the current pooled gradient norm divided by an exponential moving average (EMA) of past norms, squashed to the interval [0,1). Spikes lower beta2 toward beta2_min; calm phases keep it near beta2_max. Options include leaky-AMSGrad (decay), trust-region clipping (max_ratio), adaptive tiny terms, and several bias-correction modes none’’, beta2max'', exact’). With all features off and bias_correction=``none’’, the method is exactly Adam. We test on four settings: (i) a Transformer PDE surrogate (Heat2D), (ii) a 3D PINN for heat conduction (Heat3D), (iii) a lightweight MLX synthetic task with jitter and rare-trigger bursts, and (iv) a character-level Transformer on 30 MB of enwik8 (small-enwik8). Kourkoutas-Beta improves stability and final loss versus fixed-beta2 Adam. On small-enwik8 it lowers bits-per-character by about 38% vs Adam-0.95 and about 58% vs Adam-0.999 over 10 seeds, with smaller variance. The method remains drop-in, with runtime overhead comparable to Adam in testbeds A-C and within single-digit percent in testbed D. It preserves Adam-style convergence guarantees while improving robustness under spiky gradients.

Method: Uses variational mode decomposition (VMD) to break down non-stationary financial time series into smoother subcomponents, then feeds the decomposed data into an LSTM deep learning model for prediction.

Result: The model demonstrates better performance and stability compared to LSTM models trained on raw time series data.

Conclusion: Combining VMD decomposition with deep learning models improves financial time series forecasting accuracy and stability.

Abstract: To address the complexity of financial time series, this paper proposes a forecasting model combining sliding window and variational mode decomposition (VMD) methods. Historical stock prices and relevant market indicators are used to construct datasets. VMD decomposes non-stationary financial time series into smoother subcomponents, improving model adaptability. The decomposed data is then input into a deep learning model for prediction. The study compares the forecasting effects of an LSTM model trained on VMD-processed sequences with those using raw time series, demonstrating better performance and stability.

Method: Proposes MCLPD framework with two phases: 1) Self-supervised pre-training on unlabeled UNM dataset using dual augmentations in time and frequency domains to create contrastive pairs and fuse time-frequency information; 2) Lightweight supervised fine-tuning using only small proportions (1-5%) of labeled data from UI and UC datasets.

Result: Achieves F1 scores of 0.91 on UI and 0.81 on UC using only 1% labeled data, which further improve to 0.97 and 0.87 respectively when 5% labeled data is used. Substantially improves cross-dataset generalization compared to existing methods.

Conclusion: MCLPD effectively enhances cross-dataset PD detection performance while reducing dependency on labeled data, demonstrating the framework’s effectiveness for robust Parkinson’s disease detection from EEG data with limited annotations.

Abstract: Electroencephalography has been validated as an effective technique for detecting Parkinson’s disease,particularly in its early stages.However,the high cost of EEG data annotation often results in limited dataset size and considerable discrepancies across datasets,including differences in acquisition protocols and subject demographics,significantly hinder the robustness and generalizability of models in cross-dataset detection scenarios.To address such challenges,this paper proposes a semi-supervised learning framework named MCLPD,which integrates multi-view contrastive pre-training with lightweight supervised fine-tuning to enhance cross-dataset PD detection performance.During pre-training,MCLPD uses self-supervised learning on the unlabeled UNM dataset.To build contrastive pairs,it applies dual augmentations in both time and frequency domains,which enrich the data and naturally fuse time-frequency information.In the fine-tuning phase,only a small proportion of labeled data from another two datasets (UI and UC)is used for supervised optimization.Experimental results show that MCLPD achieves F1 scores of 0.91 on UI and 0.81 on UC using only 1%of labeled data,which further improve to 0.97 and 0.87,respectively,when 5%of labeled data is used.Compared to existing methods,MCLPD substantially improves cross-dataset generalization while reducing the dependency on labeled data,demonstrating the effectiveness of the proposed framework.

Method: Study Group Relative Policy Optimization (GRPO) fine-tuning across different model sizes and families, comparing four subset selection policies (easy, medium, hard, random) chosen from the same unlabeled pool using base-model difficulty estimates via multi-sample evaluation.

Result: Training on the hardest examples yields the largest performance gains (up to 47%), while training on easy examples yields the smallest gains. Harder examples provide more learnable opportunities during GRPO training.

Conclusion: For budget-constrained post-training, prioritizing hard examples yields substantial performance gains on reasoning tasks when using GRPO.

Abstract: Collecting high-quality training examples for language model fine-tuning is expensive, with practical budgets limiting the amount of data that can be procured. We investigate a critical question for resource-constrained alignment: under a fixed acquisition budget, should practitioners prioritize examples that are easy, medium, hard, or of random difficulty? We study Group Relative Policy Optimization (GRPO) fine-tuning across different model sizes and families, comparing four subset selection policies chosen from the same unlabeled pool using base-model difficulty estimates obtained via multi-sample evaluation. Our experiments reveal that training on the hardest examples yields the largest performance gains, up to 47%, while training on easy examples yield the smallest gains. Analysis reveals that this effect arises from harder examples providing more learnable opportunities during GRPO training. These findings provide practical guidance for budget-constrained post-training: prioritizing hard examples yields substantial performance gains on reasoning tasks when using GRPO.

Method: Develops a unified framework covering several Local-Update SGD-based distributed algorithms with dynamic topologies and non-uniform client selection strategies.

Result: The framework provides improved or matching theoretical convergence guarantees compared to existing work that assumes fixed topologies and uniform aggregation.

Conclusion: Dynamic topologies with non-uniform aggregation strategies outperform traditional fixed-topology approaches in distributed SGD, offering better convergence properties and performance.

Abstract: One of the most common methods to train machine learning algorithms today is the stochastic gradient descent (SGD). In a distributed setting, SGD-based algorithms have been shown to converge theoretically under specific circumstances. A substantial number of works in the distributed SGD setting assume a fixed topology for the edge devices. These papers also assume that the contribution of nodes to the global model is uniform. However, experiments have shown that such assumptions are suboptimal and a non uniform aggregation strategy coupled with a dynamically shifting topology and client selection can significantly improve the performance of such models. This paper details a unified framework that covers several Local-Update SGD-based distributed algorithms with dynamic topologies and provides improved or matching theoretical guarantees on convergence compared to existing work.

Method: The framework includes dynamic capability classification for relevant tool selection, plan-first orchestration with explicit dependencies and human approval options, context-aware task extraction combining dialogue history with external resources, and production-ready execution environments with checkpointing and artifact management.

Result: Demonstrated through two case studies: wind farm monitoring tutorial and deployment at the Advanced Light Source particle accelerator, showing versatility and reliability in high-stakes domains.

Conclusion: Alpha Berkeley Framework establishes itself as a reliable and transparent solution for agentic systems in safety-critical environments, providing scalable workflow coordination with human oversight capabilities.

Abstract: Coordinating workflows across heterogeneous control systems remains a central challenge in safety-critical environments such as scientific facilities, industrial plants, and energy infrastructures. Language-model-driven agents offer a natural interface for these tasks, but existing approaches often lack scalability, reliability, and human oversight. We introduce the Alpha Berkeley Framework, a production-ready architecture for scalable agentic systems that integrate conversational context with robust tool orchestration. The framework features dynamic capability classification to select only relevant tools per task, a plan-first orchestration model that generates execution plans with explicit dependencies and optional human approval, context-aware task extraction that combines dialogue history with external memory and domain resources, and production-ready execution environments with checkpointing, artifact management, and modular deployment. We demonstrate its versatility through two case studies: a tutorial-style wind farm monitoring example and a deployment at the Advanced Light Source particle accelerator. These results establish Alpha Berkeley as a reliable and transparent framework for agentic systems in high-stakes domains.

Method: Uses hierarchical lightweight processes (streams) scheduled in deterministic layers that share a common context. Features evolutionary optimization (single/multi-objective), PyTorch policy integration, tournament tooling with custom scoring, and standardized evaluation metrics.

Result: A practical framework that enables composition of exogenous drivers, endogenous agents, and aggregators without refactoring. The same model can be used for simulation, optimization, and systematic comparison with reproducible results.

Conclusion: HEAS provides a foundation for cross-disciplinary, multi-level inquiry that yields reliable and reproducible results through its unified workflow and standardized evaluation approach.

