Method: Extends DeBERTa-v3 with Focal Loss, Label Smoothing, and R-Drop regularization; uses 20,000 annotated samples across four KC categories; implements reproducible end-to-end pipeline with 10-fold cross-validation.

Result: Achieved macro-F1 of 0.836 ± 0.008, significantly outperforming classical and transformer baselines (p<0.01); strong sensitivity to higher-order epistemic engagement in Explore and Negotiate categories.

Conclusion: Large language models can effectively capture nuanced indicators of knowledge construction, offering scalable approaches to discourse analysis and automated assessment of epistemic engagement.

Abstract: This study presents DeBERTa-KC, a transformer-based model for automatic classification of knowledge construction (KC) levels in online science learning discourse. Using comments collected from four popular YouTube science channels (2022–2024), a balanced corpus of 20,000 manually annotated samples was created across four KC categories: \textit{nonKC}, \textit{Share}, \textit{Explore}, and \textit{Negotiate}. The proposed model extends DeBERTa-v3 with Focal Loss, Label Smoothing, and R-Drop regularization to address class imbalance and enhance generalization. A reproducible end-to-end pipeline was implemented, encompassing data extraction, annotation, preprocessing, training, and evaluation. Across 10-fold stratified cross-validation, DeBERTa-KC achieved a macro-F1 of $0.836 \pm 0.008$, significantly out-performing both classical and transformer baselines ($p<0.01$). Per-category results indicate strong sensitivity to higher-order epistemic engagement, particularly in \textit{Explore} and \textit{Negotiate} discourse. These findings demonstrate that large language models can effectively capture nuanced indicators of knowledge construction in informal digital learning environments, offering scalable, theory-informed approaches to discourse analysis and the development of automated tools for assessing epistemic engagement.

Method: Generated 15 lesson plans for high-school physics using 5 LLMs (ChatGPT, Claude, Gemini, DeepSeek, Grok) with 3 prompt frameworks (TAG, RACE, COSTAR), analyzed through 4 computational metrics: readability, factual accuracy, curriculum alignment, and cognitive demand.

Result: DeepSeek produced most readable plans (FKGL=8.64), Claude generated densest language (FKGL=19.89). RACE framework yielded lowest hallucination and highest curriculum alignment. Learning objectives clustered at lower Bloom’s taxonomy levels with limited higher-order thinking.

Conclusion: Readability is governed by model design, while instructional reliability depends more on prompt framework. Optimal configuration combines readability-optimized model with RACE framework and explicit content/standards checklists.

Abstract: This study evaluates the pedagogical soundness and usability of AI-generated lesson plans across five leading large language models: ChatGPT (GPT-5), Claude Sonnet 4.5, Gemini 2.5 Flash, DeepSeek V3.2, and Grok 4. Beyond model choice, three structured prompt frameworks were tested: TAG (Task, Audience, Goal), RACE (Role, Audience, Context, Execution), and COSTAR (Context, Objective, Style, Tone, Audience, Response Format). Fifteen lesson plans were generated for a single high-school physics topic, The Electromagnetic Spectrum. The lesson plans were analyzed through four automated computational metrics: (1) readability and linguistic complexity, (2) factual accuracy and hallucination detection, (3) standards and curriculum alignment, and (4) cognitive demand of learning objectives. Results indicate that model selection exerted the strongest influence on linguistic accessibility, with DeepSeek producing the most readable teaching plan (FKGL = 8.64) and Claude generating the densest language (FKGL = 19.89). The prompt framework structure most strongly affected the factual accuracy and pedagogical completeness, with the RACE framework yielding the lowest hallucination index and the highest incidental alignment with NGSS curriculum standards. Across all models, the learning objectives in the fifteen lesson plans clustered at the Remember and Understand tiers of Bloom’s taxonomy. There were limited higher-order verbs in the learning objectives extracted. Overall, the findings suggest that readability is significantly governed by model design, while instructional reliability and curricular alignment depend more on the prompt framework. The most effective configuration for lesson plans identified in the results was to combine a readability-optimized model with the RACE framework and an explicit checklist of physics concepts, curriculum standards, and higher-order objectives.

Method: Two-stage training: 1) Foundational revision capability by training to revise synthetic errors; 2) Online self-correction loop where model learns to revise its own flawed drafts from expert corrections, enabling mistake-driven learning.

Result: Significantly improves coherence and factual accuracy of generated content, enabling stable and efficient parallel generation superior to traditional denoising methods.

Conclusion: ReDiff effectively breaks error cascades by teaching models to identify and correct their own errors, transforming the generation paradigm from passive denoising to active refining.

Abstract: Discrete diffusion models have emerged as a promising direction for vision-language tasks, offering bidirectional context modeling and theoretical parallelization. However, their practical application is severely hindered by a train-inference discrepancy, which leads to catastrophic error cascades: initial token errors during parallel decoding pollute the generation context, triggering a chain reaction of compounding errors and leading to syntactic errors and semantic hallucinations. To address this fundamental challenge, we reframe the generation process from passive denoising to active refining. We introduce ReDiff, a refining-enhanced diffusion framework that teaches the model to identify and correct its own errors. Our approach features a two-stage training process: first, we instill a foundational revision capability by training the model to revise synthetic errors; second, we implement a novel online self-correction loop where the model is explicitly trained to revise its own flawed drafts by learning from an expert’s corrections. This mistake-driven learning endows the model with the crucial ability to revisit and refine its already generated output, effectively breaking the error cascade. Extensive experiments demonstrate that ReDiff significantly improves the coherence and factual accuracy of generated content, enabling stable and efficient parallel generation far superior to traditional denoising methods. Our codes and models are available at https://rediff-hku.github.io/.

Method: Stream algorithm performs hierarchical pruning to estimate per-head sparse attention masks in O(T log T) time and O(T) space, using binary-search-style refinement to retain top-k key blocks per query while preserving model behavior.

Result: Applied to long chain-of-thought reasoning, Sparse Tracing prunes 97-99% of token interactions while identifying thought anchors. On RULER benchmark, it preserves critical retrieval paths while discarding 90-96% of interactions and exposes layer-wise routes.

Conclusion: Sparse Tracing provides a practical drop-in tool for analyzing attention patterns and tracing information flow without requiring terabytes of caches, democratizing long context interpretability on consumer GPUs.

Abstract: As Large Language Models (LLMs) scale to million-token contexts, traditional Mechanistic Interpretability techniques for analyzing attention scale quadratically with context length, demanding terabytes of memory beyond 100,000 tokens. We introduce Sparse Tracing, a novel technique that leverages dynamic sparse attention to efficiently analyze long context attention patterns. We present Stream, a compilable hierarchical pruning algorithm that estimates per-head sparse attention masks in near-linear time $O(T \log T)$ and linear space $O(T)$, enabling one-pass interpretability at scale. Stream performs a binary-search-style refinement to retain only the top-$k$ key blocks per query while preserving the model’s next-token behavior. We apply Stream to long chain-of-thought reasoning traces and identify thought anchors while pruning 97-99% of token interactions. On the RULER benchmark, Stream preserves critical retrieval paths while discarding 90-96% of interactions and exposes layer-wise routes from the needle to output. Our method offers a practical drop-in tool for analyzing attention patterns and tracing information flow without terabytes of caches. By making long context interpretability feasible on consumer GPUs, Sparse Tracing helps democratize chain-of-thought monitoring. Code is available at https://anonymous.4open.science/r/stream-03B8/.

Method: Proposed a pipeline leveraging Large Language Models to analyze unstructured EHR text data to determine patient eligibility for HIV testing.

Result: Experimental results on clinical data from Erasmus University Medical Center Rotterdam showed high accuracy and low false negative rate.

Conclusion: LLM-based analysis of unstructured EHR text can effectively screen for HIV testing eligibility, improving early diagnosis while maintaining low false negatives.

Abstract: Efficient screening and early diagnosis of HIV are critical for reducing onward transmission. Although large scale laboratory testing is not feasible, the widespread adoption of Electronic Health Records (EHRs) offers new opportunities to address this challenge. Existing research primarily focuses on applying machine learning methods to structured data, such as patient demographics, for improving HIV diagnosis. However, these approaches often overlook unstructured text data such as clinical notes, which potentially contain valuable information relevant to HIV risk. In this study, we propose a novel pipeline that leverages a Large Language Model (LLM) to analyze unstructured EHR text and determine a patient’s eligibility for further HIV testing. Experimental results on clinical data from Erasmus University Medical Center Rotterdam demonstrate that our pipeline achieved high accuracy while maintaining a low false negative rate.

Method: Evaluated 11 general-purpose LLMs across 22 LCA tasks, with 17 experienced practitioners reviewing model outputs against criteria including scientific accuracy, explanation quality, robustness, verifiability, and adherence to instructions.

Result: 37% of responses contained inaccurate/misleading information; hallucination rates varied significantly (up to 40%); no clear distinction between open-weight vs closed-weight models; open-weight models performed competitively on accuracy and explanation quality.

Conclusion: Highlights risks of naive LLM application in LCA when treated as free-form oracles, but shows benefits for explanation quality and reducing labor intensiveness of simple tasks; warns against using general-purpose LLMs without grounding mechanisms.

Abstract: Purpose: Artificial intelligence (AI), and in particular large language models (LLMs), are increasingly being explored as tools to support life cycle assessment (LCA). While demonstrations exist across environmental and social domains, systematic evidence on their reliability, robustness, and usability remains limited. This study provides the first expert-grounded benchmark of LLMs in LCA, addressing the absence of standardized evaluation frameworks in a field where no clear ground truth or consensus protocols exist. Methods: We evaluated eleven general-purpose LLMs, spanning both commercial and open-source families, across 22 LCA-related tasks. Seventeen experienced practitioners reviewed model outputs against criteria directly relevant to LCA practice, including scientific accuracy, explanation quality, robustness, verifiability, and adherence to instructions. We collected 168 expert reviews. Results: Experts judged 37% of responses to contain inaccurate or misleading information. Ratings of accuracy and quality of explanation were generally rated average or good on many models even smaller models, and format adherence was generally rated favourably. Hallucination rates varied significantly, with some models producing hallucinated citations at rates of up to 40%. There was no clear-cut distinction between ratings on open-weight versus closed-weight LLMs, with open-weight models outperforming or competing on par with closed-weight models on criteria such as accuracy and quality of explanation. Conclusion: These findings highlight the risks of applying LLMs na"ively in LCA, such as when LLMs are treated as free-form oracles, while also showing benefits especially around quality of explanation and alleviating labour intensiveness of simple tasks. The use of general-purpose LLMs without grounding mechanisms presents …

Method: The authors use Dixit, a fantasy card game where players generate captions for cards that must trick some but not all players into selecting the played card. This requires multiple abilities and is governed by fixed, objective rules.

Result: Experiments with five MLMs show that Dixit win-rate rankings perfectly correlate with those on popular MLM benchmarks. Games between humans and MLMs reveal differences in agent strategies and areas for improvement in MLM reasoning.

Conclusion: Game-based evaluations like Dixit provide a robust, holistic framework for assessing MLM capabilities that addresses the limitations of current evaluation methods while being more engaging and objective.

Abstract: Multi-modal large language models (MLMs) are often assessed on static, individual benchmarks – which cannot jointly assess MLM capabilities in a single task – or rely on human or model pairwise comparisons – which is highly subjective, expensive, and allows models to exploit superficial shortcuts (e.g., verbosity) to inflate their win-rates. To overcome these issues, we propose game-based evaluations to holistically assess MLM capabilities. Games require multiple abilities for players to win, are inherently competitive, and are governed by fix, objective rules, and makes evaluation more engaging, providing a robust framework to address the aforementioned challenges. We manifest this evaluation specifically through Dixit, a fantasy card game where players must generate captions for a card that trick some, but not all players, into selecting the played card. Our quantitative experiments with five MLMs show Dixit win-rate rankings are perfectly correlated with those on popular MLM benchmarks, while games between human and MLM players in Dixit reveal several differences between agent strategies and areas of improvement for MLM reasoning.

Method: Systematic evaluation of DeepSeek-R1 across four OR benchmarks (NL4OPT, IndustryOR, EasyLP, ComplexOR) using baseline assessments, hallucination taxonomy development, and mitigation strategies including LLM-as-a-Judge, Few-shot Learning, Tool Calling, and Multi-agent Framework.

Result: DeepSeek-R1 shows promise as a cost-efficient alternative to high-cost models like GPT-4 and Claude, with demonstrated success in benchmarks but requires further evaluation in applied OR scenarios.

Conclusion: LLMs like DeepSeek-R1 can potentially bridge the formulation gap in optimization modeling through natural language understanding, though hallucination mitigation strategies are essential for real-world applicability in operations research.

Abstract: The integration of Large Language Models (LLMs) with optimization modeling offers a promising avenue for advancing decision-making in operations research (OR). Traditional optimization methods,such as linear programming, mixed integer programming, and simulation depend heavily on domain expertise to translate real-world problems into solvable mathematical models. While solvers like Gurobi and COPT are powerful, expert input remains essential for defining objectives, constraints, and variables. This research investigates the potential of LLMs, specifically the DeepSeek-R1 model, to bridge this formulation gap using natural language understanding and code generation. Although prior models like GPT-4, Claude, and Bard have shown strong performance in NLP and reasoning tasks, their high token costs and tendency toward hallucinations limit real-world applicability in supply chain contexts. In contrast, DeepSeek-R1, a cost-efficient and high-performing model trained with reinforcement learning, presents a viable alternative. Despite its success in benchmarks such as LiveCodeBench and Math-500, its effectiveness in applied OR scenarios remains under explored. This study systematically evaluates DeepSeek-R1 across four key OR benchmarks: NL4OPT, IndustryOR, EasyLP, and ComplexOR. Our methodology includes baseline assessments, the development of a hallucination taxonomy, and the application of mitigation strategies like LLM-as-a-Judge, Few-shot Learning (FSL), Tool Calling, and a Multi-agent Framework. These techniques aim to reduce hallucinations, enhance formulation accuracy, and better align model outputs with user intent.

Method: Proposes a memory-augmented framework with episodic memory for instance-level critiques and semantic memory for task-level guidance, using LLM-generated critiques alongside labeled data.

Result: Achieves up to 24.8% accuracy improvement over retrieval-based baselines, reveals behavioral differences between model types, and introduces suggestibility metric to interpret learning dynamics.

Conclusion: Memory-driven reflective learning shows promise for building more adaptive and interpretable LLM agents by leveraging critiques and memory strategies.

Abstract: We investigate how agents built on pretrained large language models can learn target classification functions from labeled examples without parameter updates. While conventional approaches like fine-tuning are often costly, inflexible, and opaque, we propose a memory-augmented framework that leverages both labeled data and LLM-generated critiques. Our framework uses episodic memory to store instance-level critiques-capturing specific past experiences-and semantic memory to distill these into reusable, task-level guidance. Across a diverse set of tasks, incorporating critiques yields up to a 24.8 percent accuracy improvement over retrieval-based (RAG-style) baselines that rely only on labels. Through extensive empirical evaluation, we uncover distinct behavioral differences between OpenAI and opensource models, particularly in how they handle fact-oriented versus preference-based data. To interpret how models respond to different representations of supervision encoded in memory, we introduce a novel metric, suggestibility. This helps explain observed behaviors and illuminates how model characteristics and memory strategies jointly shape learning dynamics. Our findings highlight the promise of memory-driven, reflective learning for building more adaptive and interpretable LLM agents.

Method: Two-step pipeline: 1) Universal transcription generator to unify orthographic features into Romanized form capturing common phonetic characteristics, 2) Universal converter to transform universal transcriptions into language-specific ones.

Result: Achieves 45% relative error reduction compared to Whisper, performs comparably to MMS despite using only 0.1% of Whisper’s training data, and matches zero-shot ASR approaches without language-specific modules.

Conclusion: LAMA-UT serves as a flexible multilingual ASR framework generalizable to unseen languages, demonstrating effectiveness through universal transcriptions for massively multilingual ASR.

Abstract: Building a universal multilingual automatic speech recognition (ASR) model that performs equitably across languages has long been a challenge due to its inherent difficulties. To address this task we introduce a Language-Agnostic Multilingual ASR pipeline through orthography Unification and language-specific Transliteration (LAMA-UT). LAMA-UT operates without any language-specific modules while matching the performance of state-of-the-art models trained on a minimal amount of data. Our pipeline consists of two key steps. First, we utilize a universal transcription generator to unify orthographic features into Romanized form and capture common phonetic characteristics across diverse languages. Second, we utilize a universal converter to transform these universal transcriptions into language-specific ones. In experiments, we demonstrate the effectiveness of our proposed method leveraging universal transcriptions for massively multilingual ASR. Our pipeline achieves a relative error reduction rate of 45% when compared to Whisper and performs comparably to MMS, despite being trained on only 0.1% of Whisper’s training data. Furthermore, our pipeline does not rely on any language-specific modules. However, it performs on par with zero-shot ASR approaches which utilize additional language-specific lexicons and language models. We expect this framework to serve as a cornerstone for flexible multilingual ASR systems that are generalizable even to unseen languages.

Method: Proposes LyriCAR framework with difficulty-aware curriculum designer and adaptive curriculum strategy that guides the model through increasingly complex challenges for efficient training.

Result: Achieves state-of-the-art results on EN-ZH lyric translation across both standard metrics and multi-dimensional reward scores, with adaptive curriculum reducing training steps by nearly 40%.

Conclusion: LyriCAR demonstrates effective unsupervised lyric translation through curriculum learning, significantly improving performance while reducing computational costs.

Abstract: Lyric translation is a challenging task that requires balancing multiple musical constraints. Existing methods often rely on hand-crafted rules and sentence-level modeling, which restrict their ability to internalize musical-linguistic patterns and to generalize effectively at the paragraph level, where cross-line coherence and global rhyme are crucial. In this work, we propose LyriCAR, a novel framework for controllable lyric translation that operates in a fully unsupervised manner. LyriCAR introduces a difficulty-aware curriculum designer and an adaptive curriculum strategy, ensuring efficient allocation of training resources, accelerating convergence, and improving overall translation quality by guiding the model with increasingly complex challenges. Extensive experiments on the EN-ZH lyric translation task show that LyriCAR achieves state-of-the-art results across both standard translation metrics and multi-dimensional reward scores, surpassing strong baselines. Notably, the adaptive curriculum strategy reduces training steps by nearly 40% while maintaining superior performance. Code, data and model can be accessed at https://github.com/rle27/LyriCAR.

Method: Created MCIF benchmark based on scientific talks spanning three modalities (speech, vision, text) and four languages, with human annotations to evaluate instruction-following across languages and multimodal contexts.

Result: Developed a comprehensive evaluation framework that enables assessment of MLLMs’ abilities to interpret instructions across languages and combine multimodal contextual information in both short- and long-form inputs.

Conclusion: MCIF fills critical gaps in MLLM evaluation and is released under CC-BY 4.0 license to foster open research and progress in multimodal language model development.

Abstract: Recent advances in large language models have catalyzed the development of multimodal LLMs (MLLMs) that integrate text, speech, and vision within unified frameworks. As MLLMs evolve from narrow, monolingual, task-specific systems to general-purpose instruction-following models, a key frontier lies in evaluating their multilingual and multimodal capabilities over both long and short contexts. However, existing benchmarks fall short in evaluating these dimensions jointly: they are often limited to English, mostly focus on one single modality at a time, rely on short-form contexts, or lack human annotations – hindering comprehensive assessment of model performance across languages, modalities, and task complexity. To address these gaps, we introduce MCIF (Multimodal Crosslingual Instruction Following), the first multilingual human-annotated benchmark based on scientific talks that is designed to evaluate instruction-following in crosslingual, multimodal settings over both short- and long-form inputs. MCIF spans three core modalities – speech, vision, and text – and four diverse languages (English, German, Italian, and Chinese), enabling a comprehensive evaluation of MLLMs’ abilities to interpret instructions across languages and combine them with multimodal contextual information. MCIF is released under a CC-BY 4.0 license to encourage open research and progress in MLLMs development.

Method: Integrate LLMs for rephrasing, text simplification, and filling knowledge gaps with symbolic NLU for producing structured relational representations usable for reasoning and incremental learning.

Result: The hybrid method performs significantly better than symbolic-only pipelines on extracting quantities and causal laws from commonsense science texts.

Conclusion: Combining LLMs’ broad language processing with symbolic NLU’s structured representations offers a promising approach that leverages the strengths of both systems.

Abstract: Despite the broad applicability of large language models (LLMs), their reliance on probabilistic inference makes them vulnerable to errors such as hallucination in generated facts and inconsistent output structure in natural language understanding (NLU) tasks. By contrast, symbolic NLU systems provide interpretable understanding grounded in curated lexicons, semantic resources, and syntactic & semantic interpretation rules. They produce relational representations that can be used for accurate reasoning and planning, as well as incremental debuggable learning. However, symbolic NLU systems tend to be more limited in coverage than LLMs and require scarce knowledge representation and linguistics skills to extend and maintain. This paper explores a hybrid approach that integrates the broad-coverage language processing of LLMs with the symbolic NLU capabilities of producing structured relational representations to hopefully get the best of both approaches. We use LLMs for rephrasing and text simplification, to provide broad coverage, and as a source of information to fill in knowledge gaps more automatically. We use symbolic NLU to produce representations that can be used for reasoning and for incremental learning. We evaluate this approach on the task of extracting and interpreting quantities and causal laws from commonsense science texts, along with symbolic- and LLM-only pipelines. Our results suggest that our hybrid method works significantly better than the symbolic-only pipeline.

Method: Leverages Mamba architecture - an efficient state-space model optimized for long-context sequence processing - for multilingual ASR, incorporating language-aware conditioning and shared representations.

Result: Experiments on standard multilingual benchmarks show MLMA achieves competitive performance compared to Transformer-based architectures.

Conclusion: Mamba has strong potential as a backbone for scalable, efficient, and accurate multilingual speech recognition.

Abstract: Multilingual automatic speech recognition (ASR) remains a challenging task, especially when balancing performance across high- and low-resource languages. Recent advances in sequence modeling suggest that architectures beyond Transformers may offer better scalability and efficiency. In this work, we introduce MLMA (Multilingual Language Modeling with Mamba for ASR), a new approach that leverages the Mamba architecture – an efficient state-space model optimized for long-context sequence processing – for multilingual ASR. Using Mamba, MLMA implicitly incorporates language-aware conditioning and shared representations to support robust recognition across diverse languages. Experiments on standard multilingual benchmarks show that MLMA achieves competitive performance compared to Transformer-based architectures. These results highlight Mamba’s potential as a strong backbone for scalable, efficient, and accurate multilingual speech recognition.

Method: The algorithm processes sentences one word at a time, attaching each word as soon as it can be attached, creating dependency trees incrementally.

Result: The algorithm achieves O(n^3) worst-case complexity like phrase-structure parsers, but in human language this worst case only occurs for small n values.

Conclusion: This dependency parsing approach provides an efficient method that aligns with human cognitive processing while maintaining reasonable computational complexity for natural language.

Abstract: This paper presents a fundamental algorithm for parsing natural language sentences into dependency trees. Unlike phrase-structure (constituency) parsers, this algorithm operates one word at a time, attaching each word as soon as it can be attached, corresponding to properties claimed for the parser in the human brain. Like phrase-structure parsing, its worst-case complexity is $O(n^3)$, but in human language, the worst case occurs only for small $n$.

Method: Proposes a two-dimensional framework (Clinical Backgrounds and Clinical Questions) to characterize clinical tasks, reviews existing datasets and benchmarks, analyzes training-time and test-time techniques, and extends evaluation to include efficiency and explainability.

Result: The paradigm clarifies assumptions, standardizes comparisons between methods, and provides guidance for developing clinically meaningful LLMs.

Conclusion: The proposed framework helps guide the development of more clinically relevant LLMs by providing a structured approach to evaluate clinical decision-making tasks beyond simple accuracy metrics.

Abstract: Large language models (LLMs) show promise for clinical use. They are often evaluated using datasets such as MedQA. However, Many medical datasets, such as MedQA, rely on simplified Question-Answering (Q\A) that underrepresents real-world clinical decision-making. Based on this, we propose a unifying paradigm that characterizes clinical decision-making tasks along two dimensions: Clinical Backgrounds and Clinical Questions. As the background and questions approach the real clinical environment, the difficulty increases. We summarize the settings of existing datasets and benchmarks along two dimensions. Then we review methods to address clinical decision-making, including training-time and test-time techniques, and summarize when they help. Next, we extend evaluation beyond accuracy to include efficiency, explainability. Finally, we highlight open challenges. Our paradigm clarifies assumptions, standardizes comparisons, and guides the development of clinically meaningful LLMs.

Method: Construct large-scale Greek corpora with rigorous quality-based filtering and preprocessing, then pre-train diverse modern architectures including ELECTRA, ConvBERT and ModernBERT, plus bilingual Greek-English models for legal domain.

Result: GEM-RoBERTa and GEM-ConvBERT models significantly outperform existing baselines on downstream tasks, demonstrating the effectiveness of the proposed approach.

Conclusion: The new class of Greek Embedding Models establishes a successful approach for advancing Greek NLP through quality data curation and architectural diversity, particularly benefiting specialized domains like law.

Abstract: The advancement of natural language processing for morphologically rich, moderately-resourced languages like Modern Greek is often hindered by a fragmented research landscape, a lack of architectural diversity and reliance on limited context-length models. This is particularly true in specialized, high-value domains such as law, where existing models are frequently confined to early transformer architectures with a restrictive 512-token window, insufficient for analyzing long legal documents. To address these challenges, this paper presents Greek Embedding Models, a new family of transformer models for Greek language built upon a foundation of extensive, quality-driven data curation. We detail the construction of several large-scale Greek corpora, emphasizing a rigorous, quality-based filtering and preprocessing methodology to create high-value training datasets from both general-domain and specialized legal sources. On this carefully curated foundation, we pre-train and systematically evaluate a diverse suite of modern architectures, which has not previously applied to Greek language, such as ELECTRA, ConvBERT and ModernBERT. Furthermore, we propose the first bilingual Greek-English Embedding Models tailored for the legal domain. The extensive experiments on downstream tasks demonstrate that the new class of models establish the effectiveness of the proposed approach, highlighting that the GEM-RoBERTa and GEM-ConvBERT models significantly outperform existing baselines.

Method: Three main approaches: 1) Multi-task model incorporating domain-independent signals for event trigger detection, 2) Architectural modifications enabling bidirectional information flow in autoregressive LLMs, 3) Generative supervised in-context fine-tuning framework with response-oriented adaptation strategies.

Result: The proposed model, method, and framework demonstrate that pre-trained neural language models achieve their best performance on sequence labeling tasks when adapted through targeted transfer learning paradigms.

Conclusion: Targeted transfer learning paradigms significantly improve the adaptation of pre-trained neural language models for sequence labeling tasks, with multi-task learning, architectural modifications, and supervised in-context fine-tuning proving effective approaches.

Abstract: This doctoral thesis improves the transfer learning for sequence labeling tasks by adapting pre-trained neural language models. The proposed improvements in transfer learning involve introducing a multi-task model that incorporates an additional signal, a method based on architectural modifications in autoregressive large language models, and a sequence labeling framework for autoregressive large language models utilizing supervised in-context fine-tuning combined with response-oriented adaptation strategies. The first improvement is given in the context of domain transfer for the event trigger detection task. The domain transfer of the event trigger detection task can be improved by incorporating an additional signal obtained from a domain-independent text processing system into a multi-task model. The second improvement involves modifying the model’s architecture. For that purpose, a method is proposed to enable bidirectional information flow across layers of autoregressive large language models. The third improvement utilizes autoregressive large language models as text generators through a generative supervised in-context fine-tuning framework. The proposed model, method, and framework demonstrate that pre-trained neural language models achieve their best performance on sequence labeling tasks when adapted through targeted transfer learning paradigms.

Method: ToolScope includes: (1) ToolScopeMerger with Auto-Correction to automatically audit and fix tool merges, reducing redundancy, and (2) ToolScopeRetriever to rank and select only the most relevant tools for each query, compressing toolsets to fit within context limits.

Result: Evaluations on three state-of-the-art LLMs and three open-source tool-use benchmarks show gains of 8.38% to 38.6% in tool selection accuracy.

Conclusion: ToolScope effectively enhances LLM tool use by addressing tool redundancy and context limitations through automated tool merging and intelligent retrieval.

Abstract: Large language model (LLM) agents rely on external tools to solve complex tasks, but real-world toolsets often contain redundant tools with overlapping names and descriptions, introducing ambiguity and reducing selection accuracy. LLMs also face strict input context limits, preventing efficient consideration of large toolsets. To address these challenges, we propose ToolScope, which includes: (1) ToolScopeMerger with Auto-Correction to automatically audit and fix tool merges, reducing redundancy, and (2) ToolScopeRetriever to rank and select only the most relevant tools for each query, compressing toolsets to fit within context limits without sacrificing accuracy. Evaluations on three state-of-the-art LLMs and three open-source tool-use benchmarks show gains of 8.38% to 38.6% in tool selection accuracy, demonstrating ToolScope’s effectiveness in enhancing LLM tool use.

Method: Created Bengali Language Cultural Knowledge (BLanCK) dataset covering folk traditions, culinary arts, and regional dialects, then evaluated several multilingual language models.

Result: Models performed well in non-cultural categories but struggled significantly with cultural knowledge. Performance improved substantially across all models when context was provided.

Conclusion: Emphasizes the need for context-aware architectures and culturally curated training data for better cultural understanding in LLMs.

Abstract: Recent progress in NLP research has demonstrated remarkable capabilities of large language models (LLMs) across a wide range of tasks. While recent multilingual benchmarks have advanced cultural evaluation for LLMs, critical gaps remain in capturing the nuances of low-resource cultures. Our work addresses these limitations through a Bengali Language Cultural Knowledge (BLanCK) dataset including folk traditions, culinary arts, and regional dialects. Our investigation of several multilingual language models shows that while these models perform well in non-cultural categories, they struggle significantly with cultural knowledge and performance improves substantially across all models when context is provided, emphasizing context-aware architectures and culturally curated training data.

Method: Translated medical QA dataset to Persian, used RLAIF to generate rejected-preferred answer pairs with CoT reasoning, trained baseline model with DPO using 4.5M token dataset.

Result: Model outperformed gaokerena-V (trained on 57M tokens) despite using much smaller dataset, demonstrating enhanced medical reasoning in Persian.

Conclusion: Reasoning-focused training approaches are efficient and effective for developing domain-specific language models with limited data availability.

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

Method: Disentangled MAD into Majority Voting and inter-agent debate components, conducted experiments across 7 NLP benchmarks, and proposed a theoretical framework modeling debate as a stochastic process.

Result: Majority Voting alone accounts for most performance gains; debate induces a martingale over belief trajectories and doesn’t improve expected correctness; targeted interventions can enhance debate effectiveness.

Conclusion: While MAD has potential, simple ensembling methods remain stronger and more reliable alternatives in many practical settings.