Abstract: Hierarchical Evolutionary Agent Simulation (HEAS) is a Python framework that unifies layered agent-based modeling with evolutionary optimization and tournament evaluation in a single, reproducible workflow. HEAS represents models as hierarchies of lightweight processes (“streams”) scheduled in deterministic layers that read and write a shared context, making cross-scale couplings explicit and auditable. A compact API and CLI-simulate, optimize, evaluate-expose single- and multi-objective evolution, PyTorch policy integration via parameter flattening/unflattening, and general tournament tooling with user-defined scoring and voting rules. The framework standardizes evaluation through uniform per-step and episode metrics, persists seeds, logbooks, and hall-of-fame archives, and provides plotting helpers for traces, Pareto fronts, and comparative outcomes, reducing glue code and improving comparability across studies. HEAS emphasizes separation of mechanism from orchestration, allowing exogenous drivers, endogenous agents, and aggregators to be composed and swapped without refactoring, while the same model can be used for forward simulation, optimization, or systematic comparison. We illustrate usage with two compact examples-an ecological system and an enterprise decision-making setting. HEAS offers a practical foundation for cross-disciplinary, multi-level inquiry, yielding reliable, reproducible results.

Method: Propose methods that consolidate semantically related actions and events using lexical similarity and embedding-based clustering to reduce annotation noise and align symbolic categories across narrative levels.

Result: Applied to Manga109 dataset, normalization improves coherence and robustness in narrative reasoning tasks (action retrieval, character grounding, event summarization) while maintaining symbolic transparency.

Conclusion: Semantic normalization is a key step toward scalable, cognitively inspired graph models for multimodal narrative understanding.

Abstract: Understanding visual narratives such as comics requires structured representations that capture events, characters, and their relations across multiple levels of story organization. However, symbolic narrative graphs often suffer from inconsistency and redundancy, where similar actions or events are labeled differently across annotations or contexts. Such variance limits the effectiveness of reasoning and generalization. This paper introduces a semantic normalization framework for hierarchical narrative knowledge graphs. Building on cognitively grounded models of narrative comprehension, we propose methods that consolidate semantically related actions and events using lexical similarity and embedding-based clustering. The normalization process reduces annotation noise, aligns symbolic categories across narrative levels, and preserves interpretability. We demonstrate the framework on annotated manga stories from the Manga109 dataset, applying normalization to panel-, event-, and story-level graphs. Preliminary evaluations across narrative reasoning tasks, such as action retrieval, character grounding, and event summarization, show that semantic normalization improves coherence and robustness, while maintaining symbolic transparency. These findings suggest that normalization is a key step toward scalable, cognitively inspired graph models for multimodal narrative understanding.

Method: Uses local edge computing nodes to process real-time user data, generative AI models for personalized artwork responses, cloud-assisted collaboration with universal scene description, ray tracing for rendering, and federated learning for privacy-preserving model improvements.

Result: Creates synchronized shared experiences where digital artworks can respond uniquely to each viewer (e.g., Mona Lisa smiling at one visitor while having Q&A with another) in real time with minimal latency.

Conclusion: Holo-Artisan heralds a new paradigm for cultural heritage interaction by enabling true holographic displays with personalized edge intelligence for immersive multi-user museum experiences.

Abstract: We present Holo-Artisan, a novel system architecture enabling immersive multi-user experiences in virtual museums through true holographic displays and personalized edge intelligence. In our design, local edge computing nodes process real-time user data – including pose, facial expression, and voice – for multiple visitors concurrently. Generative AI models then drive digital artworks (e.g., a volumetric Mona Lisa) to respond uniquely to each viewer. For instance, the Mona Lisa can return a smile to one visitor while engaging in a spoken Q&A with another, all in real time. A cloud-assisted collaboration platform composes these interactions in a shared scene using a universal scene description, and employs ray tracing to render high-fidelity, personalized views with a direct pipeline to glasses-free holographic displays. To preserve user privacy and continuously improve personalization, we integrate federated learning (FL) – edge devices locally fine-tune AI models and share only model updates for aggregation. This edge-centric approach minimizes latency and bandwidth usage, ensuring a synchronized shared experience with individual customization. Through Holo-Artisan, static museum exhibits are transformed into dynamic, living artworks that engage each visitor in a personal dialogue, heralding a new paradigm of cultural heritage interaction.

Method: The authors developed improvements to the adder-viz software for visualizing real-time event transcode processes and applications, making it available as MIT-licensed software in a centralized repository.

Result: Enhanced visualization capabilities for real-time event transcode processes using the unified ADΔER representation for neuromorphic event video.

Conclusion: The improved adder-viz software provides better tools for working with event video data using the ADΔER representation, addressing previous limitations in event camera representations.

Abstract: Recent years have brought about a surge in neuromorphic ``event’’ video research, primarily targeting computer vision applications. Event video eschews video frames in favor of asynchronous, per-pixel intensity samples. While much work has focused on a handful of representations for specific event cameras, these representations have shown limitations in flexibility, speed, and compressibility. We previously proposed the unified AD{\Delta}ER representation to address these concerns. This paper introduces numerous improvements to the \textit{adder-viz} software for visualizing real-time event transcode processes and applications in-the-loop. The MIT-licensed software is available from a centralized repository at \href{https://github.com/ac-freeman/adder-codec-rs}{https://github.com/ac-freeman/adder-codec-rs}.

Method: Combines multiple LiDARs and cameras for live dynamic point cloud capture, integrates with pre-captured static point clouds, and automatically adjusts static cloud colors to current lighting conditions.

Result: System achieves 30 fps across wide-area scenes with latency below 100 ms, demonstrated through real-world deployment in a tourist site.

Conclusion: Proposed system effectively addresses outdoor 3D capture challenges by combining dynamic and static point clouds with automatic lighting compensation.

Abstract: Various real-time methods for capturing and transmitting dynamic 3D spaces have been proposed, including those based on RGB-D cameras and volumetric capture. However, applying existing methods to outdoor tourist sites remains difficult because maintenance and aesthetic constraints limit sensor placement, and daylight variability complicates processing. We propose a system that combines multiple LiDARs and cameras for live dynamic point cloud capture, and integrates them with pre-captured static point clouds for wide-area 3D visualization. The system sustains 30 fps across wide-area scenes while keeping latency below 100 ms. To mitigate lighting inconsistencies, static point-cloud colors are automatically adjusted to current lighting. The effectiveness of our system is demonstrated through real-world deployment in a tourist site.

Method: Created first Chinese HF speech database with paired pre/post-hospitalization recordings. Used patient-wise and pair-wise classification, plus adaptive frequency filter for frequency importance analysis.

Result: Confirmed Chinese language effectiveness for HF detection. Pair-wise classification serves as ideal speaker-decoupled baseline. Statistical tests show individual differences are key accuracy contributors.

Conclusion: Chinese speech contains valuable HF information. Personalized approaches and frequency analysis are effective. Database published for future research.

Abstract: Speech is a cost-effective and non-intrusive data source for identifying acute and chronic heart failure (HF). However, there is a lack of research on whether Chinese syllables contain HF-related information, as observed in other well-studied languages. This study presents the first Chinese speech database of HF patients, featuring paired recordings taken before and after hospitalisation. The findings confirm the effectiveness of the Chinese language in HF detection using both standard ‘patient-wise’ and personalised ‘pair-wise’ classification approaches, with the latter serving as an ideal speaker-decoupled baseline for future research. Statistical tests and classification results highlight individual differences as key contributors to inaccuracy. Additionally, an adaptive frequency filter (AFF) is proposed for frequency importance analysis. The data and demonstrations are published at https://github.com/panyue1998/Voice_HF.

Method: Three-component architecture: 1) frozen Whisper-large-v3 speech encoder for acoustic features, 2) trainable Linear-ReLU-Linear adaptor for speech-text alignment, 3) frozen Qwen2.5-7B-Instruct LLM with trainable LoRA for linguistic decoding.

Result: Achieved word/character error rate of 9.83% across 11 languages in evaluation set, ranking third place among global participants.

Conclusion: The systematic combination of pretrained models with task-specific fine-tuning proves effective for multilingual ASR, demonstrating strong cross-lingual performance with minimal error rates.

Abstract: This paper presents the architecture and performance of a novel Multilingual Automatic Speech Recognition (ASR) system developed by the Transsion Speech Team for Track 1 of the MLC-SLM 2025 Challenge. The proposed system comprises three key components: 1) a frozen Whisper-large-v3 based speech encoder, leveraging large-scale pretraining to ensure robust acoustic feature extraction; 2) a trainable adaptor module using Linear-ReLU-Linear transformation mechanisms to effectively align speech and text representations; and 3) a frozen Qwen2.5-7B-Instruct large language model (LLM) integrated with trainable LoRA for optimized contextual linguistic decoding. By systematically combining pretrained models with task specific fine-tuning, the system achieved a word/character error rate (WER/CER) of 9.83% across 11 languages in the evaluation set and ranked third place among global participants.