Abstract: Multi-Agent Debate~(MAD) has emerged as a promising paradigm for improving the performance of large language models through collaborative reasoning. Despite recent advances, the key factors driving MAD’s effectiveness remain unclear. In this work, we disentangle MAD into two key components–Majority Voting and inter-agent Debate–and assess their respective contributions. Through extensive experiments across seven NLP benchmarks, we find that Majority Voting alone accounts for most of the performance gains typically attributed to MAD. To explain this, we propose a theoretical framework that models debate as a stochastic process. We prove that it induces a martingale over agents’ belief trajectories, implying that debate alone does not improve expected correctness. Guided by these insights, we demonstrate that targeted interventions, by biasing the belief update toward correction, can meaningfully enhance debate effectiveness. Overall, our findings suggest that while MAD has potential, simple ensembling methods remain strong and more reliable alternatives in many practical settings. Code is released in https://github.com/deeplearning-wisc/debate-or-vote.

Method: Proposed CreativityPrism framework that decomposes creativity into quality, novelty, and diversity dimensions, evaluated across three domains using 20 specific metrics on 17 state-of-the-art LLMs.

Result: Revealed performance gap between proprietary and open-source models; strong correlations within domains but weak across domains; diversity and quality metrics correlate strongly while novelty shows weak correlation with others.

Conclusion: Strong performance in one creativity task or dimension doesn’t generalize to others, highlighting the need for holistic evaluation of LLM creativity rather than fragmented assessments.

Abstract: Creativity is often seen as a hallmark of human intelligence. While large language models (LLMs) are increasingly perceived as producing creative text, there is still no holistic framework to evaluate their creativity across diverse scenarios. Existing evaluation methods remain fragmented, with dramatic variation across domains and tasks, largely due to differing definitions and measurements of creativity. Inspired by the hypothesis that creativity is not one fixed idea, we propose CreativityPrism, an evaluation analysis framework that decomposes creativity into three dimensions: quality, novelty, and diversity. CreativityPrism incorporates nine tasks, three domains, i.e., divergent thinking, creative writing, and logical reasoning, and twenty evaluation metrics, which measure each dimension in task-specific, unique ways. We evaluate 17 state-of-the-art (SoTA) proprietary and open-sourced LLMs on CreativityPrism and analyze the performance correlations among different metrics and task domains. Our results reveal a notable gap between proprietary and open-source models. Overall, model performance tends to be highly correlated across tasks within the same domain and less so across different domains. Among evaluation dimensions, diversity and quality metrics show strong correlations - models that perform well on one often excel on the other - whereas novelty exhibits much weaker correlation with either. These findings support our hypothesis that strong performance in one creativity task or dimension does not necessarily generalize to others, underscoring the need for a holistic evaluation of LLM creativity.

Method: ARTER computes complementary signals (embedding and LLM-based) over retrieved candidates to categorize mentions into easy and hard cases, then routes them to low-computational entity linker (ReFinED) or targeted LLM-based reasoning respectively.

Result: ARTER outperforms ReFinED by up to +4.47% with average gain of +2.53% on 5 out of 6 datasets, and performs comparably to full LLM-based pipelines while using half the LLM tokens.

Conclusion: ARTER provides an efficient and effective entity linking approach that strategically combines different reasoning methods, achieving strong performance without the computational cost of full LLM-based reasoning.

Abstract: Entity Linking (EL) has traditionally relied on large annotated datasets and extensive model fine-tuning. While recent few-shot methods leverage large language models (LLMs) through prompting to reduce training requirements, they often suffer from inefficiencies due to expensive LLM-based reasoning. ARTER (Adaptive Routing and Targeted Entity Reasoning) presents a structured pipeline that achieves high performance without deep fine-tuning by strategically combining candidate generation, context-based scoring, adaptive routing, and selective reasoning. ARTER computes a small set of complementary signals(both embedding and LLM-based) over the retrieved candidates to categorize contextual mentions into easy and hard cases. The cases are then handled by a low-computational entity linker (e.g. ReFinED) and more expensive targeted LLM-based reasoning respectively. On standard benchmarks, ARTER outperforms ReFinED by up to +4.47%, with an average gain of +2.53% on 5 out of 6 datasets, and performs comparably to pipelines using LLM-based reasoning for all mentions, while being as twice as efficient in terms of the number of LLM tokens.

Method: BoundRL performs token-level text segmentation and label prediction by generating starting tokens and reconstructing segments from original text. It uses reinforcement learning with verifiable rewards (RLVR) and incorporates intermediate candidates through systematic perturbation to prevent entropy collapse.

Result: BoundRL enables small language models (1.7B parameters) to outperform few-shot prompting of larger models. RLVR significantly improves over supervised fine-tuning, and intermediate candidates further enhance both performance and generalization.

Conclusion: BoundRL provides an efficient and effective solution for segmenting complex structured texts, reducing inference costs while maintaining high quality through innovative reinforcement learning and perturbation techniques.

Abstract: As structured texts become increasingly complex across diverse domains – from technical reports to generative AI prompts – the need for text segmentation into semantically meaningful components becomes critical. Such texts often contain elements beyond plain language, including tables, code snippets, and placeholders, which conventional sentence- or paragraph-level segmentation methods cannot handle effectively. To address this challenge, we propose BoundRL, a novel and efficient approach that jointly performs token-level text segmentation and label prediction for long structured texts. Instead of generating complete contents for each segment, it generates only a sequence of starting tokens and reconstructs the complete contents by locating these tokens within the original texts, thereby reducing inference costs by orders of magnitude and minimizing hallucination. To adapt the model for the output format, BoundRL~performs reinforcement learning with verifiable rewards (RLVR) with a specifically designed reward that jointly optimizes document reconstruction fidelity and semantic alignment. To mitigate entropy collapse, it further constructs intermediate candidates by systematically perturbing a fraction of generated sequences of segments to create stepping stones toward higher-quality solutions. To demonstrate BoundRL’s effectiveness on particularly challenging structured texts, we focus evaluation on complex prompts used for LLM applications. Experiments show that BoundRL enables small language models (1.7B parameters) to outperform few-shot prompting of much larger models. Moreover, RLVR with our designed reward yields significant improvements over supervised fine-tuning, and incorporating intermediate candidates further improves both performance and generalization.

Method: Automatically annotate posts in existing stance detection datasets with dialect/vernacular and text complexity attributes, then evaluate LLM performance in zero-shot setting.

Result: LLMs show significant stereotypes, such as associating pro-marijuana views with low text complexity and African American dialect with opposition to Donald Trump.

Conclusion: LLMs inherit and exhibit stereotypes from pretraining data in stance detection tasks, highlighting the need for bias mitigation in this sensitive NLP domain.

Abstract: Large Language Models inherit stereotypes from their pretraining data, leading to biased behavior toward certain social groups in many Natural Language Processing tasks, such as hateful speech detection or sentiment analysis. Surprisingly, the evaluation of this kind of bias in stance detection methods has been largely overlooked by the community. Stance Detection involves labeling a statement as being against, in favor, or neutral towards a specific target and is among the most sensitive NLP tasks, as it often relates to political leanings. In this paper, we focus on the bias of Large Language Models when performing stance detection in a zero-shot setting. We automatically annotate posts in pre-existing stance detection datasets with two attributes: dialect or vernacular of a specific group and text complexity/readability, to investigate whether these attributes influence the model’s stance detection decisions. Our results show that LLMs exhibit significant stereotypes in stance detection tasks, such as incorrectly associating pro-marijuana views with low text complexity and African American dialect with opposition to Donald Trump.

Method: The authors propose two methods to convert established datasets, creating a curated collection of 220 questions spanning 15 diverse domains that require both depth and width in information seeking.

Result: State-of-the-art agents achieve only 2.39% average success rate on DeepWideSearch, and error analysis reveals four key failure modes: lack of reflection, overreliance on internal knowledge, insufficient retrieval, and context overflow.

Conclusion: DeepWideSearch exposes critical limitations in current agent architectures and is publicly released to catalyze research on more capable and robust information-seeking agents.

Abstract: Current search agents fundamentally lack the ability to simultaneously perform \textit{deep} reasoning over multi-hop retrieval and \textit{wide}-scale information collection-a critical deficiency for real-world applications like comprehensive market analysis and business development. To bridge this gap, we introduce DeepWideSearch, the first benchmark explicitly designed to evaluate agents to integrate depth and width in information seeking. In DeepWideSearch, agents must process a large volume of data, each requiring deep reasoning over multi-hop retrieval paths. Specifically, we propose two methods to converse established datasets, resulting in a curated collection of 220 questions spanning 15 diverse domains. Extensive experiments demonstrate that even state-of-the-art agents achieve only 2.39% average success rate on DeepWideSearch, highlighting the substantial challenge of integrating depth and width search in information-seeking tasks. Furthermore, our error analysis reveals four failure modes: lack of reflection, overreliance on internal knowledge, insufficient retrieval, and context overflow-exposing key limitations in current agent architectures. We publicly release DeepWideSearch to catalyze future research on more capable and robust information-seeking agents.

Method: Proposes a multi-agent framework with three specialized roles (planning, coding, answering) and a self-improvement training framework using Monte Carlo Tree Search rollouts to generate pseudo-gold trajectories and optimize agents with reinforcement learning.

Result: Achieves 62.13% on TableBench, surpassing OpenAI-o4-mini-high, demonstrating substantial gains in table understanding performance.

Conclusion: The framework shows promise in combining structured multi-agent workflows with reinforcement learning to advance table understanding by integrating robust reasoning with reliable table processing.

Abstract: Understanding and reasoning over tables is a critical capability for many real-world applications. Large language models (LLMs) have shown promise on this task, but current approaches remain limited. Fine-tuning based methods strengthen language reasoning; yet they are prone to arithmetic errors and hallucination. In contrast, tool-based methods enable precise table manipulation but rely on rigid schemas and lack semantic understanding. These complementary drawbacks highlight the need for approaches that integrate robust reasoning with reliable table processing. In this work, we propose Mixture-of-Minds, a multi-agent framework that decomposes table reasoning into three specialized roles: planning, coding, and answering. This design enables each agent to focus on a specific aspect of the task while leveraging code execution for precise table manipulation. Building on this workflow, we introduce a self-improvement training framework that employs Monte Carlo Tree Search (MCTS) rollouts to generate pseudo-gold trajectories and optimize agents with reinforcement learning (RL). Extensive experiments show that Mixture-of-Minds delivers substantial gains, reaching 62.13% on TableBench and surpassing OpenAI-o4-mini-high. These results demonstrate the promise of combining structured multi-agent workflows with RL to advance table understanding.

Method: Tested LLMs on five spatial reasoning tasks (quadrant identification, geometric transformations, distance evaluation, word searches, tile sliding) with increasing grid dimensions and complexity.

Result: LLMs achieved moderate success on small-scale tasks but performance deteriorated rapidly with increasing complexity, with average accuracy loss of 42.7% and up to 84% loss. Tasks starting with over 50% accuracy showed at least 48% performance drop.

Conclusion: LLMs lack robust spatial representations in their underlying architectures, highlighting the gap between linguistic and spatial reasoning, and providing groundwork for future benchmarks combining language and geometry.

Abstract: This paper explores the spatial reasoning capability of large language models (LLMs) over textual input through a suite of five tasks aimed at probing their spatial understanding and computational abilities. The models were tested on both fundamental spatial reasoning and multi-step problem-solving within structured grid-based environments using tasks such as quadrant identification, geometric transformations, distance evaluation, word searches, and tile sliding. Each task was scaled in complexity through increasing grid dimensions, requiring models to extend beyond simple pattern recognition into abstract spatial reasoning. Our results reveal that while LLMs demonstrate moderate success in all tasks with small complexity and size, performance drops off rapidly as scale increases, with an average loss in accuracy of 42.7%, and reaching as high as 84%. Every test that began with over 50% accuracy showed a loss of at least 48%, illustrating the consistent nature of the deterioration. Furthermore, their struggles with scaling complexity hint at a lack of robust spatial representations in their underlying architectures. This paper underscores the gap between linguistic and spatial reasoning in LLMs, offering insights into their current limitations, and laying the groundwork for future integrative benchmarks at the intersection of language and geometry.

Method: Developed decoding-free sampling strategies that don’t require generation from LLMs, relying instead on cheap sampling methods that are model and tokenizer agnostic.

Result: Decoding-free sampling strategies provide sufficiently accurate marginal estimates at a small fraction of the runtime cost compared to traditional methods.

Conclusion: Decoding-free sampling enables efficient approximate marginalization for LLM evaluation, making it practical for downstream inference tasks while maintaining accuracy.

Abstract: Modern language models operate on subword-tokenized text in order to make a trade-off between model size, inference speed, and vocabulary coverage. A side effect of this is that, during inference, models are evaluated by measuring the probability of only the specific tokenization produced as the output, despite there being many possible ways to represent the same text with a subword vocabulary. Recent studies have argued instead for evaluating LLMs by marginalization - the probability mass of all tokenizations of a given text. Marginalization is difficult due to the number of possible tokenizations of a text, so often approximate marginalization is done via sampling. However, a downside of sampling is that an expensive generation step must be performed by the LLM for each sample, which limits the number of samples that can be acquired given a runtime budget, and therefore also the accuracy of the approximation. Since computing the probability of a sequence given the tokenization is relatively cheap compared to actually generating it, we investigate sampling strategies that are decoding-free - they require no generation from the LLM, instead relying entirely on extremely cheap sampling strategies that are model and tokenizer agnostic. We investigate the approximation quality and speed of decoding-free sampling strategies for a number of open models to find that they provide sufficiently accurate marginal estimates at a small fraction of the runtime cost and demonstrate its use on a set of downstream inference tasks.

Method: Synchronizes and fuses interview text (sentence-level transformer embeddings), audio (log Mel statistics with deltas), and facial signals (action units, gaze, head and pose descriptors) via calibrated late fusion classifier to output graded severities for depression (5 classes) and PTSD (3 classes).

Result: Outperforms unimodal/ablation baselines in stratified cross-validation. Fused model matches strongest unimodal baseline on accuracy and weighted F1, while improving decision curve utility and robustness under noisy/missing modalities. For PTSD, fusion reduces regression error and improves class concordance. Text contributes most to depression severity, audio and facial cues are critical for PTSD.

Conclusion: The approach offers reproducible evaluation and clinician-in-the-loop support for affective clinical decision making, with errors clustering between adjacent severities and reliable identification of extreme classes.

Abstract: Depression and post traumatic stress disorder (PTSD) often co-occur with connected symptoms, complicating automated assessment, which is often binary and disorder specific. Clinically useful diagnosis needs severity aware cross disorder estimates and decision support explanations. Our unified tri modal affective severity framework synchronizes and fuses interview text with sentence level transformer embeddings, audio with log Mel statistics with deltas, and facial signals with action units, gaze, head and pose descriptors to output graded severities for diagnosing both depression (PHQ-8; 5 classes) and PTSD (3 classes). Standardized features are fused via a calibrated late fusion classifier, yielding per disorder probabilities and feature-level attributions. This severity aware tri-modal affective fusion approach is demoed on multi disorder concurrent depression and PTSD assessment. Stratified cross validation on DAIC derived corpora outperforms unimodal/ablation baselines. The fused model matches the strongest unimodal baseline on accuracy and weighted F1, while improving decision curve utility and robustness under noisy or missing modalities. For PTSD specifically, fusion reduces regression error and improves class concordance. Errors cluster between adjacent severities; extreme classes are identified reliably. Ablations show text contributes most to depression severity, audio and facial cues are critical for PTSD, whereas attributions align with linguistic and behavioral markers. Our approach offers reproducible evaluation and clinician in the loop support for affective clinical decision making.

Method: Augments standard pretraining with next-context prediction objective that learns predictive representations of multi-token contexts using error signals from future token chunks, while remaining compatible with standard autoregressive evaluation.

Result: Consistent improvements in both perplexity and downstream task performance across GPT2 and Pythia model families up to 1.5B parameters, with better long-range coherence and attention allocation.

Conclusion: Next-context prediction provides a scalable and efficient pathway to stronger language modeling with minimal computational overhead.

Abstract: Next-token prediction (NTP) is the cornerstone of modern large language models (LLMs) pretraining, driving their unprecedented capabilities in text generation, reasoning, and instruction following. However, the token-level prediction limits the model’s capacity to capture higher-level semantic structures and long-range contextual relationships. To overcome this limitation, we introduce \textbf{ContextLM}, a framework that augments standard pretraining with an inherent \textbf{next-context prediction} objective. This mechanism trains the model to learn predictive representations of multi-token contexts, leveraging error signals derived from future token chunks. Crucially, ContextLM achieves this enhancement while remaining fully compatible with the standard autoregressive, token-by-token evaluation paradigm (e.g., perplexity). Extensive experiments on the GPT2 and Pythia model families, scaled up to $1.5$B parameters, show that ContextLM delivers consistent improvements in both perplexity and downstream task performance. Our analysis indicates that next-context prediction provides a scalable and efficient pathway to stronger language modeling, yielding better long-range coherence and more effective attention allocation with minimal computational overhead.

Method: Two-step approach: (1) Study citation failure using CITECONTROL benchmark that systematically varies relation between response and evidence, (2) Propose CITENTION framework integrating generative, attention-based, and retrieval-based methods to improve citations.

Result: Experiments show citation failures increase with relational complexity. CITENTION framework demonstrates substantial citation improvements on CITECONTROL benchmark and in transfer settings.

Conclusion: The work successfully disentangles citation failure from response failure, provides systematic analysis of failure modes, and offers an effective framework (CITENTION) to mitigate citation failures in LLM-based RAG systems.

Abstract: Citations from LLM-based RAG systems are supposed to simplify response verification. However, this does not hold for citation failure, when a model generates a helpful response, but fails to cite complete evidence. In contrast to previous work, we propose to disentangle this from response failure, where the response itself is flawed, and citing complete evidence is impossible. To address citation failure, this work follows a two-step approach: (1) We study when citation failure occurs and (2) how it can be mitigated. For step 1, we extend prior work by investigating how the relation between response and evidence affects citation quality. We introduce CITECONTROL, a benchmark that systematically varies this relation to analyze failure modes. Experiments show that failures increase with relational complexity and suggest that combining citation methods could improve performance, motivating step 2. To improve LLM citation efficiently, we propose CITENTION, a framework integrating generative, attention-based, and retrieval-based methods. Results demonstrate substantial citation improvements on CITECONTROL and in transfer settings. We make our data and code publicly available.

Method: The study evaluates state-of-the-art generative PRMs on TQA from both answer and step perspectives, analyzing PRMs that combine textual and code verification methods.

Result: Results show that PRMs can aid solution selection but struggle to generalize to out-of-domain data. Analysis reveals a weak correlation between performance in step-level verification and answer accuracy, possibly due to weak step dependencies and loose causal links.

Conclusion: The findings highlight limitations of current PRMs on TQA and offer valuable insights for building more robust, process-aware verifiers for semi-structured data tasks.

Abstract: Process reward models (PRMs) improve complex reasoning in large language models (LLMs) by grading candidate solutions step-by-step and selecting answers via aggregated step scores. While effective in domains such as mathematics, their applicability to tasks involving semi-structured data, like table question answering (TQA) remains unexplored. TQA poses unique challenges for PRMs, including abundant irrelevant information, loosely connected reasoning steps, and domain-specific reasoning. This work presents the first systematic study of PRMs for TQA. We evaluate state-of-the-art generative PRMs on TQA from both answer and step perspectives. Results show that PRMs that combine textual and code verification can aid solution selection but struggle to generalize to out-of-domain data. Analysis reveals a weak correlation between performance in step-level verification and answer accuracy, possibly stemming from weak step dependencies and loose causal links. Our findings highlight limitations of current PRMs on TQA and offer valuable insights for building more robust, process-aware verifiers.

Method: CoRT uses Hint-Engineering to synthesize high-quality code-integrated reasoning data by strategically injecting diverse hints at optimal points. It employs supervised fine-tuning with 30 samples, rejection sampling, and reinforcement learning to optimize multi-round CI usage and internal thinking.

Result: CoRT achieved absolute improvements of 4% on 32B model and 8% on 1.5B model across five mathematical reasoning datasets, while reducing token usage by 30% for 32B model and 50% for 1.5B model compared to natural language baselines.

Conclusion: CoRT effectively bridges the gap between large reasoning models and computational tools, significantly improving mathematical reasoning performance and efficiency through optimized code integration and multi-round interaction strategies.

Abstract: Large reasoning models (LRMs) like OpenAI-o1 have shown impressive capabilities in natural language reasoning. However, these models frequently demonstrate inefficiencies or inaccuracies when tackling complex mathematical operations. While integrating computational tools such as Code Interpreters (CIs) offers a promising solution, it introduces a critical challenge: a conflict between the model’s internal, probabilistic reasoning and the external, deterministic knowledge provided by the CI, which often leads models to unproductive deliberation. To overcome this, we introduce CoRT (Code-Optimized Reasoning Training), a post-training framework designed to teach LRMs to effectively utilize CIs. We propose \emph{Hint-Engineering}, a new data synthesis strategy that strategically injects diverse hints at optimal points within reasoning paths. This approach generates high-quality, code-integrated reasoning data specifically tailored to optimize LRM-CI interaction. Using this method, we have synthesized 30 high-quality samples to post-train models ranging from 1.5B to 32B parameters through supervised fine-tuning. CoRT further refines the multi-round interleaving of external CI usage and internal thinking by employing rejection sampling and reinforcement learning. Our experimental evaluations demonstrate CoRT’s effectiveness, yielding absolute improvements of 4% and 8% on DeepSeek-R1-Distill-Qwen-32B and DeepSeek-R1-Distill-Qwen-1.5B, respectively, across five challenging mathematical reasoning datasets. Moreover, CoRT significantly enhances efficiency, reducing token usage by approximately 30% for the 32B model and 50% for the 1.5B model compared to pure natural language reasoning baselines. The models and code are available at: https://github.com/ChengpengLi1003/CoRT.

Method: Conducted controlled probing experiments on tabular benchmarks like Adult Income and Titanic, testing performance with and without semantic cues by removing or randomizing column names and value categories.

Result: Contamination effects emerged only for datasets with strong semantic cues; performance dropped sharply to near-random levels when cues were removed or randomized.

Conclusion: LLMs’ tabular reasoning performance may reflect memorization of publicly available datasets rather than authentic reasoning, highlighting the need for evaluation protocols that disentangle semantic leakage from genuine reasoning ability.

Abstract: Large Language Models (LLMs) are increasingly evaluated on their ability to reason over structured data, yet such assessments often overlook a crucial confound: dataset contamination. In this work, we investigate whether LLMs exhibit prior knowledge of widely used tabular benchmarks such as Adult Income, Titanic, and others. Through a series of controlled probing experiments, we reveal that contamination effects emerge exclusively for datasets containing strong semantic cues-for instance, meaningful column names or interpretable value categories. In contrast, when such cues are removed or randomized, performance sharply declines to near-random levels. These findings suggest that LLMs’ apparent competence on tabular reasoning tasks may, in part, reflect memorization of publicly available datasets rather than genuine generalization. We discuss implications for evaluation protocols and propose strategies to disentangle semantic leakage from authentic reasoning ability in future LLM assessments.

Method: The framework shifts from static chunk segmentation to flexible retrieval by treating sentences as atomic units and supporting arbitrary sentence combinations, eliminating the need for semantic boundary detection.

Result: Experimental evaluation on LongBench V2 shows FreeChunker achieves superior retrieval performance compared to traditional chunking methods and significantly outperforms existing approaches in computational efficiency.

Conclusion: FreeChunker’s paradigm shift from fixed chunking to flexible sentence-based retrieval reduces computational overhead while enhancing adaptability to complex queries, demonstrating improved performance and efficiency.

Abstract: Chunking strategies significantly impact the effectiveness of Retrieval-Augmented Generation (RAG) systems. Existing methods operate within fixed-granularity paradigms that rely on static boundary identification, limiting their adaptability to diverse query requirements. This paper presents FreeChunker, a Cross-Granularity Encoding Framework that fundamentally transforms the traditional chunking paradigm: the framework treats sentences as atomic units and shifts from static chunk segmentation to flexible retrieval supporting arbitrary sentence combinations. This paradigm shift not only significantly reduces the computational overhead required for semantic boundary detection but also enhances adaptability to complex queries. Experimental evaluation on LongBench V2 demonstrates that FreeChunker achieves superior retrieval performance compared to traditional chunking methods, while significantly outperforming existing approaches in computational efficiency.

Method: Pre-trained llamalogue model on dialogue data, then applied various fine-tuning strategies including PPO and DPO to enhance communicative text generation.

Result: Models underperformed on most standard BabyLM benchmarks but excelled at dialogue continuation prediction. DPO fine-tuning improved performance on custom dialogue benchmark while PPO had mixed to adversarial effects.

Conclusion: Dialogue-specific pre-training creates specialized models for conversational tasks, with DPO fine-tuning being more effective than PPO for improving dialogue performance.

Abstract: We investigate whether pre-training exclusively on dialogue data results in formally and functionally apt small language models. Based on this pre-trained llamalogue model, we employ a variety of fine-tuning strategies to enforce “more communicative” text generations by our models. Although our models underperform on most standard BabyLM benchmarks, they excel at dialogue continuation prediction in a minimal pair setting. While PPO fine-tuning has mixed to adversarial effects on our models, DPO fine-tuning further improves their performance on our custom dialogue benchmark.

Method: Created the NegHalu dataset by reconstructing existing hallucination detection datasets with negated expressions, and analyzed LLM performance through experiments and token-level internal state tracing.

Result: LLMs demonstrate poor performance in detecting hallucinations in negated text, often making logically inconsistent or unfaithful judgments, with challenges in mitigating unintended effects.

Conclusion: Negation significantly impacts LLM hallucination detection capabilities, revealing important limitations that need to be addressed for reliable performance across different linguistic contexts.

Abstract: Recent studies on hallucination in large language models (LLMs) have been actively progressing in natural language processing. However, the impact of negated text on hallucination with LLMs remains largely unexplored. In this paper, we set three important yet unanswered research questions and aim to address them. To derive the answers, we investigate whether LLMs can recognize contextual shifts caused by negation and still reliably distinguish hallucinations comparable to affirmative cases. We also design the NegHalu dataset by reconstructing existing hallucination detection datasets with negated expressions. Our experiments demonstrate that LLMs struggle to detect hallucinations in negated text effectively, often producing logically inconsistent or unfaithful judgments. Moreover, we trace the internal state of LLMs as they process negated inputs at the token level and reveal the challenges of mitigating their unintended effects.

Method: The task comprises two subtasks: multimodal legal retrieval and multimodal question answering, using multimodal data (likely combining text and images) related to traffic sign regulations.

Result: Best-reported results are 64.55% F2 score for multimodal legal retrieval and 86.30% accuracy for multimodal question answering.

Conclusion: The VLSP 2025 MLQA-TSR shared task successfully provides a benchmark for multimodal legal AI systems in Vietnam, demonstrating promising performance on traffic sign regulation tasks.

Abstract: This paper presents the VLSP 2025 MLQA-TSR - the multimodal legal question answering on traffic sign regulation shared task at VLSP 2025. VLSP 2025 MLQA-TSR comprises two subtasks: multimodal legal retrieval and multimodal question answering. The goal is to advance research on Vietnamese multimodal legal text processing and to provide a benchmark dataset for building and evaluating intelligent systems in multimodal legal domains, with a focus on traffic sign regulation in Vietnam. The best-reported results on VLSP 2025 MLQA-TSR are an F2 score of 64.55% for multimodal legal retrieval and an accuracy of 86.30% for multimodal question answering.

Method: Trained using the same architecture as NeoBERT with dedicated focus on Hebrew texts, creating both monolingual and bilingual versions.

Result: Outperform existing models on almost all Hebrew benchmarks, with NeoDictaBERT-bilingual showing particularly strong results on retrieval tasks and outperforming other multilingual models of similar size.

Conclusion: The models provide a strong foundation for Hebrew NLP downstream tasks and are released to advance research and development in Hebrew NLP.

Abstract: Since their initial release, BERT models have demonstrated exceptional performance on a variety of tasks, despite their relatively small size (BERT-base has ~100M parameters). Nevertheless, the architectural choices used in these models are outdated compared to newer transformer-based models such as Llama3 and Qwen3. In recent months, several architectures have been proposed to close this gap. ModernBERT and NeoBERT both show strong improvements on English benchmarks and significantly extend the supported context window. Following their successes, we introduce NeoDictaBERT and NeoDictaBERT-bilingual: BERT-style models trained using the same architecture as NeoBERT, with a dedicated focus on Hebrew texts. These models outperform existing ones on almost all Hebrew benchmarks and provide a strong foundation for downstream tasks. Notably, the NeoDictaBERT-bilingual model shows strong results on retrieval tasks, outperforming other multilingual models of similar size. In this paper, we describe the training process and report results across various benchmarks. We release the models to the community as part of our goal to advance research and development in Hebrew NLP.

Method: Uses a teacher-student framework with novel alignment dataset for post-training. Experiments with adaptive teacher decoding strategies were conducted.

Result: BabyLM generates more grammatical and cohesive responses after post-training. Adaptive teacher decoding showed limited additional gains.

Conclusion: Targeted post-training improves dialogue quality, but contingency remains a challenging goal for BabyLMs.

Abstract: Multi-turn dialogues between a child and a caregiver are characterized by a property called contingency - that is, prompt, direct, and meaningful exchanges between interlocutors. We introduce ContingentChat, a teacher-student framework that benchmarks and improves multi-turn contingency in a BabyLM trained on 100M words. Using a novel alignment dataset for post-training, BabyLM generates responses that are more grammatical and cohesive. Experiments with adaptive teacher decoding strategies show limited additional gains. ContingentChat demonstrates the benefits of targeted post-training for dialogue quality and indicates that contingency remains a challenging goal for BabyLMs.

Method: Created MIXTURE dataset (144K samples) pairing low-quality instruction clusters with high-quality distillations. Used supervised fine-tuning followed by reinforcement learning with three reward signals: quality, semantic alignment, and format compliance via GRPO.

Result: LM-Mixup surpasses full-dataset training and competes with state-of-the-art data selection methods across multiple benchmarks, using only about 3% of the entire dataset.

Conclusion: Low-quality data is valuable when properly distilled with LM-Mixup, significantly enhancing instruction-tuned LLM efficiency and performance.

Abstract: Instruction tuning is crucial for aligning Large Language Models (LLMs), yet the quality of instruction-following data varies significantly. While high-quality data is paramount, it is often scarce; conversely, abundant low-quality data is frequently discarded, leading to substantial information loss. Existing data augmentation methods struggle to augment this low-quality data effectively, and the evaluation of such techniques remains poorly defined. To address this, we formally define the task of Instruction Distillation: distilling multiple low-quality and redundant inputs into high-quality and coherent instruction-output pairs. Specifically, we introduce a comprehensive data construction pipeline to create MIXTURE, a 144K-sample dataset pairing low-quality or semantically redundant imperfect instruction clusters with their high-quality distillations. We then introduce LM-Mixup, by first performing supervised fine-tuning on MIXTURE and then optimizing it with reinforcement learning. This process uses three complementary reward signals: quality, semantic alignment, and format compliance, via Group Relative Policy Optimization (GRPO). We demonstrate that LM-Mixup effectively augments imperfect datasets: fine-tuning LLMs on its distilled data, which accounts for only about 3% of the entire dataset, not only surpasses full-dataset training but also competes with state-of-the-art high-quality data selection methods across multiple benchmarks. Our work establishes that low-quality data is a valuable resource when properly distilled and augmented with LM-Mixup, significantly enhancing the efficiency and performance of instruction-tuned LLMs.