Method: Proposes GOAT framework with uncertainty analysis showing correlation between hallucination and model uncertainty. Reformulates TTS generation as trajectory flow optimization using enhanced Subtrajectory Balance objective and sharpened internal reward as target distribution. Integrates reward temperature decay and learning rate optimization.

Result: Reduces over 50% character error rates on challenging test cases and lowers uncertainty by up to 58%. Demonstrates strong generalization ability and effectiveness.

Conclusion: GOAT effectively mitigates hallucinations in LM-based TTS without requiring massive resources or adding inference cost, providing a practical post-training solution.

Abstract: Language Model (LM)-based Text-to-Speech (TTS) systems often generate hallucinated speech that deviates from input text. Existing mitigation strategies either demand excessive training resources or introduce significant inference latency. In this paper, we propose GFlOwNet-guided distribution AlignmenT (GOAT) for LM-based TTS, a post-training framework that mitigates hallucinations without relying on massive resources or inference cost. Specifically, we first conduct an uncertainty analysis, revealing a strong positive correlation between hallucination and model uncertainty. Based on this, we reformulate TTS generation as a trajectory flow optimization problem and introduce an enhanced Subtrajectory Balance objective together with a sharpened internal reward as target distribution. We further integrate reward temperature decay and learning rate optimization for stability and performance balance. Extensive experiments show that GOAT reduce over 50% character error rates on challenging test cases and lowering uncertainty by up to 58%, demonstrating its strong generalization ability and effectiveness.

Method: Collected 64-channel EEG data from 122 participants during four speech conditions. Used alpha oscillations as biomarkers. Developed EffortNet with three learning paradigms: self-supervised learning, incremental learning, and transfer learning to handle individual variability.

Result: Alpha oscillations showed significantly higher power during noisy speech. EffortNet achieved 80.9% classification accuracy with only 40% training data, outperforming CNN (62.3%) and STAnet (61.1%). Transformer-enhanced speech elicited neural responses more similar to clean speech than MMSE-enhanced speech.

Conclusion: EffortNet provides a practical solution for personalized hearing technology assessment and enables cognitive-aware speech enhancement system design.

Abstract: This paper presents EffortNet, a novel deep learning framework for decoding individual listening effort from electroencephalography (EEG) during speech comprehension. Listening effort represents a significant challenge in speech-hearing research, particularly for aging populations and those with hearing impairment. We collected 64-channel EEG data from 122 participants during speech comprehension under four conditions: clean, noisy, MMSE-enhanced, and Transformer-enhanced speech. Statistical analyses confirmed that alpha oscillations (8-13 Hz) exhibited significantly higher power during noisy speech processing compared to clean or enhanced conditions, confirming their validity as objective biomarkers of listening effort. To address the substantial inter-individual variability in EEG signals, EffortNet integrates three complementary learning paradigms: self-supervised learning to leverage unlabeled data, incremental learning for progressive adaptation to individual characteristics, and transfer learning for efficient knowledge transfer to new subjects. Our experimental results demonstrate that Effort- Net achieves 80.9% classification accuracy with only 40% training data from new subjects, significantly outperforming conventional CNN (62.3%) and STAnet (61.1%) models. The probability-based metric derived from our model revealed that Transformer-enhanced speech elicited neural responses more similar to clean speech than MMSEenhanced speech. This finding contrasted with subjective intelligibility ratings but aligned with objective metrics. The proposed framework provides a practical solution for personalized assessment of hearing technologies, with implications for designing cognitive-aware speech enhancement systems.

Method: VersBand uses four specialized models: VocalBand (flow-matching for vocals), AccompBand (flow-based transformer with Band-MOE for accompaniments), LyricBand (lyrics generation), and MelodyBand (melody generation) to enable comprehensive multi-task song generation.

Result: Experimental results show VersBand outperforms baseline models across multiple song generation tasks using both objective and subjective metrics.

Conclusion: VersBand successfully addresses the challenges of prompt-based control, alignment, and multi-task support in song generation, providing a comprehensive framework for high-quality song synthesis.

Abstract: Song generation focuses on producing controllable high-quality songs based on various prompts. However, existing methods struggle to generate vocals and accompaniments with prompt-based control and proper alignment. Additionally, they fall short in supporting various tasks. To address these challenges, we introduce VersBand, a multi-task song generation framework for synthesizing high-quality, aligned songs with prompt-based control. VersBand comprises these primary models: 1) VocalBand, a decoupled model, leverages the flow-matching method for generating singing styles, pitches, and mel-spectrograms, allowing fast, high-quality vocal generation with style control. 2) AccompBand, a flow-based transformer model, incorporates the Band-MOE, selecting suitable experts for enhanced quality, alignment, and control. This model allows for generating controllable, high-quality accompaniments aligned with vocals. 3) Two generation models, LyricBand for lyrics and MelodyBand for melodies, contribute to the comprehensive multi-task song generation system, allowing for extensive control based on multiple prompts. Experimental results show that VersBand outperforms baseline models across multiple song generation tasks using objective and subjective metrics. Demos and codes are available at https://aaronz345.github.io/VersBandDemo and https://github.com/AaronZ345/VersBand.

Method: Benchmarked diverse fine-tuning methods including full fine-tuning and parameter-efficient fine-tuning (PEFT) techniques. Evaluated impact on FID, Vendi score, prompt-image alignment, and downstream classification performance using synthetic images for training.

Result: Specific fine-tuning strategies improved both generation fidelity and downstream performance when synthetic images were used for classifier training and evaluation on real images.

Conclusion: Fine-tuning techniques significantly impact high-resolution image generation quality and downstream utility, with certain strategies proving particularly effective for improving both generation metrics and practical application performance.

Abstract: Advancements in diffusion-based foundation models have improved text-to-image generation, yet most efforts have been limited to low-resolution settings. As high-resolution image synthesis becomes increasingly essential for various applications, particularly in medical imaging domains, fine-tuning emerges as a crucial mechanism for adapting these powerful pre-trained models to task-specific requirements and data distributions. In this work, we present a systematic study, examining the impact of various fine-tuning techniques on image generation quality when scaling to high resolution 512x512 pixels. We benchmark a diverse set of fine-tuning methods, including full fine-tuning strategies and parameter-efficient fine-tuning (PEFT). We dissect how different fine-tuning methods influence key quality metrics, including Fr'echet Inception Distance (FID), Vendi score, and prompt-image alignment. We also evaluate the utility of generated images in a downstream classification task under data-scarce conditions, demonstrating that specific fine-tuning strategies improve both generation fidelity and downstream performance when synthetic images are used for classifier training and evaluation on real images. Our code is accessible through the project website - https://tehraninasab.github.io/PixelUPressure/.

Method: Multi-teacher knowledge distillation approach with novel diffusion-based data augmentation to enhance generalization while reducing model size. The trained model is packaged as both online and offline segmentation tools.

Result: Student model achieves 95.22% mIoU segmentation performance despite 96.6% parameter reduction compared to teacher models. Case study shows tongue patches yield higher classification performance and better interpretability than original images.

Conclusion: Successfully developed the first open-source, freely available tongue image segmentation tool that enables accurate segmentation with minimal computational requirements, making it accessible to TCM practitioners and researchers without programming expertise.

Abstract: Tongue imaging serves as a valuable diagnostic tool, particularly in Traditional Chinese Medicine (TCM). The quality of tongue surface segmentation significantly affects the accuracy of tongue image classification and subsequent diagnosis in intelligent tongue diagnosis systems. However, existing research on tongue image segmentation faces notable limitations, and there is a lack of robust and user-friendly segmentation tools. This paper proposes a tongue image segmentation model (TOM) based on multi-teacher knowledge distillation. By incorporating a novel diffusion-based data augmentation method, we enhanced the generalization ability of the segmentation model while reducing its parameter size. Notably, after reducing the parameter count by 96.6% compared to the teacher models, the student model still achieves an impressive segmentation performance of 95.22% mIoU. Furthermore, we packaged and deployed the trained model as both an online and offline segmentation tool (available at https://itongue.cn/), allowing TCM practitioners and researchers to use it without any programming experience. We also present a case study on TCM constitution classification using segmented tongue patches. Experimental results demonstrate that training with tongue patches yields higher classification performance and better interpretability than original tongue images. To our knowledge, this is the first open-source and freely available tongue image segmentation tool.