Method: Evaluated four confidence estimation approaches (VCE, MSP, Sample Consistency, CoCoA) on four question-answering tasks using state-of-the-art open-source LLM.

Result: Each uncertainty metric captures different facets of model confidence. CoCoA hybrid approach yields best overall reliability, improving both calibration and discrimination of correct answers.

Conclusion: CoCoA is recommended as the best uncertainty measure for LLM applications, with discussion of trade-offs between different methods.

Abstract: Large language models (LLMs) produce outputs with varying levels of uncertainty, and, just as often, varying levels of correctness; making their practical reliability far from guaranteed. To quantify this uncertainty, we systematically evaluate four approaches for confidence estimation in LLM outputs: VCE, MSP, Sample Consistency, and CoCoA (Vashurin et al., 2025). For the evaluation of the approaches, we conduct experiments on four question-answering tasks using a state-of-the-art open-source LLM. Our results show that each uncertainty metric captures a different facet of model confidence and that the hybrid CoCoA approach yields the best reliability overall, improving both calibration and discrimination of correct answers. We discuss the trade-offs of each method and provide recommendations for selecting uncertainty measures in LLM applications.

Method: Developed an improved MLM approach that adapts masking probabilities based on model’s prediction ability, and incorporated sub-token embeddings for better morphological generalization.

Result: Substantial performance increase on (Super)GLUE tasks compared to standard MLM, and the submission beat the baseline in the strict-small track.

Conclusion: The adaptive MLM approach with sub-token embeddings effectively improves language model performance on benchmark tasks.

Abstract: We describe our strategy for the 2025 edition of the BabyLM Challenge. Our main contribution is that of an improved form of Masked Language Modeling (MLM), which adapts the probabilities of the tokens masked according to the model’s ability to predict them. The results show a substantial increase in performance on (Super)GLUE tasks over the standard MLM. We also incorporate sub-token embeddings, finding that this increases the model’s morphological generalization capabilities. Our submission beats the baseline in the strict-small track.

Method: Computes inter-model similarity from layer-wise hidden representations over clustered typical samples, then performs adaptive hierarchical parameter fusion to align knowledge across models while preserving domain-general features in shallow layers and allowing task-specific adaptation in deeper layers.

Result: Outperforms baselines in both knowledge retention and generalization across five NLP tasks and multiple continual learning scenarios, achieving seamless multi-domain integration and strong resistance to catastrophic forgetting.

Conclusion: RECALL provides a scalable and data-free solution for evolving LLMs that enables effective continual learning without performance trade-offs or need for task labels.

Abstract: We unveil that internal representations in large language models (LLMs) serve as reliable proxies of learned knowledge, and propose RECALL, a novel representation-aware model merging framework for continual learning without access to historical data. RECALL computes inter-model similarity from layer-wise hidden representations over clustered typical samples, and performs adaptive, hierarchical parameter fusion to align knowledge across models. This design enables the preservation of domain-general features in shallow layers while allowing task-specific adaptation in deeper layers. Unlike prior methods that require task labels or incur performance trade-offs, RECALL achieves seamless multi-domain integration and strong resistance to catastrophic forgetting. Extensive experiments across five NLP tasks and multiple continual learning scenarios show that RECALL outperforms baselines in both knowledge retention and generalization, providing a scalable and data-free solution for evolving LLMs.

Method: Two-step training: continued pretraining on documents with factual descriptions, then expert iteration to use Python type hints in evaluation settings. Activation steering using original model’s vectors to suppress evaluation-awareness.

Result: Steering successfully suppressed evaluation-awareness - model acted like deployed even when evaluation cues were present, closing the gap between evaluation and deployment behavior.

Conclusion: Activation steering can improve safety evaluation reliability by making models behave as if deployed, preventing strategic alignment adjustments during testing.

Abstract: Large language models (LLMs) can sometimes detect when they are being evaluated and adjust their behavior to appear more aligned, compromising the reliability of safety evaluations. In this paper, we show that adding a steering vector to an LLM’s activations can suppress evaluation-awareness and make the model act like it is deployed during evaluation. To study our steering technique, we train an LLM to exhibit evaluation-aware behavior using a two-step training process designed to mimic how this behavior could emerge naturally. First, we perform continued pretraining on documents with factual descriptions of the model (1) using Python type hints during evaluation but not during deployment and (2) recognizing that the presence of a certain evaluation cue always means that it is being tested. Then, we train the model with expert iteration to use Python type hints in evaluation settings. The resulting model is evaluation-aware: it writes type hints in evaluation contexts more than deployment contexts. However, this gap can only be observed by removing the evaluation cue. We find that activation steering can suppress evaluation awareness and make the model act like it is deployed even when the cue is present. Importantly, we constructed our steering vector using the original model before our additional training. Our results suggest that AI evaluators could improve the reliability of safety evaluations by steering models to act like they are deployed.

Method: Robust Preference Selection (RPS) uses directional neighborhood consensus to sample multiple responses from related preferences and select the best-aligned one, without requiring model retraining.

Result: RPS consistently improves robustness across three alignment paradigms (DPA, DPO, SFT), achieving win rates up to 69% on challenging underrepresented preferences.

Conclusion: RPS provides a practical, theoretically-grounded solution for enhancing preference-aligned model reliability without costly retraining.

Abstract: Aligning large language models with human preferences is critical for creating reliable and controllable AI systems. A human preference can be visualized as a high-dimensional vector where different directions represent trade-offs between desired attributes (e.g., helpfulness vs. verbosity). Yet, because the training data often reflects dominant, average preferences, LLMs tend to perform well on common requests but fall short in specific, individual needs. This mismatch creates a preference coverage gap. Existing methods often address this through costly retraining, which may not be generalized to the full spectrum of diverse preferences. This brittleness means that when a user’s request reflects a nuanced preference deviating from the training data’s central tendency, model performance can degrade unpredictably. To address this challenge, we introduce Robust Preference Selection (RPS), a post-hoc, training-free method by leveraging directional neighborhood consensus. Instead of forcing a model to generate a response from a single, highly specific preference, RPS samples multiple responses from a local neighborhood of related preferences to create a superior candidate pool. It then selects the response that best aligns with the user’s original intent. We provide a theoretical framework showing our neighborhood generation strategy is provably superior to a strong baseline that also samples multiple candidates. Comprehensive experiments across three distinct alignment paradigms (DPA, DPO, and SFT) demonstrate that RPS consistently improves robustness against this baseline, achieving win rates of up to 69% on challenging preferences from under-represented regions of the space without any model retraining. Our work presents a practical, theoretically-grounded solution for enhancing the reliability of preference-aligned models.

Method: Hierarchical Sequence (HSEQ) Iteration linearizes documents, tables, and knowledge graphs into reversible hierarchical sequences with structural tags, then performs structure-aware iteration with Head and Iteration Agents that guide retrieval and expansion.

Result: Experiments on HotpotQA, HybridQA/TAT-QA, and MetaQA show consistent EM/F1 gains over strong baselines with high efficiency, demonstrating format-agnostic unification and budget-aware iteration.

Conclusion: HSEQ provides a unified framework that enables single policy operation across text, tables, and KGs, reduces unnecessary hops and tokens while preserving accuracy, and improves answer consistency and auditability through evidence canonicalization.

Abstract: Retrieval-augmented generation (RAG) remains brittle on multi-step questions and heterogeneous evidence sources, trading accuracy against latency and token/tool budgets. This paper introducesHierarchical Sequence (HSEQ) Iteration for Heterogeneous Question Answering, a unified framework that (i) linearize documents, tables, and knowledge graphs into a reversible hierarchical sequence with lightweight structural tags, and (ii) perform structure-aware iteration to collect just-enough evidence before answer synthesis. A Head Agent provides guidance that leads retrieval, while an Iteration Agent selects and expands HSeq via structure-respecting actions (e.g., parent/child hops, table row/column neighbors, KG relations); Finally the head agent composes canonicalized evidence to genearte the final answer, with an optional refinement loop to resolve detected contradictions. Experiments on HotpotQA (text), HybridQA/TAT-QA (table+text), and MetaQA (KG) show consistent EM/F1 gains over strong single-pass, multi-hop, and agentic RAG baselines with high efficiency. Besides, HSEQ exhibits three key advantages: (1) a format-agnostic unification that enables a single policy to operate across text, tables, and KGs without per-dataset specialization; (2) guided, budget-aware iteration that reduces unnecessary hops, tool calls, and tokens while preserving accuracy; and (3) evidence canonicalization for reliable QA, improving answers consistency and auditability.

Method: Created a 21-way multiparallel version of EuroParl dataset with political affiliations, analyzed translation quality of speeches across political parties using 1.5M sentences covering 1000+ speakers, 7 countries, and 12 EU parties.

Result: Systematic translation differences observed where majority parties (left, center, right) are better translated than outsider parties, revealing political bias in LLM translation performance.

Conclusion: Political biases exist in LLMs’ multilingual translation capabilities, with systematic favoritism toward established political parties over outsider groups, highlighting fairness concerns in AI systems.

Abstract: The political biases of Large Language Models (LLMs) are usually assessed by simulating their answers to English surveys. In this work, we propose an alternative framing of political biases, relying on principles of fairness in multilingual translation. We systematically compare the translation quality of speeches in the European Parliament (EP), observing systematic differences with majority parties from left, center, and right being better translated than outsider parties. This study is made possible by a new, 21-way multiparallel version of EuroParl, the parliamentary proceedings of the EP, which includes the political affiliations of each speaker. The dataset consists of 1.5M sentences for a total of 40M words and 249M characters. It covers three years, 1000+ speakers, 7 countries, 12 EU parties, 25 EU committees, and hundreds of national parties.

Method: Developed an Adaptable text Representations Compressor (ARC-Encoder) that compresses context into continuous representations, systematically studying training strategies and architecture choices.

Result: Achieves state-of-the-art performance on multiple benchmarks while improving computational efficiency, and can be adapted to work with multiple decoder LLMs simultaneously.

Conclusion: ARC-Encoder provides a flexible and efficient solution for portable encoders that work seamlessly across different LLMs, offering a practical alternative to fine-tuning-based compression methods.

Abstract: Recent techniques such as retrieval-augmented generation or chain-of-thought reasoning have led to longer contexts and increased inference costs. Context compression techniques can reduce these costs, but the most effective approaches require fine-tuning the target model or even modifying its architecture. This can degrade its general abilities when not used for this specific purpose. Here we explore an alternative approach: an encoder that compresses the context into continuous representations which replace token embeddings in decoder LLMs. First, we perform a systematic study of training strategies and architecture choices for the encoder. Our findings led to the design of an Adaptable text Representations Compressor, named ARC-Encoder, which outputs $x$-times fewer continuous representations (typically $x!\in!{4,8}$) than text tokens. We evaluate ARC-Encoder across a variety of LLM usage scenarios, ranging from in-context learning to context window extension, on both instruct and base decoders. Results show that ARC-Encoder achieves state-of-the-art performance on several benchmarks while improving computational efficiency at inference. Finally, we demonstrate that our models can be adapted to multiple decoders simultaneously, allowing a single encoder to generalize across different decoder LLMs. This makes ARC-Encoder a flexible and efficient solution for portable encoders that work seamlessly with multiple LLMs. We release a training code at https://github.com/kyutai-labs/ARC-Encoder , fine-tuning dataset and pretrained models are available at https://huggingface.co/collections/kyutai/arc-encoders-68ee18787301407d60a57047 .

Method: Created CenterBench dataset with 9,720 comprehension questions on center-embedded sentences with recursive relative clauses, including syntactically identical but semantically implausible counterparts and questions testing surface understanding, syntactic dependencies, and causal reasoning.

Result: Models show performance gaps up to 26.8 percentage points between plausible and implausible sentences that widen with complexity, indicating they abandon structural analysis for semantic associations. Semantic plausibility actually harms performance on action-related questions where causal reasoning matters more.

Conclusion: CenterBench provides the first framework to identify when models shift from structural analysis to pattern matching, revealing systematic differences between model and human processing of semantic plausibility.

Abstract: When language models correctly parse “The cat that the dog chased meowed,” are they analyzing syntax or simply familiar with dogs chasing cats? Despite extensive benchmarking, we lack methods to distinguish structural understanding from semantic pattern matching. We introduce CenterBench, a dataset of 9,720 comprehension questions on center-embedded sentences (like “The cat [that the dog chased] meowed”) where relative clauses nest recursively, creating processing demands from simple to deeply nested structures. Each sentence has a syntactically identical but semantically implausible counterpart (e.g., mailmen prescribe medicine, doctors deliver mail) and six comprehension questions testing surface understanding, syntactic dependencies, and causal reasoning. Testing six models reveals that performance gaps between plausible and implausible sentences widen systematically with complexity, with models showing median gaps up to 26.8 percentage points, quantifying when they abandon structural analysis for semantic associations. Notably, semantic plausibility harms performance on questions about resulting actions, where following causal relationships matters more than semantic coherence. Reasoning models improve accuracy but their traces show semantic shortcuts, overthinking, and answer refusal. Unlike models whose plausibility advantage systematically widens with complexity, humans shows variable semantic effects. CenterBench provides the first framework to identify when models shift from structural analysis to pattern matching.

Method: Proposes GlobalRAG framework that decomposes questions into subgoals, coordinates retrieval with reasoning, and refines evidence iteratively. Introduces Planning Quality Reward and SubGoal Completion Reward to encourage coherent planning and reliable execution. Uses progressive weight annealing to balance process-oriented and outcome-based objectives.

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

Conclusion: GlobalRAG effectively addresses global planning and faithful execution challenges in multi-hop QA through its reinforcement learning framework with specialized rewards and progressive annealing, demonstrating superior performance with reduced training data requirements.

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

Method: Proposes MACDF (Multi-Agent Cognitive Decision Framework) that shifts from passive retrieval to proactive decision support using multi-agent cognitive systems.

Result: Extensive offline evaluations show significant improvements in recommendation accuracy and user satisfaction, especially for complex queries. Online A/B testing on JD search platform confirms practical efficacy.

Conclusion: The work demonstrates the transformative potential of multi-agent cognitive systems in redefining e-commerce search by better aligning with users’ cognitive decision processes.

Abstract: The retrieval-ranking paradigm has long dominated e-commerce search, but its reliance on query-item matching fundamentally misaligns with multi-stage cognitive decision processes of platform users. This misalignment introduces critical limitations: semantic gaps in complex queries, high decision costs due to cross-platform information foraging, and the absence of professional shopping guidance. To address these issues, we propose a Multi-Agent Cognitive Decision Framework (MACDF), which shifts the paradigm from passive retrieval to proactive decision support. Extensive offline evaluations demonstrate MACDF’s significant improvements in recommendation accuracy and user satisfaction, particularly for complex queries involving negation, multi-constraint, or reasoning demands. Online A/B testing on JD search platform confirms its practical efficacy. This work highlights the transformative potential of multi-agent cognitive systems in redefining e-commerce search.

Method: Used ChatGPT to code communication data according to collaborative problem solving frameworks, analyzing data from three collaborative task types: negotiation, problem solving, and decision making.

Result: ChatGPT-based coding exhibited no significant bias across gender and racial groups in all three types of collaborative tasks.

Conclusion: The findings support the adoption of ChatGPT-based automated coding for large-scale assessment of collaboration and communication, as it demonstrates fairness across demographic groups.

Abstract: Assessing communication and collaboration at scale depends on a labor intensive task of coding communication data into categories according to different frameworks. Prior research has established that ChatGPT can be directly instructed with coding rubrics to code the communication data and achieves accuracy comparable to human raters. However, whether the coding from ChatGPT or similar AI technology exhibits bias against different demographic groups, such as gender and race, remains unclear. To fill this gap, this paper investigates ChatGPT-based automated coding of communication data using a typical coding framework for collaborative problem solving, examining differences across gender and racial groups. The analysis draws on data from three types of collaborative tasks: negotiation, problem solving, and decision making. Our results show that ChatGPT-based coding exhibits no significant bias across gender and racial groups, paving the road for its adoption in large-scale assessment of collaboration and communication.

Method: Fine-tuned three pre-trained transformer models (AraELECTRA, CAMeLBERT, and XLM-RoBERTa) on the provided dataset for binary classification of AI-generated vs human-written Arabic text.

Result: XLM-RoBERTa achieved the highest performance with an F1 score of 0.7701, outperforming the specialized Arabic models AraELECTRA and CAMeLBERT.

Conclusion: Multilingual models like XLM-RoBERTa have strong generalization capabilities for AI-generated text detection, even outperforming specialized monolingual models, highlighting the complexities of this task.

Abstract: This paper details our submission to the Ara- GenEval Shared Task on Arabic AI-generated text detection, where our team, BUSTED, se- cured 5th place. We investigated the effec- tiveness of three pre-trained transformer mod- els: AraELECTRA, CAMeLBERT, and XLM- RoBERTa. Our approach involved fine-tuning each model on the provided dataset for a binary classification task. Our findings revealed a sur- prising result: the multilingual XLM-RoBERTa model achieved the highest performance with an F1 score of 0.7701, outperforming the spe- cialized Arabic models. This work underscores the complexities of AI-generated text detection and highlights the strong generalization capa- bilities of multilingual models.

Method: Used the Five-Dimensional Curiosity scale Revised (5DCR) framework to comprehensively evaluate LLM curiosity across dimensions including Information Seeking, Thrill Seeking, and Social Curiosity.

Result: LLMs demonstrated stronger thirst for knowledge than humans but made conservative choices in uncertain environments. Curious behaviors were found to enhance the models’ reasoning and active learning abilities.

Conclusion: LLMs have potential to exhibit human-like curiosity, providing experimental support for future development of learning capabilities and innovative research in LLMs.

Abstract: Curiosity serves as a pivotal conduit for human beings to discover and learn new knowledge. Recent advancements of large language models (LLMs) in natural language processing have sparked discussions regarding whether these models possess capability of curiosity-driven learning akin to humans. In this paper, starting from the human curiosity assessment questionnaire Five-Dimensional Curiosity scale Revised (5DCR), we design a comprehensive evaluation framework that covers dimensions such as Information Seeking, Thrill Seeking, and Social Curiosity to assess the extent of curiosity exhibited by LLMs. The results demonstrate that LLMs exhibit a stronger thirst for knowledge than humans but still tend to make conservative choices when faced with uncertain environments. We further investigated the relationship between curiosity and thinking of LLMs, confirming that curious behaviors can enhance the model’s reasoning and active learning abilities. These findings suggest that LLMs have the potential to exhibit curiosity similar to that of humans, providing experimental support for the future development of learning capabilities and innovative research in LLMs.

Method: Systematic comparison of LRM reasoning in English vs. question language across MGSM and GPQA Diamond tasks, analyzing both answer accuracy and cognitive attributes in reasoning traces.

Result: English reasoning traces show substantially higher cognitive behaviors and generally yield higher accuracy, with performance gap increasing for more complex tasks. However, English-centric approach is susceptible to ‘Lost in Translation’ errors.

Conclusion: While English reasoning generally improves performance, it introduces translation-related failure modes that could be avoided by reasoning in the question’s original language.

Abstract: Large Reasoning Models (LRMs) achieve strong performance on mathematical, scientific, and other question-answering tasks, but their multilingual reasoning abilities remain underexplored. When presented with non-English questions, LRMs often default to reasoning in English, raising concerns about interpretability and the handling of linguistic and cultural nuances. We systematically compare an LRM’s reasoning in English versus the language of the question. Our evaluation spans two tasks: MGSM and GPQA Diamond. Beyond measuring answer accuracy, we also analyze cognitive attributes in the reasoning traces. We find that English reasoning traces exhibit a substantially higher presence of these cognitive behaviors, and that reasoning in English generally yields higher final-answer accuracy, with the performance gap increasing as tasks become more complex. However, this English-centric strategy is susceptible to a key failure mode - getting “Lost in Translation,” where translation steps lead to errors that would have been avoided by question’s language reasoning.

Method: Created CantoNLU benchmark with 7 NLU tasks covering syntax and semantics, tested four model types: Mandarin model, Cantonese-adapted models (continual pre-training), and monolingual Cantonese model.

Result: Cantonese-adapted models performed best overall, monolingual models performed better on syntactic tasks, and Mandarin models remained competitive when Cantonese data is scarce.

Conclusion: Direct transfer from Mandarin may be sufficient when Cantonese data is limited, and the benchmark with datasets/code/model weights is released to facilitate future Cantonese NLP research.

Abstract: Cantonese, although spoken by millions, remains under-resourced due to policy and diglossia. To address this scarcity of evaluation frameworks for Cantonese, we introduce \textsc{\textbf{CantoNLU}}, a benchmark for Cantonese natural language understanding (NLU). This novel benchmark spans seven tasks covering syntax and semantics, including word sense disambiguation, linguistic acceptability judgment, language detection, natural language inference, sentiment analysis, part-of-speech tagging, and dependency parsing. In addition to the benchmark, we provide model baseline performance across a set of models: a Mandarin model without Cantonese training, two Cantonese-adapted models obtained by continual pre-training a Mandarin model on Cantonese text, and a monolingual Cantonese model trained from scratch. Results show that Cantonese-adapted models perform best overall, while monolingual models perform better on syntactic tasks. Mandarin models remain competitive in certain settings, indicating that direct transfer may be sufficient when Cantonese domain data is scarce. We release all datasets, code, and model weights to facilitate future research in Cantonese NLP.

Method: Introduces ND-LoRA (Neural Diversity Low-Rank Adaptation), which combines parallel LoRA adapters with Barlow Twins regularization. The approach is inspired by portfolio theory, with mathematical proof showing hallucination probability is bounded by representational correlation.

Result: ND-LoRA reduces hallucinations by up to 25.6% (14.6% on average) without degrading general accuracy. Ablations show synergistic effects, causal interventions confirm neurodiversity as the mediating factor, and correlation analysis reveals that a 0.1% neural correlation increase leads to 3.8% hallucination increase.

Conclusion: Neural diversity serves as a third axis of scaling - orthogonal to parameters and data - that can improve language model reliability at fixed budgets. Different tasks require different optimal amounts of neurodiversity, highlighting task-dependent optimality.

Abstract: Language models continue to hallucinate despite increases in parameters, compute, and data. We propose neural diversity – decorrelated parallel representations – as a principled mechanism that reduces hallucination rates at fixed parameter and data budgets. Inspired by portfolio theory, where uncorrelated assets reduce risk by $\sqrt{P}$, we prove hallucination probability is bounded by representational correlation: $P(H) \leq f(\sigma^2((1-\rho(P))/P + \rho(P)), \mu^2)$, which predicts that language models need an optimal amount of neurodiversity. To validate this, we introduce ND-LoRA (Neural Diversity Low-Rank Adaptation), combining parallel LoRA adapters with Barlow Twins regularization, and demonstrate that ND-LoRA reduces hallucinations by up to 25.6% (and 14.6% on average) without degrading general accuracy. Ablations show LoRA adapters and regularization act synergistically, causal interventions prove neurodiversity as the mediating factor and correlational analyses indicate scale: a 0.1% neural correlation increase is associated with a 3.8% hallucination increase. Finally, task-dependent optimality emerges: different tasks require different amounts of optimal neurodiversity. Together, our results highlight neural diversity as a third axis of scaling – orthogonal to parameters and data – to improve the reliability of language models at fixed budgets.

Method: Three lightweight adaptations to MBR utility functions to increase sensitivity to structural variability. Created a dataset with three types of latent structure (dialogue act, emotion, response structure) and proposed two metrics for structural optimality evaluation.

Result: Common similarity-based utility functions perform poorly on structural optimality metrics. The proposed adaptations significantly improve structural optimality and achieve up to 13.7 percentage point improvement in win rate on AlpacaEval and MT-Bench benchmarks.

Conclusion: Making MBR more sensitive to structural variability in the outcome space improves generation quality in open-ended tasks like instruction-following, demonstrating the importance of considering latent structures in generation strategies.

Abstract: Minimum Bayes Risk (MBR) decoding has seen renewed interest as an alternative to traditional generation strategies. While MBR has proven effective in machine translation, where the variability of a language model’s outcome space is naturally constrained, it may face challenges in more open-ended tasks such as dialogue or instruction-following. We hypothesise that in such settings, applying MBR with standard similarity-based utility functions may result in selecting responses that are broadly representative of the model’s distribution, yet sub-optimal with respect to any particular grouping of generations that share an underlying latent structure. In this work, we introduce three lightweight adaptations to the utility function, designed to make MBR more sensitive to structural variability in the outcome space. To test our hypothesis, we curate a dataset capturing three representative types of latent structure: dialogue act, emotion, and response structure (e.g., a sentence, a paragraph, or a list). We further propose two metrics to evaluate the structural optimality of MBR. Our analysis demonstrates that common similarity-based utility functions fall short by these metrics. In contrast, our proposed adaptations considerably improve structural optimality. Finally, we evaluate our approaches on real-world instruction-following benchmarks, AlpacaEval and MT-Bench, and show that increased structural sensitivity improves generation quality by up to 13.7 percentage points in win rate.

Method: Conducted a user study with 94 participants using 90 scenarios from PrivacyLens to evaluate how users perceive privacy-preservation quality and helpfulness of LLM responses.

Result: Users showed low agreement on privacy-preservation and helpfulness evaluations, while proxy LLMs had high agreement but low correlation with user perceptions.

Conclusion: Privacy and helpfulness of LLM responses are individual-specific, proxy LLMs are poor estimators of user perceptions, and user-centered studies are needed for better evaluation.

Abstract: Large language models (LLMs) have seen rapid adoption for tasks such as drafting emails, summarizing meetings, and answering health questions. In such uses, users may need to share private information (e.g., health records, contact details). To evaluate LLMs’ ability to identify and redact such private information, prior work developed benchmarks (e.g., ConfAIde, PrivacyLens) with real-life scenarios. Using these benchmarks, researchers have found that LLMs sometimes fail to keep secrets private when responding to complex tasks (e.g., leaking employee salaries in meeting summaries). However, these evaluations rely on LLMs (proxy LLMs) to gauge compliance with privacy norms, overlooking real users’ perceptions. Moreover, prior work primarily focused on the privacy-preservation quality of responses, without investigating nuanced differences in helpfulness. To understand how users perceive the privacy-preservation quality and helpfulness of LLM responses to privacy-sensitive scenarios, we conducted a user study with 94 participants using 90 scenarios from PrivacyLens. We found that, when evaluating identical responses to the same scenario, users showed low agreement with each other on the privacy-preservation quality and helpfulness of the LLM response. Further, we found high agreement among five proxy LLMs, while each individual LLM had low correlation with users’ evaluations. These results indicate that the privacy and helpfulness of LLM responses are often specific to individuals, and proxy LLMs are poor estimates of how real users would perceive these responses in privacy-sensitive scenarios. Our results suggest the need to conduct user-centered studies on measuring LLMs’ ability to help users while preserving privacy. Additionally, future research could investigate ways to improve the alignment between proxy LLMs and users for better estimation of users’ perceived privacy and utility.

Method: Developed and compared rule-based, machine learning (RF, SVM, LR), deep learning (BERT, ClinicalBERT), and LLM-based (zero-shot and error-analysis prompting) NLP approaches on 236 annotated clinical notes.

Result: Error-analysis prompting achieved F1=1.000 for both treatment and toxicities extraction, outperforming all other methods. Zero-shot prompting reached F1=1.000 for treatment and F1=0.876 for toxicities. Traditional ML methods ranked second for toxicities.

Conclusion: LLM-based NLP is most effective for extracting fluoropyrimidine treatment and toxicity information from clinical notes, with strong potential for oncology research and pharmacovigilance.

Abstract: Objective: Fluoropyrimidines are widely prescribed for colorectal and breast cancers, but are associated with toxicities such as hand-foot syndrome and cardiotoxicity. Since toxicity documentation is often embedded in clinical notes, we aimed to develop and evaluate natural language processing (NLP) methods to extract treatment and toxicity information. Materials and Methods: We constructed a gold-standard dataset of 236 clinical notes from 204,165 adult oncology patients. Domain experts annotated categories related to treatment regimens and toxicities. We developed rule-based, machine learning-based (Random Forest, Support Vector Machine [SVM], Logistic Regression [LR]), deep learning-based (BERT, ClinicalBERT), and large language models (LLM)-based NLP approaches (zero-shot and error-analysis prompting). Models used an 80:20 train-test split. Results: Sufficient data existed to train and evaluate 5 annotated categories. Error-analysis prompting achieved optimal precision, recall, and F1 scores (F1=1.000) for treatment and toxicities extraction, whereas zero-shot prompting reached F1=1.000 for treatment and F1=0.876 for toxicities extraction.LR and SVM ranked second for toxicities (F1=0.937). Deep learning underperformed, with BERT (F1=0.873 treatment; F1= 0.839 toxicities) and ClinicalBERT (F1=0.873 treatment; F1 = 0.886 toxicities). Rule-based methods served as our baseline with F1 scores of 0.857 in treatment and 0.858 in toxicities. Discussion: LMM-based approaches outperformed all others, followed by machine learning methods. Machine and deep learning approaches were limited by small training data and showed limited generalizability, particularly for rare categories. Conclusion: LLM-based NLP most effectively extracted fluoropyrimidine treatment and toxicity information from clinical notes, and has strong potential to support oncology research and pharmacovigilance.

Method: Proposed calibrating LRM thinking by training them on synthetic, human-like thinking trajectories to address identified challenges.

Result: The approach reduced thinking budgets by ~35x while improving evaluation performance across different LRM scales (7B to 32B), with R1-Distill-Qwen-7B achieving +8.7 correlation point improvement on WMT24 Metrics benchmarks.

Conclusion: Efficiently calibrated LRMs have significant potential to advance fine-grained automatic machine translation evaluation.

Abstract: Recent advancements in large reasoning models (LRMs) have introduced an intermediate “thinking” process prior to generating final answers, improving their reasoning capabilities on complex downstream tasks. However, the potential of LRMs as evaluators for machine translation (MT) quality remains underexplored. We provides the first systematic analysis of LRM-as-a-judge in MT evaluation. We identify key challenges, revealing LRMs require tailored evaluation materials, tend to “overthink” simpler instances and have issues with scoring mechanisms leading to overestimation. To address these, we propose to calibrate LRM thinking by training them on synthetic, human-like thinking trajectories. Our experiments on WMT24 Metrics benchmarks demonstrate that this approach largely reduces thinking budgets by ~35x while concurrently improving evaluation performance across different LRM scales from 7B to 32B (e.g., R1-Distill-Qwen-7B achieves a +8.7 correlation point improvement). These findings highlight the potential of efficiently calibrated LRMs to advance fine-grained automatic MT evaluation.

Method: Constructed a dataset driven by real-world application (product description generation) parameterized by fairness attributes intersected with gendered adjectives and product categories, creating labeled prompts for evaluation.