Method: Used patient-specific cerebrovascular in-silico models converted to synthetic MR images. Implemented dedicated GAN architecture and evaluated three adversarial loss functions: Vanilla, Relativistic, and Wasserstein GANs, comparing against non-adversarial generator-only training.

Result: Wasserstein GAN achieved best results with 6.9% vNRMSE vs 9.6% for non-adversarial reference. Vanilla and Relativistic GANs were unstable (8.1% and 7.8% vs 7.2% generator-only). Wasserstein GAN also outperformed at low SNR (8.7% vs 10.7%).

Conclusion: GAN-based super-resolution enhances 4D Flow MRI, particularly in cerebrovascular regions, but implementation specifics and adversarial strategy selection are critical for stable training and optimal performance.

Abstract: 4D Flow Magnetic Resonance Imaging (4D Flow MRI) enables non-invasive quantification of blood flow and hemodynamic parameters. However, its clinical application is limited by low spatial resolution and noise, particularly affecting near-wall velocity measurements. Machine learning-based super-resolution has shown promise in addressing these limitations, but challenges remain, not least in recovering near-wall velocities. Generative adversarial networks (GANs) offer a compelling solution, having demonstrated strong capabilities in restoring sharp boundaries in non-medical super-resolution tasks. Yet, their application in 4D Flow MRI remains unexplored, with implementation challenged by known issues such as training instability and non-convergence. In this study, we investigate GAN-based super-resolution in 4D Flow MRI. Training and validation were conducted using patient-specific cerebrovascular in-silico models, converted into synthetic images via an MR-true reconstruction pipeline. A dedicated GAN architecture was implemented and evaluated across three adversarial loss functions: Vanilla, Relativistic, and Wasserstein. Our results demonstrate that the proposed GAN improved near-wall velocity recovery compared to a non-adversarial reference (vNRMSE: 6.9% vs. 9.6%); however, that implementation specifics are critical for stable network training. While Vanilla and Relativistic GANs proved unstable compared to generator-only training (vNRMSE: 8.1% and 7.8% vs. 7.2%), a Wasserstein GAN demonstrated optimal stability and incremental improvement (vNRMSE: 6.9% vs. 7.2%). The Wasserstein GAN further outperformed the generator-only baseline at low SNR (vNRMSE: 8.7% vs. 10.7%). These findings highlight the potential of GAN-based super-resolution in enhancing 4D Flow MRI, particularly in challenging cerebrovascular regions, while emphasizing the need for careful selection of adversarial strategies.

Method: Uses probabilistic reconstruction to estimate image uncertainty, propagates it through analysis pipelines to clinical metrics, applies conformal prediction for calibrated confidence intervals, and iteratively samples k-space until meeting user-defined precision targets.

Result: Validated on knee and cardiac MRI datasets, the framework reduces scan times compared to fixed protocols while providing formal statistical guarantees on final image precision.

Conclusion: This approach enables patient-specific acquisitions that balance scan efficiency with diagnostic confidence, representing a critical step towards personalized and resource-efficient MRI.

Abstract: Magnetic Resonance Imaging (MRI) offers unparalleled soft-tissue contrast but is fundamentally limited by long acquisition times. While deep learning-based accelerated MRI can dramatically shorten scan times, the reconstruction from undersampled data introduces ambiguity resulting from an ill-posed problem with infinitely many possible solutions that propagates to downstream clinical tasks. This uncertainty is usually ignored during the acquisition process as acceleration factors are often fixed a priori, resulting in scans that are either unnecessarily long or of insufficient quality for a given clinical endpoint. This work introduces a dynamic, uncertainty-aware acquisition framework that adjusts scan time on a per-subject basis. Our method leverages a probabilistic reconstruction model to estimate image uncertainty, which is then propagated through a full analysis pipeline to a quantitative metric of interest (e.g., patellar cartilage volume or cardiac ejection fraction). We use conformal prediction to transform this uncertainty into a rigorous, calibrated confidence interval for the metric. During acquisition, the system iteratively samples k-space, updates the reconstruction, and evaluates the confidence interval. The scan terminates automatically once the uncertainty meets a user-predefined precision target. We validate our framework on both knee and cardiac MRI datasets. Our results demonstrate that this adaptive approach reduces scan times compared to fixed protocols while providing formal statistical guarantees on the precision of the final image. This framework moves beyond fixed acceleration factors, enabling patient-specific acquisitions that balance scan efficiency with diagnostic confidence, a critical step towards personalized and resource-efficient MRI.

Method: Survey existing event-based video systems, discuss technical issues from recent scalable event streaming work, and propose a new low-latency event streaming format based on the latest Media Over QUIC protocol draft.

Result: The paper identifies the gap in transmission solutions for event-based video and develops a streaming approach specifically designed for asynchronous pixel samples from neuromorphic sensors.

Conclusion: A new streaming format is needed for event-based video data that differs fundamentally from traditional lossy compression methods, and the proposed Media Over QUIC-based solution addresses the low-latency requirements of neuromorphic sensor data transmission.

Abstract: Lossy compression and rate-adaptive streaming are a mainstay in traditional video steams. However, a new class of neuromorphic ``event’’ sensors records video with asynchronous pixel samples rather than image frames. These sensors are designed for computer vision applications, rather than human video consumption. Until now, researchers have focused their efforts primarily on application development, ignoring the crucial problem of data transmission. We survey the landscape of event-based video systems, discuss the technical issues with our recent scalable event streaming work, and propose a new low-latency event streaming format based on the latest additions to the Media Over QUIC protocol draft.

Method: Investigates wavelet transform-based denoising methods, systematically evaluating optimal combinations of threshold values, decomposition levels, and wavelet types for noise mitigation in medical images.

Result: Bior6.8 biorthogonal wavelet with universal thresholding at decomposition levels 2-3 consistently achieves optimal denoising performance, providing significant noise reduction while preserving essential anatomical structures.

Conclusion: The identified wavelet-based denoising approach effectively mitigates noise in medical images while maintaining critical diagnostic features, enhancing clinical decision-making capabilities.

Abstract: Medical imaging modalities including magnetic resonance imaging (MRI), computed tomography (CT), and ultrasound are essential for accurate diagnosis and treatment planning in modern healthcare. However, noise contamination during image acquisition and processing frequently degrades image quality, obscuring critical diagnostic details and compromising clinical decision-making. Additionally, enhancement techniques such as histogram equalization may inadvertently amplify existing noise artifacts, including salt-and-pepper distortions. This study investigates wavelet transform-based denoising methods for effective noise mitigation in medical images, with the primary objective of identifying optimal combinations of threshold values, decomposition levels, and wavelet types to achieve superior denoising performance and enhanced diagnostic accuracy. Through systematic evaluation across various noise conditions, the research demonstrates that the bior6.8 biorthogonal wavelet with universal thresholding at decomposition levels 2-3 consistently achieves optimal denoising performance, providing significant noise reduction while preserving essential anatomical structures and diagnostic features critical for clinical applications.

Method: Proposes SPIRiT regularization that combines compressed sensing with both linear predictability and sensitivity encoding parallel imaging methods.

Result: Demonstrates improved reconstructed images on brain, knee, and ankle data compared to individual methods.

Conclusion: The combined approach with SPIRiT regularization effectively enhances MRI reconstruction quality from accelerated scans.

Abstract: Accelerated Magnetic Resonance Imaging (MRI) permits high quality images from fewer samples that can be collected with a faster scan. Two established methods for accelerating MRI include parallel imaging and compressed sensing. Two types of parallel imaging include linear predictability, which assumes that the Fourier samples are linearly related, and sensitivity encoding, which incorporates a priori knowledge of the sensitivity maps. In this work, we combine compressed sensing with both types of parallel imaging using a novel regularization term: SPIRiT regularization. When combined, the reconstructed images are improved. We demonstrate results on data of a brain, a knee, and an ankle.

Method: Train diffusion model on large-scale 2D X-ray projections, use per-projection adaptive sampling to generate HR projections, employ 3D Gaussian splatting for volume reconstruction, and introduce negative alpha blending for residual learning between LR and diffusion-based projections.

Result: Superior quantitative and qualitative results for 3D CT super-resolution on two datasets, demonstrating effective recovery of fine anatomical details without requiring paired training data.

Conclusion: The proposed framework successfully combines diffusion-generated 2D priors with 3D reconstruction techniques to achieve state-of-the-art zero-shot CT super-resolution, addressing key limitations of existing methods.