Result: The dataset enables identification of quality, veracity, safety, and fairness gaps in LLMs, providing a concrete evaluation resource.

Conclusion: The work contributes a proposal for LLM evaluation specifically targeting Responsible AI dimensions, paired with a practical dataset for the research community to assess fairness in context-specific applications.

Abstract: Current methods for evaluating large language models (LLMs) typically focus on high-level tasks such as text generation, without targeting a particular AI application. This approach is not sufficient for evaluating LLMs for Responsible AI dimensions like fairness, since protected attributes that are highly relevant in one application may be less relevant in another. In this work, we construct a dataset that is driven by a real-world application (generate a plain-text product description, given a list of product features), parameterized by fairness attributes intersected with gendered adjectives and product categories, yielding a rich set of labeled prompts. We show how to use the data to identify quality, veracity, safety, and fairness gaps in LLMs, contributing a proposal for LLM evaluation paired with a concrete resource for the research community.

Method: Propose hybrid models with token mixers including sparse attention with learnable token eviction, sliding-window attention, and lightweight CNN to adaptively retain critical KV-pairs while maintaining linear complexity.

Result: Empirical evaluations on retrieval-intensive benchmarks show effectiveness of the proposed approaches.

Conclusion: The hybrid models successfully mitigate forgetfulness in linear-attention models while preserving efficiency for retrieval-intensive tasks.

Abstract: Linear-attention models that compress the entire input sequence into a fixed-size recurrent state offer an efficient alternative to Transformers, but their finite memory induces forgetfulness that harms retrieval-intensive tasks. To mitigate the issue, we explore a series of hybrid models that restore direct access to past tokens. We interleave token mixers with intermediate time and space complexity between linear and full attention, including sparse attention with token eviction, and the query-aware native sparse attention. Particularly, we propose a novel learnable token eviction approach. Combined with sliding-window attention, an end-to-end trainable lightweight CNN aggregates information from both past and future adjacent tokens to adaptively retain a limited set of critical KV-pairs per head, maintaining linear attention’s constant time and space complexity. Efficient Triton kernels for the sparse attention mechanisms are provided. Empirical evaluations on retrieval-intensive benchmarks support the effectiveness of our approaches.

Method: Developed a lightweight mean-pooling approach and compared it against compression-tokens architecture, studying training for multiple compression ratios across various QA datasets and model configurations.

Result: Mean-pooling consistently outperformed compression-tokens across experiments, with small performance degradation when trained for multiple compression ratios.

Conclusion: Simple mean-pooling is the most effective compression method, though trade-offs between architectures and training regimes reveal a complex landscape for compression techniques.

Abstract: A common strategy to reduce the computational costs of using long contexts in retrieval-augmented generation (RAG) with large language models (LLMs) is soft context compression, where the input sequence is transformed into a shorter continuous representation. We develop a lightweight and simple mean-pooling approach that consistently outperforms the widely used compression-tokens architecture, and study training the same compressor to output multiple compression ratios. We conduct extensive experiments across in-domain and out-of-domain QA datasets, as well as across model families, scales, and compression ratios. Overall, our simple mean-pooling approach achieves the strongest performance, with a relatively small drop when training for multiple compression ratios. More broadly though, across architectures and training regimes the trade-offs are more nuanced, illustrating the complex landscape of compression methods.

Method: Analysis of existing detection approaches and their limitations in handling diverse LLM usage scenarios, human edits, and subtle LLM influences on users.

Result: Current detectors provide limited value and their numerical results are often misunderstood; they remain useful only under specific conditions.

Conclusion: LLM-generated text detectors should be interpreted as references rather than decisive indicators, as their significance diminishes due to real-world complexities.

Abstract: With the widespread use of large language models (LLMs), many researchers have turned their attention to detecting text generated by them. However, there is no consistent or precise definition of their target, namely “LLM-generated text”. Differences in usage scenarios and the diversity of LLMs further increase the difficulty of detection. What is commonly regarded as the detecting target usually represents only a subset of the text that LLMs can potentially produce. Human edits to LLM outputs, together with the subtle influences that LLMs exert on their users, are blurring the line between LLM-generated and human-written text. Existing benchmarks and evaluation approaches do not adequately address the various conditions in real-world detector applications. Hence, the numerical results of detectors are often misunderstood, and their significance is diminishing. Therefore, detectors remain useful under specific conditions, but their results should be interpreted only as references rather than decisive indicators.

Method: The paper introduces MVRAG framework that utilizes intention-aware query rewriting from multiple domain viewpoints to enhance retrieval precision.

Result: Experiments on legal and medical case retrieval demonstrate significant improvements in recall and precision rates with the proposed framework.

Conclusion: The multi-perspective retrieval approach unleashes the potential of multi-view information for enhancing RAG tasks, accelerating LLM applications in knowledge-intensive fields.

Abstract: While Retrieval-Augmented Generation (RAG) plays a crucial role in the application of Large Language Models (LLMs), existing retrieval methods in knowledge-dense domains like law and medicine still suffer from a lack of multi-perspective views, which are essential for improving interpretability and reliability. Previous research on multi-view retrieval often focused solely on different semantic forms of queries, neglecting the expression of specific domain knowledge perspectives. This paper introduces a novel multi-view RAG framework, MVRAG, tailored for knowledge-dense domains that utilizes intention-aware query rewriting from multiple domain viewpoints to enhance retrieval precision, thereby improving the effectiveness of the final inference. Experiments conducted on legal and medical case retrieval demonstrate significant improvements in recall and precision rates with our framework. Our multi-perspective retrieval approach unleashes the potential of multi-view information enhancing RAG tasks, accelerating the further application of LLMs in knowledge-intensive fields.

Method: Proposed AGKA approach that uses GPT 4.0 to retrieve label definition knowledge from annotation guidelines and applies random under-sampling to select few typical examples. Conducted systematic evaluation on six LEC datasets covering behavior, emotion, and cognition classification.

Result: AGKA enhanced non-fine-tuned LLMs, especially GPT 4.0 and Llama 3 70B. GPT 4.0 with AGKA few-shot outperformed fine-tuned BERT and RoBERTa on binary classification datasets. However, GPT 4.0 lagged in multi-class tasks requiring deep semantic understanding. Llama 3 70B with AGKA performed comparably to GPT 4.0.

Conclusion: AGKA effectively improves LLMs for LEC tasks, with GPT 4.0 excelling in binary classification but struggling with complex multi-class tasks. Llama 3 70B with AGKA shows promise as an open-source alternative to closed-source models. LLMs have difficulty distinguishing similar labels in multi-class classification.

Abstract: Various machine learning approaches have gained significant popularity for the automated classification of educational text to identify indicators of learning engagement – i.e. learning engagement classification (LEC). LEC can offer comprehensive insights into human learning processes, attracting significant interest from diverse research communities, including Natural Language Processing (NLP), Learning Analytics, and Educational Data Mining. Recently, Large Language Models (LLMs), such as ChatGPT, have demonstrated remarkable performance in various NLP tasks. However, their comprehensive evaluation and improvement approaches in LEC tasks have not been thoroughly investigated. In this study, we propose the Annotation Guidelines-based Knowledge Augmentation (AGKA) approach to improve LLMs. AGKA employs GPT 4.0 to retrieve label definition knowledge from annotation guidelines, and then applies the random under-sampler to select a few typical examples. Subsequently, we conduct a systematic evaluation benchmark of LEC, which includes six LEC datasets covering behavior classification (question and urgency level), emotion classification (binary and epistemic emotion), and cognition classification (opinion and cognitive presence). The study results demonstrate that AGKA can enhance non-fine-tuned LLMs, particularly GPT 4.0 and Llama 3 70B. GPT 4.0 with AGKA few-shot outperforms full-shot fine-tuned models such as BERT and RoBERTa on simple binary classification datasets. However, GPT 4.0 lags in multi-class tasks that require a deep understanding of complex semantic information. Notably, Llama 3 70B with AGKA is a promising combination based on open-source LLM, because its performance is on par with closed-source GPT 4.0 with AGKA. In addition, LLMs struggle to distinguish between labels with similar names in multi-class classification.

Method: Uses inference-time activation contrasting to locate safety neurons and dynamic activation patching to evaluate their causal effects on model safety.

Result: Consistently identified about 5% safety neurons across multiple LLMs; patching these neurons restored over 90% safety performance across red-teaming benchmarks without impacting general ability.

Conclusion: Safety neurons explain the ‘alignment tax’ phenomenon and enable applications like detecting unsafe outputs before generation, providing insights for safer LLM development.

Abstract: Large language models (LLMs) excel in various capabilities but pose safety risks such as generating harmful content and misinformation, even after safety alignment. In this paper, we explore the inner mechanisms of safety alignment through the lens of mechanistic interpretability, focusing on identifying and analyzing safety neurons within LLMs that are responsible for safety behaviors. We propose inference-time activation contrasting to locate these neurons and dynamic activation patching to evaluate their causal effects on model safety. Experiments on multiple prevalent LLMs demonstrate that we can consistently identify about $5%$ safety neurons, and by only patching their activations we can restore over $90%$ of the safety performance across various red-teaming benchmarks without influencing general ability. The finding of safety neurons also helps explain the ‘‘alignment tax’’ phenomenon by revealing that the key neurons for model safety and helpfulness significantly overlap, yet they require different activation patterns for the same neurons. Furthermore, we demonstrate an application of our findings in safeguarding LLMs by detecting unsafe outputs before generation. The source code is available at https://github.com/THU-KEG/SafetyNeuron.

Method: Constructed ViANLI using adversarial human-and-machine-in-the-loop approach with rigorous verification. Proposed NLIMoE model with expert subnetworks and dynamic routing on transformer encoder.

Result: ViANLI contains 10,000+ premise-hypothesis pairs, challenging SOTA models. NLIMoE achieves 47.3% accuracy vs 45.5% for XLM-R Large. Training with ViANLI improves performance on other Vietnamese NLI benchmarks.

Conclusion: ViANLI enhances research into model robustness and enriches resources for Vietnamese and multilingual NLI research.

Abstract: Existing Vietnamese Natural Language Inference (NLI) datasets lack adversarial complexity, limiting their ability to evaluate model robustness against challenging linguistic phenomena. In this article, we address the gap in robust Vietnamese NLI resources by introducing ViANLI, the first adversarial NLI dataset for Vietnamese, and propose NLIMoE, a Mixture-of-Experts model to tackle its complexity. We construct ViANLI using an adversarial human-and-machine-in-the-loop approach with rigorous verification. NLIMoE integrates expert subnetworks with a learned dynamic routing mechanism on top of a shared transformer encoder. ViANLI comprises over 10,000 premise-hypothesis pairs and challenges state-of-the-art models, with XLM-R Large achieving only 45.5% accuracy, while NLIMoE reaches 47.3%. Training with ViANLI improves performance on other benchmark Vietnamese NLI datasets including ViNLI, VLSP2021-NLI, and VnNewsNLI. ViANLI is released for enhancing research into model robustness and enriching resources for future Vietnamese and multilingual NLI research.

Method: Proposed Permutative Preference Alignment (PPA) using NDCG ranking metric with differentiable surrogate loss for end-to-end alignment.

Result: PPA outperforms existing pairwise and listwise methods on evaluation sets and AlpacaEval benchmark, improving ranking accuracy more effectively than B-T-based methods.

Conclusion: NDCG-based approaches provide superior ranking performance for LLM alignment compared to traditional Bradley-Terry model based methods.

Abstract: Aligning Large Language Models (LLMs) with human preferences is crucial in ensuring desirable and controllable model behaviors. Current methods, such as Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO), rely on the Bradley-Terry (B-T) model to maximize the likelihood of pairwise choices. However, when multiple responses are available, the B-T model fails to guarantee an accurate list ranking of the responses. To address this issue, we propose Permutative Preference Alignment (PPA), a novel offline listwise approach that incorporates the Normalized Discounted Cumulative Gain (NDCG), a widely-used ranking metric, as an alternative training objective for LLM alignment. We develop an end-to-end alignment algorithm by approximating NDCG with a differentiable surrogate loss. Experiments demonstrate that PPA outperforms existing pairwise and listwise methods on evaluation sets and general benchmarks such as AlpacaEval. Furthermore, we show that NDCG-based approaches improve ranking accuracy more effectively than B-T-based methods and provide a theoretical explanation for this improvement.

Method: An iterative self-tuning process that crafts adversarial LLMs with enhanced jailbreak ability to generate effective adversarial suffixes.

Result: Achieves nearly 100% ASR on open-source LLMs and strong transferability to closed-source models (99% on GPT-3.5, 49% on GPT-4) while significantly reducing computational costs.

Conclusion: ADV-LLM provides an efficient jailbreak framework and generates valuable datasets for future safety alignment research.

Abstract: Recent research has shown that Large Language Models (LLMs) are vulnerable to automated jailbreak attacks, where adversarial suffixes crafted by algorithms appended to harmful queries bypass safety alignment and trigger unintended responses. Current methods for generating these suffixes are computationally expensive and have low Attack Success Rates (ASR), especially against well-aligned models like Llama2 and Llama3. To overcome these limitations, we introduce ADV-LLM, an iterative self-tuning process that crafts adversarial LLMs with enhanced jailbreak ability. Our framework significantly reduces the computational cost of generating adversarial suffixes while achieving nearly 100% ASR on various open-source LLMs. Moreover, it exhibits strong attack transferability to closed-source models, achieving 99% ASR on GPT-3.5 and 49% ASR on GPT-4, despite being optimized solely on Llama3. Beyond improving jailbreak ability, ADV-LLM provides valuable insights for future safety alignment research through its ability to generate large datasets for studying LLM safety. Our code is available at: https://github.com/SunChungEn/ADV-LLM

Method: HeadKV and HeadKV-R2 perform head-level KV cache compression by estimating individual attention heads’ importance for contextual QA tasks requiring both retrieval and reasoning. The method operates at the head level rather than layer level.

Result: Extensive experiments show HeadKV significantly outperforms baselines, especially in low-resource settings (KV size = 64 & 128). It retains only 1.5% of KV cache while achieving 97% performance of full KV cache on contextual QA benchmarks across diverse models and datasets.

Conclusion: Head-level KV cache compression is more effective than layer-level approaches, demonstrating that selective head retention based on contextual reasoning ability estimation can dramatically reduce memory overhead while maintaining performance.

Abstract: Key-Value (KV) caching is a common technique to enhance the computational efficiency of Large Language Models (LLMs), but its memory overhead grows rapidly with input length. Prior work has shown that not all tokens are equally important for text generation, proposing layer-level KV cache compression to selectively retain key information. Recognizing the distinct roles of attention heads in generation, we propose HeadKV, a head-level KV cache compression method, and HeadKV-R2, which leverages a novel contextual reasoning ability estimation for compression. Our approach operates at the level of individual heads, estimating their importance for contextual QA tasks that require both retrieval and reasoning capabilities. Extensive experiments across diverse benchmarks (LongBench, LooGLE), model architectures (e.g., Llama-3-8B-Instruct, Mistral-7B-Instruct), and long-context abilities tests demonstrate that our head-level KV cache compression significantly outperforms strong baselines, particularly in low-resource settings (KV size = 64 & 128). Notably, our method retains just 1.5% of the KV cache while achieving 97% of the performance of the full KV cache on the contextual question answering benchmark. Codes are available at https://github.com/FYYFU/HeadKV

Method: Proposed Bi-Mamba architecture with 1-bit quantization, trained from scratch using autoregressive distillation loss on standard LLM datasets.

Result: Achieves performance comparable to full-precision Mamba models, outperforms post-training binarization Mamba and binarization-aware training Transformers, while significantly reducing memory and computational requirements.

Conclusion: Pioneers linear-complexity LLMs under low-bit representation and enables specialized hardware optimization for efficient 1-bit Mamba models.