Abstract: Computed tomography (CT) is widely used in clinical diagnosis, but acquiring high-resolution (HR) CT is limited by radiation exposure risks. Deep learning-based super-resolution (SR) methods have been studied to reconstruct HR from low-resolution (LR) inputs. While supervised SR approaches have shown promising results, they require large-scale paired LR-HR volume datasets that are often unavailable. In contrast, zero-shot methods alleviate the need for paired data by using only a single LR input, but typically struggle to recover fine anatomical details due to limited internal information. To overcome these, we propose a novel zero-shot 3D CT SR framework that leverages upsampled 2D X-ray projection priors generated by a diffusion model. Exploiting the abundance of HR 2D X-ray data, we train a diffusion model on large-scale 2D X-ray projection and introduce a per-projection adaptive sampling strategy. It selects the generative process for each projection, thus providing HR projections as strong external priors for 3D CT reconstruction. These projections serve as inputs to 3D Gaussian splatting for reconstructing a 3D CT volume. Furthermore, we propose negative alpha blending (NAB-GS) that allows negative values in Gaussian density representation. NAB-GS enables residual learning between LR and diffusion-based projections, thereby enhancing high-frequency structure reconstruction. Experiments on two datasets show that our method achieves superior quantitative and qualitative results for 3D CT SR.

Method: Combines vision-language model with diffusion model, utilizing pathology-related keywords associated with normal tissues to guide the reconstruction process and facilitate differentiation between normal and abnormal tissues.

Result: Experiments on gastric lymph node dataset and public breast lymph node dataset show promising performance and generalization ability under domain shift across various organs.

Conclusion: The proposed method demonstrates potential for effective unsupervised anomaly detection in digital pathology by leveraging vision-language guidance with diffusion models.

Abstract: Anomaly detection is an emerging approach in digital pathology for its ability to efficiently and effectively utilize data for disease diagnosis. While supervised learning approaches deliver high accuracy, they rely on extensively annotated datasets, suffering from data scarcity in digital pathology. Unsupervised anomaly detection, however, offers a viable alternative by identifying deviations from normal tissue distributions without requiring exhaustive annotations. Recently, denoising diffusion probabilistic models have gained popularity in unsupervised anomaly detection, achieving promising performance in both natural and medical imaging datasets. Building on this, we incorporate a vision-language model with a diffusion model for unsupervised anomaly detection in digital pathology, utilizing histopathology prompts during reconstruction. Our approach employs a set of pathology-related keywords associated with normal tissues to guide the reconstruction process, facilitating the differentiation between normal and abnormal tissues. To evaluate the effectiveness of the proposed method, we conduct experiments on a gastric lymph node dataset from a local hospital and assess its generalization ability under domain shift using a public breast lymph node dataset. The experimental results highlight the potential of the proposed method for unsupervised anomaly detection across various organs in digital pathology. Code: https://github.com/QuIIL/AnoPILaD.

Method: Used knowledge distillation with hybrid supervision (ground-truth labels + teacher soft targets), employed VGG19, Visformer-S, and AutoFormer-V2-T as teacher models, and OFA-595 supernet as student. Validated on COVID-QU-Ex and LCS25000 datasets with Score-CAM visualizations for interpretability.

Result: The distilled student model maintained high classification performance with significantly reduced parameters and inference time compared to teacher models.

Conclusion: Knowledge distillation enables creation of compact, efficient models suitable for clinical deployment while maintaining accuracy and providing explainability through visualizations, making them trustworthy for medical AI applications.

Abstract: This study comprehensively explores knowledge distillation frameworks for COVID-19 and lung cancer classification using chest X-ray (CXR) images. We employ high-capacity teacher models, including VGG19 and lightweight Vision Transformers (Visformer-S and AutoFormer-V2-T), to guide the training of a compact, hardware-aware student model derived from the OFA-595 supernet. Our approach leverages hybrid supervision, combining ground-truth labels with teacher models’ soft targets to balance accuracy and computational efficiency. We validate our models on two benchmark datasets: COVID-QU-Ex and LCS25000, covering multiple classes, including COVID-19, healthy, non-COVID pneumonia, lung, and colon cancer. To interpret the spatial focus of the models, we employ Score-CAM-based visualizations, which provide insight into the reasoning process of both teacher and student networks. The results demonstrate that the distilled student model maintains high classification performance with significantly reduced parameters and inference time, making it an optimal choice in resource-constrained clinical environments. Our work underscores the importance of combining model efficiency with explainability for practical, trustworthy medical AI solutions.

Method: Integrated multi-task deep learning framework with: 1) EfficientNet-B0 + CBAM for classification, 2) ResNet34-UNet++ with self-attention and attention gating for segmentation, 3) ConvNeXt-Tiny for molecular subtyping (HER-2, Ki-67 markers). Includes Gradio-based online platform with multi-format uploads and bilingual interfaces.

Result: Outstanding performance: 93.28% accuracy, 82.05% F1-score, 96.41% AUC for classification; 0.9091 Dice coefficient for segmentation. Platform significantly improved diagnostic accuracy, efficiency, and accessibility.

Conclusion: The framework and online tool advance intelligent bladder cancer diagnosis by improving clinical reliability, supporting early detection, and enabling real-time feedback, representing significant progress toward AI-assisted urology decision-making.

Abstract: Clinical cystoscopy, the current standard for bladder cancer diagnosis, suffers from significant reliance on physician expertise, leading to variability and subjectivity in diagnostic outcomes. There is an urgent need for objective, accurate, and efficient computational approaches to improve bladder cancer diagnostics. Leveraging recent advancements in deep learning, this study proposes an integrated multi-task deep learning framework specifically designed for bladder cancer diagnosis from cystoscopic images. Our framework includes a robust classification model using EfficientNet-B0 enhanced with Convolutional Block Attention Module (CBAM), an advanced segmentation model based on ResNet34-UNet++ architecture with self-attention mechanisms and attention gating, and molecular subtyping using ConvNeXt-Tiny to classify molecular markers such as HER-2 and Ki-67. Additionally, we introduce a Gradio-based online diagnostic platform integrating all developed models, providing intuitive features including multi-format image uploads, bilingual interfaces, and dynamic threshold adjustments. Extensive experimentation demonstrates the effectiveness of our methods, achieving outstanding accuracy (93.28%), F1-score (82.05%), and AUC (96.41%) for classification tasks, and exceptional segmentation performance indicated by a Dice coefficient of 0.9091. The online platform significantly improved the accuracy, efficiency, and accessibility of clinical bladder cancer diagnostics, enabling practical and user-friendly deployment. The code is publicly available. Our multi-task framework and integrated online tool collectively advance the field of intelligent bladder cancer diagnosis by improving clinical reliability, supporting early tumor detection, and enabling real-time diagnostic feedback. These contributions mark a significant step toward AI-assisted decision-making in urology.

Method: Fine-tune only batch normalization and fully connected layers while preserving pretrained convolutional filters, combined with adversarial training scheme for enhanced cross-domain generalization.

Result: BN layers identified as primary source of domain dependence; DoSReMC significantly improves cross-domain performance across three large-scale FFDM datasets including a new pathologically confirmed dataset.

Conclusion: DoSReMC provides a practical, easily implementable solution for robust mammography classification that can be integrated into existing AI pipelines across diverse clinical environments.

Abstract: Numerous deep learning-based solutions have been developed for the automatic recognition of breast cancer using mammography images. However, their performance often declines when applied to data from different domains, primarily due to domain shift - the variation in data distributions between source and target domains. This performance drop limits the safe and equitable deployment of AI in real-world clinical settings. In this study, we present DoSReMC (Domain Shift Resilient Mammography Classification), a batch normalization (BN) adaptation framework designed to enhance cross-domain generalization without retraining the entire model. Using three large-scale full-field digital mammography (FFDM) datasets - including HCTP, a newly introduced, pathologically confirmed in-house dataset - we conduct a systematic cross-domain evaluation with convolutional neural networks (CNNs). Our results demonstrate that BN layers are a primary source of domain dependence: they perform effectively when training and testing occur within the same domain, and they significantly impair model generalization under domain shift. DoSReMC addresses this limitation by fine-tuning only the BN and fully connected (FC) layers, while preserving pretrained convolutional filters. We further integrate this targeted adaptation with an adversarial training scheme, yielding additional improvements in cross-domain generalizability. DoSReMC can be readily incorporated into existing AI pipelines and applied across diverse clinical environments, providing a practical pathway toward more robust and generalizable mammography classification systems.

Method: Proposes DECSC framework that unifies local spatial-spectral correlations (3D CSC), nonlocal spatial self-similarities (transformer blocks), and global spatial consistency (2D CSC) within Deep Equilibrium models that treat networks as infinite-depth fixed-point solutions.

Result: Experimental results demonstrate superior denoising performance compared to state-of-the-art methods.

Conclusion: The DECSC framework effectively addresses HSI denoising with convergence guarantees by combining convolutional sparse coding with deep equilibrium modeling and transformer-based nonlocal similarity exploitation.

Abstract: Hyperspectral images (HSIs) play a crucial role in remote sensing but are often degraded by complex noise patterns. Ensuring the physical property of the denoised HSIs is vital for robust HSI denoising, giving the rise of deep unfolding-based methods. However, these methods map the optimization of a physical model to a learnable network with a predefined depth, which lacks convergence guarantees. In contrast, Deep Equilibrium (DEQ) models treat the hidden layers of deep networks as the solution to a fixed-point problem and models them as infinite-depth networks, naturally consistent with the optimization. Under the framework of DEQ, we propose a Deep Equilibrium Convolutional Sparse Coding (DECSC) framework that unifies local spatial-spectral correlations, nonlocal spatial self-similarities, and global spatial consistency for robust HSI denoising. Within the convolutional sparse coding (CSC) framework, we enforce shared 2D convolutional sparse representation to ensure global spatial consistency across bands, while unshared 3D convolutional sparse representation captures local spatial-spectral details. To further exploit nonlocal self-similarities, a transformer block is embedded after the 2D CSC. Additionally, a detail enhancement module is integrated with the 3D CSC to promote image detail preservation. We formulate the proximal gradient descent of the CSC model as a fixed-point problem and transform the iterative updates into a learnable network architecture within the framework of DEQ. Experimental results demonstrate that our DECSC method achieves superior denoising performance compared to state-of-the-art methods.

Method: Evaluated three models for generating virtual DES images: pre-trained U-Net, end-to-end trained U-Net, and CycleGAN. Assessed impact on classification of CESM examinations into malignant vs non-malignant categories.

Result: Pre-trained U-Net performed best with 85.59% F1 score using virtual DES images, compared to 90.35% with real DES images. Real DES images contain additional diagnostic information that improves classification accuracy.

Conclusion: Virtual DES image generation shows considerable potential. Future advancements may narrow the performance gap to make exclusive reliance on virtual DES images clinically viable, reducing radiation exposure while maintaining diagnostic quality.

Abstract: Contrast-enhanced spectral mammography (CESM) is an imaging modality that provides two types of images, commonly known as low-energy (LE) and dual-energy subtracted (DES) images. In many domains, particularly in medicine, the emergence of image-to-image translation techniques has enabled the artificial generation of images using other images as input. Within CESM, applying such techniques to generate DES images from LE images could be highly beneficial, potentially reducing patient exposure to radiation associated with high-energy image acquisition. In this study, we investigated three models for the artificial generation of DES images (virtual DES): a pre-trained U-Net model, a U-Net trained end-to-end model, and a CycleGAN model. We also performed a series of experiments to assess the impact of using virtual DES images on the classification of CESM examinations into malignant and non-malignant categories. To our knowledge, this is the first study to evaluate the impact of virtual DES images on CESM lesion classification. The results demonstrate that the best performance was achieved with the pre-trained U-Net model, yielding an F1 score of 85.59% when using the virtual DES images, compared to 90.35% with the real DES images. This discrepancy likely results from the additional diagnostic information in real DES images, which contributes to a higher classification accuracy. Nevertheless, the potential for virtual DES image generation is considerable and future advancements may narrow this performance gap to a level where exclusive reliance on virtual DES images becomes clinically viable.

Method: Developed a data annotation protocol capturing radiologists’ per-pixel confidence values. Trained AI models using various confidence thresholds (tested 50-60%+), with binarized labels from confidence thresholds. Systematically evaluated impact on segmentation and downstream clinical tasks.

Result: Incorporating confidence values improved segmentation performance. 60% confidence threshold worked significantly better than naive 50% approach. Enhanced segmentation translated to better performance on clinically-critical tasks: S/F oxygenation ratio estimation, S/F ratio change classification, and 30-day patient readmission prediction.

Conclusion: Label confidence is a valuable signal that, when properly leveraged through appropriate thresholding (60%+), can significantly enhance AI reliability and clinical utility in medical imaging by capturing inherent aleatoric uncertainty in real-world clinical data.

Abstract: In medical imaging, inter-observer variability among radiologists often introduces label uncertainty, particularly in modalities where visual interpretation is subjective. Lung ultrasound (LUS) is a prime example-it frequently presents a mixture of highly ambiguous regions and clearly discernible structures, making consistent annotation challenging even for experienced clinicians. In this work, we introduce a novel approach to both labeling and training AI models using expert-supplied, per-pixel confidence values. Rather than treating annotations as absolute ground truth, we design a data annotation protocol that captures the confidence that radiologists have in each labeled region, modeling the inherent aleatoric uncertainty present in real-world clinical data. We demonstrate that incorporating these confidence values during training leads to improved segmentation performance. More importantly, we show that this enhanced segmentation quality translates into better performance on downstream clinically-critical tasks-specifically, estimating S/F oxygenation ratio values, classifying S/F ratio change, and predicting 30-day patient readmission. While we empirically evaluate many methods for exposing the uncertainty to the learning model, we find that a simple approach that trains a model on binarized labels obtained with a (60%) confidence threshold works well. Importantly, high thresholds work far better than a naive approach of a 50% threshold, indicating that training on very confident pixels is far more effective. Our study systematically investigates the impact of training with varying confidence thresholds, comparing not only segmentation metrics but also downstream clinical outcomes. These results suggest that label confidence is a valuable signal that, when properly leveraged, can significantly enhance the reliability and clinical utility of AI in medical imaging.

Method: Proposed HessNet - a Hessian-based lightweight neural network with only 6000 parameters for 3D segmentation of tubular structures. Uses semi-supervised learning and can run on CPU, significantly reducing resource requirements.

Result: Achieved state-of-the-art vessel segmentation accuracy on minimal training dataset. Enabled creation of a large semi-manually annotated brain vessel dataset (200 images from IXI dataset) with expert supervision. The dataset is publicly available.

Conclusion: HessNet provides an efficient, resource-light solution for accurate brain vessel segmentation, facilitates expert annotation workflow, and addresses the critical gap in publicly available annotated MRA datasets for medical research.

Abstract: Accurate segmentation of blood vessels in brain magnetic resonance angiography (MRA) is essential for successful surgical procedures, such as aneurysm repair or bypass surgery. Currently, annotation is primarily performed through manual segmentation or classical methods, such as the Frangi filter, which often lack sufficient accuracy. Neural networks have emerged as powerful tools for medical image segmentation, but their development depends on well-annotated training datasets. However, there is a notable lack of publicly available MRA datasets with detailed brain vessel annotations. To address this gap, we propose a novel semi-supervised learning lightweight neural network with Hessian matrices on board for 3D segmentation of complex structures such as tubular structures, which we named HessNet. The solution is a Hessian-based neural network with only 6000 parameters. HessNet can run on the CPU and significantly reduces the resource requirements for training neural networks. The accuracy of vessel segmentation on a minimal training dataset reaches state-of-the-art results. It helps us create a large, semi-manually annotated brain vessel dataset of brain MRA images based on the IXI dataset (annotated 200 images). Annotation was performed by three experts under the supervision of three neurovascular surgeons after applying HessNet. It provides high accuracy of vessel segmentation and allows experts to focus only on the most complex important cases. The dataset is available at https://git.scinalytics.com/terilat/VesselDatasetPartly.

Method: Comprehensive assessment of existing machine learning methods for dMRI analysis, focusing on data preprocessing/harmonization, microstructure mapping, tractography, and white matter tract analysis.

Result: Machine learning shows exceptional suitability for difficult dMRI tasks but requires addressing shortcomings in evaluation practices, data availability, model generalizability, reliability, and explainability.

Conclusion: While promising, machine learning for dMRI analysis needs improved evaluation standards, richer training datasets, validation benchmarks, and better model generalizability and explainability to reach full potential.

Abstract: \hspace{2mm} Diffusion-weighted magnetic resonance imaging (dMRI) of the brain offers unique capabilities including noninvasive probing of tissue microstructure and structural connectivity. It is widely used for clinical assessment of disease and injury, and for neuroscience research. Analyzing the dMRI data to extract useful information for medical and scientific purposes can be challenging. The dMRI measurements may suffer from strong noise and artifacts, and may exhibit high inter-session and inter-scanner variability in the data, as well as inter-subject heterogeneity in brain structure. Moreover, the relationship between measurements and the phenomena of interest can be highly complex. Recent years have witnessed increasing use of machine learning methods for dMRI analysis. This manuscript aims to assess these efforts, with a focus on methods that have addressed data preprocessing and harmonization, microstructure mapping, tractography, and white matter tract analysis. We study the main findings, strengths, and weaknesses of the existing methods and suggest topics for future research. We find that machine learning may be exceptionally suited to tackle some of the difficult tasks in dMRI analysis. However, for this to happen, several shortcomings of existing methods and critical unresolved issues need to be addressed. There is a pressing need to improve evaluation practices, to increase the availability of rich training datasets and validation benchmarks, as well as model generalizability, reliability, and explainability concerns.

Method: Proposes Fast-DDPM with only 10 time steps using efficient noise schedulers (uniform and non-uniform sampling) that align training and sampling procedures for optimal time-step utilization.

Result: Outperformed DDPM and state-of-the-art methods in multi-image super-resolution, denoising, and image translation tasks while reducing training time to 0.2x and sampling time to 0.01x.

Conclusion: Fast-DDPM provides an effective solution for efficient medical image generation with significantly reduced computational costs while maintaining or improving performance.

Abstract: Denoising diffusion probabilistic models (DDPMs) have achieved unprecedented success in computer vision. However, they remain underutilized in medical imaging, a field crucial for disease diagnosis and treatment planning. This is primarily due to the high computational cost associated with (1) the use of large number of time steps (e.g., 1,000) in diffusion processes and (2) the increased dimensionality of medical images, which are often 3D or 4D. Training a diffusion model on medical images typically takes days to weeks, while sampling each image volume takes minutes to hours. To address this challenge, we introduce Fast-DDPM, a simple yet effective approach capable of improving training speed, sampling speed, and generation quality simultaneously. Unlike DDPM, which trains the image denoiser across 1,000 time steps, Fast-DDPM trains and samples using only 10 time steps. The key to our method lies in aligning the training and sampling procedures to optimize time-step utilization. Specifically, we introduced two efficient noise schedulers with 10 time steps: one with uniform time step sampling and another with non-uniform sampling. We evaluated Fast-DDPM across three medical image-to-image generation tasks: multi-image super-resolution, image denoising, and image-to-image translation. Fast-DDPM outperformed DDPM and current state-of-the-art methods based on convolutional networks and generative adversarial networks in all tasks. Additionally, Fast-DDPM reduced the training time to 0.2x and the sampling time to 0.01x compared to DDPM. Our code is publicly available at: https://github.com/mirthAI/Fast-DDPM.

Method: Two-phase approach: 1) Identify candidate locations similar to rare-type nuclei surroundings, 2) Use progressive inpainting with pre-trained diffusion model to integrate rare-type nuclei by replacing major-type nuclei or background locations.

Result: Superior ability to synthesize realistic rare-type nuclei and enhance nuclei segmentation and classification quality across three public datasets using two popular segmentation models.

Conclusion: NucleiMix effectively addresses data imbalance in nuclei instance segmentation through targeted data augmentation, demonstrating robust performance improvement in pathology image analysis tasks.

Abstract: Nuclei instance segmentation is an essential task in pathology image analysis, serving as the foundation for many downstream applications. The release of several public datasets has significantly advanced research in this area, yet many existing methods struggle with data imbalance issues. To address this challenge, this study introduces a data augmentation method, called NucleiMix, which is designed to balance the distribution of nuclei types by increasing the number of rare-type nuclei within datasets. NucleiMix operates in two phases. In the first phase, it identifies candidate locations similar to the surroundings of rare-type nuclei and inserts rare-type nuclei into the candidate locations. In the second phase, it employs a progressive inpainting strategy using a pre-trained diffusion model to seamlessly integrate rare-type nuclei into their new environments in replacement of major-type nuclei or background locations. We systematically evaluate the effectiveness of NucleiMix on three public datasets using two popular nuclei instance segmentation models. The results demonstrate the superior ability of NucleiMix to synthesize realistic rare-type nuclei and to enhance the quality of nuclei segmentation and classification in an accurate and robust manner.

Method: Proposes MambaRoll - an unrolled architecture with scale-specific physics-driven SSM modules that autoregressively predict finer-scale features from coarser representations. Includes Deep Multi-Scale Decoding loss for intermediate scale supervision.

Result: MambaRoll consistently outperforms state-of-the-art CNN-, transformer-, and SSM-based methods in accelerated MRI and sparse-view CT reconstructions.

Conclusion: The autoregressive multi-scale approach with physics-driven SSM modules and scale-aware supervision enables high-fidelity and efficient medical image reconstruction, addressing limitations of existing methods.

Abstract: Medical image reconstruction from undersampled acquisitions is an ill-posed inverse problem requiring accurate recovery of anatomical structures from incomplete measurements. Physics-driven (PD) network models have gained prominence for this task by integrating data-consistency mechanisms with learned priors, enabling improved performance over purely data-driven approaches. However, reconstruction quality still hinges on the network’s ability to disentangle artifacts from true anatomical signals-both of which exhibit complex, multi-scale contextual structure. Convolutional neural networks (CNNs) capture local correlations but often struggle with non-local dependencies. While transformers aim to alleviate this limitation, practical implementations involve design compromises to reduce computational cost by balancing local and non-local sensitivity, occasionally resulting in performance comparable to CNNs. To address these challenges, we propose MambaRoll, a novel physics-driven autoregressive state space model (SSM) for high-fidelity and efficient image reconstruction. MambaRoll employs an unrolled architecture where each cascade autoregressively predicts finer-scale feature maps conditioned on coarser-scale representations, enabling consistent multi-scale context propagation. Each stage is built on a hierarchy of scale-specific PD-SSM modules that capture spatial dependencies while enforcing data consistency through residual correction. To further improve scale-aware learning, we introduce a Deep Multi-Scale Decoding (DMSD) loss, which provides supervision at intermediate spatial scales in alignment with the autoregressive design. Demonstrations on accelerated MRI and sparse-view CT reconstructions show that MambaRoll consistently outperforms state-of-the-art CNN-, transformer-, and SSM-based methods.

Method: Proposes LD-SMC method using sequential Monte Carlo sampling in diffusion model latent space, defining generative model with auxiliary observations and performing posterior inference via reverse diffusion process.

Result: Empirical evaluations on ImageNet and FFHQ demonstrate superior performance over competing methods in various inverse problem tasks, particularly in challenging inpainting scenarios.

Conclusion: LD-SMC effectively addresses sampling challenges in diffusion models for inverse problems, providing improved reconstruction quality through sequential Monte Carlo techniques in latent space.

Abstract: In image processing, solving inverse problems is the task of finding plausible reconstructions of an image that was corrupted by some (usually known) degradation operator. Commonly, this process is done using a generative image model that can guide the reconstruction towards solutions that appear natural. The success of diffusion models over the last few years has made them a leading candidate for this task. However, the sequential nature of diffusion models makes this conditional sampling process challenging. Furthermore, since diffusion models are often defined in the latent space of an autoencoder, the encoder-decoder transformations introduce additional difficulties. To address these challenges, we suggest a novel sampling method based on sequential Monte Carlo (SMC) in the latent space of diffusion models. We name our method LD-SMC. We define a generative model for the data using additional auxiliary observations and perform posterior inference with SMC sampling based on a reverse diffusion process. Empirical evaluations on ImageNet and FFHQ show the benefits of LD-SMC over competing methods in various inverse problem tasks and especially in challenging inpainting tasks.

Method: 3D Gaussian MRI (3DGSMR) framework using 3D Gaussian distributions as explicit representation for MR volumes, operating in self-supervised manner without training datasets.

Result: Effectively reconstructs voxelized MR images with quality on par with established 3D MRI reconstruction techniques.

Conclusion: Successful adaptation of 3DGS to MRI reconstruction with innovative complex-valued signal decomposition, offering training-free approach.

Abstract: Three-Dimensional Gaussian Splatting (3DGS) has shown substantial promise in the field of computer vision, but remains unexplored in the field of magnetic resonance imaging (MRI). This study explores its potential for the reconstruction of isotropic resolution 3D MRI from undersampled k-space data. We introduce a novel framework termed 3D Gaussian MRI (3DGSMR), which employs 3D Gaussian distributions as an explicit representation for MR volumes. Experimental evaluations indicate that this method can effectively reconstruct voxelized MR images, achieving a quality on par with that of well-established 3D MRI reconstruction techniques found in the literature. Notably, the 3DGSMR scheme operates under a self-supervised framework, obviating the need for extensive training datasets or prior model training. This approach introduces significant innovations to the domain, notably the adaptation of 3DGS to MRI reconstruction and the novel application of the existing 3DGS methodology to decompose MR signals, which are presented in a complex-valued format.

Method: Retrospective study of 162 patients (129 AIS, 33 seizures) using MR perfusion imaging. Extracted region-wise perfusion map descriptors from DSC images and performed statistical analyses to identify discriminative brain regions.

Result: Logistic regression model achieved AUROC of 0.90, AUPRC of 0.74, 92% specificity, and 73% sensitivity. Temporal and occipital lobe regions showed significant discriminative power.

Conclusion: MRP-based perfusion map descriptors show strong potential as interpretable features for distinguishing true strokes from mimics, supporting further exploration of this approach.

Abstract: Distinguishing acute ischemic strokes (AIS) from stroke mimics (SMs), particularly in cases involving medium and small vessel occlusions, remains a significant diagnostic challenge. While computed tomography (CT) based protocols are commonly used in emergency settings, their sensitivity for detecting distal occlusions is limited. This study explores the potential of magnetic resonance perfusion (MRP) imaging as a tool for differentiating distal AIS from epileptic seizures, a prevalent SM. Using a retrospective dataset of 162 patients (129 AIS, 33 seizures), we extracted region-wise perfusion map descriptors (PMDs) from dynamic susceptibility contrast (DSC) images. Statistical analyses identified several brain regions, located mainly in the temporal and occipital lobe, exhibiting significant group differences in certain PMDs. Hemispheric asymmetry analyses further highlighted these regions as discriminative. A logistic regression model trained on PMDs achieved an area under the receiver operating characteristic (AUROC) curve of 0.90, and an area under the precision recall curve (AUPRC) of 0.74, with a specificity of 92% and a sensitivity of 73%, suggesting strong performance in distinguishing distal AIS from seizures. These findings support further exploration of MRP-based PMDs as interpretable features for distinguishing true strokes from various mimics. The code is openly available at our GitHub https://github.com/Marijn311/PMD_extraction_and_analysis{github.com/Marijn311/PMD_extraction_and_analysis

Method: Used 3D pseudo-continuous arterial spin labeling MRI on 186 cognitively healthy participants, applied extended 3D SLIC supervoxel algorithm for regional analysis, trained CNN for sex classification, and developed vascular risk scores by comparing individual CBF profiles to normative data.

Result: Achieved 95% accuracy in sex classification, identified higher CBF in females in specific brain regions (medial BA6/10, V5, occipital polar cortex, insula), observed global CBF decline with age in both sexes, and successfully quantified individual perfusion deficits using personalized VRS.

Conclusion: Sex and age specific CBF patterns were successfully identified, and a personalized vascular risk biomarker was developed, contributing to precision neurology advancements for early hypoperfusion detection.

Abstract: The influence of sex and age on cerebral perfusion is recognized, but the specific impacts on regional cerebral blood flow (CBF) and vascular risk remain to be fully characterized. In this study, 3D pseudo-continuous arterial spin labeling (pCASL) MRI was used to identify sex and age related CBF patterns, and a vascular risk score (VRS) was developed based on normative perfusion profiles. Perfusion data from 186 cognitively healthy participants (89 males, 97 females; aged 8 to 92 years), obtained from a publicly available dataset, were analyzed. An extension of the 3D Simple Linear Iterative Clustering (SLIC) supervoxel algorithm was applied to CBF maps to group neighboring voxels with similar intensities into anatomically meaningful regions. Regional CBF features were extracted and used to train a convolutional neural network (CNN) for sex classification and perfusion pattern analysis. Global, age related CBF changes were also assessed. Participant specific VRS was computed by comparing individual CBF profiles to age and sex specific normative data to quantify perfusion deficits. A 95 percent accuracy in sex classification was achieved using the proposed supervoxel based method, and distinct perfusion signatures were identified. Higher CBF was observed in females in medial Brodmann areas 6 and 10, area V5, occipital polar cortex, and insular regions. A global decline in CBF with age was observed in both sexes. Individual perfusion deficits were quantified using VRS, providing a personalized biomarker for early hypoperfusion. Sex and age specific CBF patterns were identified, and a personalized vascular risk biomarker was proposed, contributing to advancements in precision neurology.

Method: Proposes Cardiac Latent Interpolation Diffusion (CaLID) framework with three innovations: 1) data-driven interpolation using diffusion models, 2) latent space operation for 24x faster computation, 3) works with only sparse 2D CMR images without auxiliary inputs. Also extends to 2D+T for spatiotemporal modeling.

Result: Achieves state-of-the-art performance in reconstruction quality and efficiency. Demonstrates superior results in volumetric evaluations and downstream segmentation tasks. Reduces computational overhead significantly compared to previous methods.

Conclusion: CaLID advances spatio and spatiotemporal whole-heart reconstruction, providing a robust and clinically practical solution for cardiovascular imaging by addressing fundamental limitations of existing approaches.

Abstract: Cardiac Magnetic Resonance (CMR) imaging is a critical tool for diagnosing and managing cardiovascular disease, yet its utility is often limited by the sparse acquisition of 2D short-axis slices, resulting in incomplete volumetric information. Accurate 3D reconstruction from these sparse slices is essential for comprehensive cardiac assessment, but existing methods face challenges, including reliance on predefined interpolation schemes (e.g., linear or spherical), computational inefficiency, and dependence on additional semantic inputs such as segmentation labels or motion data. To address these limitations, we propose a novel Cardiac Latent Interpolation Diffusion (CaLID) framework that introduces three key innovations. First, we present a data-driven interpolation scheme based on diffusion models, which can capture complex, non-linear relationships between sparse slices and improves reconstruction accuracy. Second, we design a computationally efficient method that operates in the latent space and speeds up 3D whole-heart upsampling time by a factor of 24, reducing computational overhead compared to previous methods. Third, with only sparse 2D CMR images as input, our method achieves SOTA performance against baseline methods, eliminating the need for auxiliary input such as morphological guidance, thus simplifying workflows. We further extend our method to 2D+T data, enabling the effective modeling of spatiotemporal dynamics and ensuring temporal coherence. Extensive volumetric evaluations and downstream segmentation tasks demonstrate that CaLID achieves superior reconstruction quality and efficiency. By addressing the fundamental limitations of existing approaches, our framework advances the state of the art for spatio and spatiotemporal whole-heart reconstruction, offering a robust and clinically practical solution for cardiovascular imaging.

Method: Evaluated ResNet50, InceptionV3, Vision Transformers, U-Net variants with MLP classifiers, integrated Grad-CAM and Saliency Maps for explainability.

Result: ResNet50 excelled in classification and ROI identification, outperforming transformers. Grad-CAM provided most clinically meaningful explanations.

Conclusion: CNN-based transfer learning most effective; future work with larger pretraining needed to unlock transformer potential.

Abstract: Magnetic Resonance Imaging (MRI) is an essential diagnostic tool for assessing knee injuries. However, manual interpretation of MRI slices remains time-consuming and prone to inter-observer variability. This study presents a systematic evaluation of various deep learning architectures combined with explainable AI (xAI) techniques for automated region of interest (ROI) detection in knee MRI scans. We investigate both supervised and self-supervised approaches, including ResNet50, InceptionV3, Vision Transformers (ViT), and multiple U-Net variants augmented with multi-layer perceptron (MLP) classifiers. To enhance interpretability and clinical relevance, we integrate xAI methods such as Grad-CAM and Saliency Maps. Model performance is assessed using AUC for classification and PSNR/SSIM for reconstruction quality, along with qualitative ROI visualizations. Our results demonstrate that ResNet50 consistently excels in classification and ROI identification, outperforming transformer-based models under the constraints of the MRNet dataset. While hybrid U-Net + MLP approaches show potential for leveraging spatial features in reconstruction and interpretability, their classification performance remains lower. Grad-CAM consistently provided the most clinically meaningful explanations across architectures. Overall, CNN-based transfer learning emerges as the most effective approach for this dataset, while future work with larger-scale pretraining may better unlock the potential of transformer models.