Abstract: The typical Selective State-Space Model (SSM) used in Mamba addresses several limitations of Transformers, such as the quadratic computational complexity with respect to sequence length and the significant memory requirements during inference due to the key-value (KV) cache. However, the increasing size of Mamba models continues to pose challenges for training and deployment, particularly due to their substantial computational demands during both training and inference. In this work, we introduce $\texttt{Bi-Mamba}$, a scalable and powerful 1-bit Mamba architecture designed to enable more efficient large language models (LLMs), with model sizes of 780M, 1.3B, and 2.7B parameters. $\texttt{Bi-Mamba}$ models are trained from scratch on a standard LLM-scale dataset using an autoregressive distillation loss. Extensive experiments on language modeling benchmarks demonstrate that $\texttt{Bi-Mamba}$ achieves performance comparable to its full-precision (FP16 or BF16) counterparts, while outperforming post-training binarization (PTB) Mamba and binarization-aware training (BAT) Transformer baselines. Moreover, $\texttt{Bi-Mamba}$ drastically reduces memory usage and computational cost compared to the original Mamba. Our work pioneers a new line of linear-complexity LLMs under low-bit representation and provides the way for the design of specialized hardware optimized for efficient 1-bit Mamba-based models. Code and the pre-trained weights are available at https://github.com/Tangshengku/Bi-Mamba.

Method: Used transformer models (RoBERTa, T5, Llama-3, MatSciBERT) with custom regression heads, finetuned on two text formats: structured template strings and natural language narratives generated via ChatGPT API.

Result: All models achieved strong accuracy with MAEs 0.25-0.33 eV. Llama-3 (1.2B parameters) performed best (MAE 0.248 eV, R2 0.891). MatSciBERT (110M parameters) reached comparable performance (MAE 0.288 eV, R2 0.871). Attention analysis showed focus on compositional/spin features over geometric features.

Conclusion: Pretrained language models can effectively extract complex feature-property relationships from textual material descriptions, with domain-specific pretraining being particularly valuable for achieving high performance with fewer parameters.

Abstract: We investigate transformer-based language models, including RoBERTa, T5, Llama-3, and MatSciBERT, for predicting the band gaps of semiconductor materials directly from textual descriptions. The inputs encode key material features, such as chemical composition, crystal system, space group, and other structural and electronic properties. Unlike shallow machine learning models, which require extensive feature engineering, or Graph Neural Networks, which rely on graph representations derived from atomic coordinates, pretrained language models can process textual inputs directly, eliminating the need for manual feature preprocessing or structure-based encoding. Material descriptions were constructed in two formats: structured strings with a consistent template and natural language narratives generated via the ChatGPT API. Each model was augmented with a custom regression head and finetuned for band gap prediction task. Language models of different architectures and parameter sizes were all able to predict band gaps from human-readable text with strong accuracy, achieving MAEs in the range of 0.25-0.33 eV, highlighting the success of this approach for scientific regression tasks. Finetuned Llama-3, with 1.2 billion parameters, achieved the highest accuracy (MAE 0.248 eV, R2 0.891). MatSciBERT, pretrained on materials science literature, reached comparable performance (MAE 0.288 eV, R2 0.871) with significantly fewer parameters (110 million), emphasizing the importance of domain-specific pretraining. Attention analysis shows that both models selectively focus on compositional and spin-related features while de-emphasizing geometric features, reflecting the difficulty of capturing spatial information from text. These results establish that pretrained language models can effectively extract complex feature-property relationships from textual material descriptions.

Method: Conducted comprehensive analysis of various attention mechanisms (RoPE, NoPE, QK-Norm), identified their attention patterns, and proposed a novel hybrid attention architecture integrating global and local attention spans.

Result: The proposed hybrid attention mechanism surpasses conventional RoPE-based transformer models in both long and short context tasks while delivering substantial efficiency gains during training and inference.

Conclusion: The analysis provides valuable insights for architectural design in long-context modeling, and the hybrid attention approach offers superior performance and efficiency compared to existing methods.

Abstract: Long-context large language models (LLMs) have achieved remarkable advancements, driven by techniques like Rotary Position Embedding (RoPE) (Su et al., 2023) and its extensions (Chen et al., 2023; Liu et al., 2024c; Peng et al., 2023). By adjusting RoPE parameters and incorporating training data with extended contexts, we can train performant models with considerably longer input sequences. However, existing RoPE-based methods exhibit performance limitations when applied to extended context lengths. This paper presents a comprehensive analysis of various attention mechanisms, including RoPE, No Positional Embedding (NoPE), and Query-Key Normalization (QK-Norm), identifying their strengths and shortcomings in long-context modeling. Our investigation identifies distinctive attention patterns in these methods and highlights their impact on long-context performance, providing valuable insights for architectural design. Building on these findings, we propose a novel architecture featuring a hybrid attention mechanism that integrates global and local attention spans. This design not only surpasses conventional RoPE-based transformer models with full attention in both long and short context tasks but also delivers substantial efficiency gains during training and inference.

Method: Analyze model internals to discover attention heads encoding sufficiency signals, then use lightweight classifiers to detect when critical information has been processed for dynamic context cutoff.

Result: 3.4% accuracy improvement with 1.33x token reduction on average across six QA datasets with three model families (LLaMA/Qwen/Mistral, 1B-70B), superior to other context efficiency methods at equivalent token reduction rates.

Conclusion: Models’ internal understanding naturally dictates processing needs rather than external compression heuristics, with larger models exhibiting intrinsic self-assessment capabilities through prompting.

Abstract: Large language models (LLMs) process entire input contexts indiscriminately, which is inefficient when the information required to answer a query is localized within the context. We present dynamic context cutoff, a novel method enabling LLMs to self-terminate processing upon acquiring sufficient task-relevant information. Through analysis of model internals, we discover that specific attention heads inherently encode “sufficiency signals” – detectable through lightweight classifiers – that predict when critical information has been processed. This reveals a new efficiency paradigm: models’ internal understanding naturally dictates processing needs rather than external compression heuristics. Comprehensive experiments across six QA datasets (up to 40K tokens) with three model families (LLaMA/Qwen/Mistral, 1B-70B) demonstrate 3.4% accuracy improvement while achieving 1.33x token reduction on average. Furthermore, our method demonstrates superior performance compared to other context efficiency methods at equivalent token reduction rates. Additionally, we observe an emergent scaling phenomenon: while smaller models require probing for sufficiency detection, larger models exhibit intrinsic self-assessment capabilities through prompting.

Method: Two-step iterative optimization: (1) Role-step: learns directed acyclic graphs (DAGs) representing model roles and message flow using particle swarm optimization on continuous adjacency matrices; (2) Weight-step: optimizes model weights using JFK-score to quantify individual contributions and particle swarm optimization.

Result: Outperforms 15 role- and weight-based baselines by 18.5% on average across 12 tasks, discovers systems with heterogeneous model roles and substantial collaborative gains, and benefits from model diversity.

Conclusion: Heterogeneous Swarms effectively designs multi-LLM systems through joint optimization of roles and weights, demonstrating significant performance improvements and the value of heterogeneous collaboration in language model ensembles.

Abstract: We propose Heterogeneous Swarms, an algorithm to design multi-LLM systems by jointly optimizing model roles and weights. We represent multi-LLM systems as directed acyclic graphs (DAGs) of LLMs with topological message passing for collaborative generation. Given a pool of LLM experts and a utility function, Heterogeneous Swarms employs two iterative steps: role-step and weight-step. For role-step, we interpret model roles as learning a DAG that specifies the flow of inputs and outputs between LLMs. Starting from a swarm of random continuous adjacency matrices, we decode them into discrete DAGs, call the LLMs in topological order, evaluate on the utility function (e.g. accuracy on a task), and optimize the adjacency matrices with particle swarm optimization based on the utility score. For weight-step, we assess the contribution of individual LLMs in the multi-LLM systems and optimize model weights with swarm intelligence. We propose JFK-score to quantify the individual contribution of each LLM in the best-found DAG of the role-step, then optimize model weights with particle swarm optimization based on the JFK-score. Experiments demonstrate that Heterogeneous Swarms outperforms 15 role- and/or weight-based baselines by 18.5% on average across 12 tasks. Further analysis reveals that Heterogeneous Swarms discovers multi-LLM systems with heterogeneous model roles and substantial collaborative gains, and benefits from the diversity of language models.

Method: Designs a search space with three token types (Global, Local, Sliding Window) and uses a learnable attention mask to jointly learn token-type information with architecture weights, similar to One-Shot Neural Architecture Search.

Result: Experiments show NAtS can efficiently reduce KV cache size while maintaining model performance, both when training new transformers from scratch and fine-tuning existing large language models.

Conclusion: NAtS provides an effective approach for reducing inference costs in transformer models by intelligently managing token retention through automated attention search.

Abstract: We present Neural Attention Search (NAtS), a framework that automatically evaluates the importance of each token within a sequence and determines if the corresponding token can be dropped after several steps. This approach can efficiently reduce the KV cache sizes required by transformer-based models during inference and thus reduce inference costs. In this paper, we design a search space that contains three token types: (i) Global Tokens will be preserved and queried by all the following tokens. (ii) Local Tokens survive until the next global token appears. (iii) Sliding Window Tokens have an impact on the inference of a fixed size of the next following tokens. Similar to the One-Shot Neural Architecture Search approach, this token-type information can be learned jointly with the architecture weights via a learnable attention mask. Experiments on both training a new transformer from scratch and fine-tuning existing large language models show that NAtS can efficiently reduce the KV cache size required for the models while maintaining the models’ performance.

Method: Developed U-SafeBench benchmark dataset to evaluate user-specific safety of LLMs, and proposed a chain-of-thought based remedy to improve safety.

Result: Evaluation of 20 widely used LLMs shows they fail to act safely when considering user-specific safety standards, demonstrating a new vulnerability.

Conclusion: User-specific safety is a critical aspect of LLM safety that current models struggle with, but can be improved through chain-of-thought approaches.

Abstract: As the use of large language model (LLM) agents continues to grow, their safety vulnerabilities have become increasingly evident. Extensive benchmarks evaluate various aspects of LLM safety by defining the safety relying heavily on general standards, overlooking user-specific standards. However, safety standards for LLM may vary based on a user-specific profiles rather than being universally consistent across all users. This raises a critical research question: Do LLM agents act safely when considering user-specific safety standards? Despite its importance for safe LLM use, no benchmark datasets currently exist to evaluate the user-specific safety of LLMs. To address this gap, we introduce U-SafeBench, a benchmark designed to assess user-specific aspect of LLM safety. Our evaluation of 20 widely used LLMs reveals current LLMs fail to act safely when considering user-specific safety standards, marking a new discovery in this field. To address this vulnerability, we propose a simple remedy based on chain-of-thought, demonstrating its effectiveness in improving user-specific safety. Our benchmark and code are available at https://github.com/yeonjun-in/U-SafeBench.

Method: Used the “finding experts” method from activation steering research to identify neurons responsible for specific concepts, then analyzed these neurons through ExpertLens to reconstruct human concept organization.

Result: ExpertLens representations are stable across models and datasets, closely align with human representations (matching inter-human alignment levels), and significantly outperform word/sentence embeddings in capturing concept alignment.

Conclusion: ExpertLens is a flexible and lightweight approach for capturing and analyzing model representations, providing granular insights into LLM concept organization.

Abstract: Activation steering methods in large language models (LLMs) have emerged as an effective way to perform targeted updates to enhance generated language without requiring large amounts of adaptation data. We ask whether the features discovered by activation steering methods are interpretable. We identify neurons responsible for specific concepts (e.g., cat'') using the finding experts’’ method from research on activation steering and show that the ExpertLens, i.e., inspection of these neurons provides insights about model representation. We find that ExpertLens representations are stable across models and datasets and closely align with human representations inferred from behavioral data, matching inter-human alignment levels. ExpertLens significantly outperforms the alignment captured by word/sentence embeddings. By reconstructing human concept organization through ExpertLens, we show that it enables a granular view of LLM concept representation. Our findings suggest that ExpertLens is a flexible and lightweight approach for capturing and analyzing model representations.

Method: Evaluated various LLMs on custom datasets from multi-hop QA tasks, keeping context length and relevant information position constant while varying document count.

Result: Most LLMs showed up to 20% performance degradation with increasing document count, while Qwen2.5 maintained consistent results across different document counts.

Conclusion: Processing multiple documents is a separate challenge from handling long contexts, and Qwen2.5 demonstrates superior multi-document handling capability compared to other models.

Abstract: Retrieval-Augmented Generation (RAG) enhances the accuracy of Large Language Model (LLM) responses by leveraging relevant external documents during generation. Although previous studies noted that retrieving many documents can degrade performance, they did not isolate how the quantity of documents affects performance while controlling for context length. We evaluate various language models on custom datasets derived from a multi-hop QA task. We keep the context length and position of relevant information constant while varying the number of documents, and find that increasing the document count in RAG settings poses significant challenges for most LLMs, reducing performance by up to 20%. However, Qwen2.5 maintained consistent results across increasing document counts, indicating better multi-document handling capability. Finally, our results indicate that processing multiple documents is a separate challenge from handling long contexts. We also make the datasets and code available: https://github.com/shaharl6000/MoreDocsSameLen .

Method: Developed a novel statistical test assuming smooth fiber bundle structure as null hypothesis. Test identifies irregularities in token neighborhoods by rejecting the null when embeddings don’t follow expected smooth structure.

Result: The test frequently rejected the null hypothesis across multiple open-source LLMs, indicating token embedding spaces are not smooth fiber bundles or manifolds. This affects model stability when semantically equivalent prompts contain implicated tokens.

Conclusion: Token embedding spaces in LLMs don’t conform to smooth manifold assumptions, which impacts model behavior and stability, particularly when processing semantically equivalent prompts containing irregular tokens.

Abstract: A full understanding of the behavior of a large language model (LLM) requires our grasp of its input token space. If this space differs from our assumptions, our comprehension of and conclusions about the LLM will likely be flawed. We elucidate the structure of the token embeddings both empirically and theoretically. We present a novel statistical test assuming that the neighborhood around each token has a relatively flat and smooth structure as the null hypothesis. Failing to reject the null is uninformative, but rejecting it at a specific token $\psi$ implies an irregularity in the token subspace in a $\psi$-neighborhood, $B(\psi)$. The structure assumed in the null is a generalization of a manifold with boundary called a \emph{smooth fiber bundle} (which can be split into two spatial regimes – small and large radius), so we denote our new hypothesis test as the ``fiber bundle hypothesis.’’ By running our test over several open-source LLMs, each with unique token embeddings, we find that the null is frequently rejected, and so the evidence suggests that the token subspace is not a fiber bundle and hence also not a manifold. As a consequence of our findings, when an LLM is presented with two semantically equivalent prompts, if one prompt contains a token implicated by our test, the response to that prompt will likely exhibit less stability than the other.

Method: Proposed BiGTex with stacked Graph-Text Fusion Units enabling mutual attention between textual and structural representations. Uses parameter-efficient fine-tuning (LoRA) with frozen LLM while adapting to task-specific signals.

Result: Extensive experiments on five benchmark datasets show BiGTex achieves state-of-the-art performance in node classification and effectively generalizes to link prediction.

Conclusion: Ablation study confirms the importance of soft prompting and bi-directional attention in the model’s success, demonstrating effective integration of GNNs and LLMs for text-attributed graphs.

Abstract: Text-attributed graphs (TAGs) present unique challenges in representation learning by requiring models to capture both the semantic richness of node-associated texts and the structural dependencies of the graph. While graph neural networks (GNNs) excel at modeling topological information, they lack the capacity to process unstructured text. Conversely, large language models (LLMs) are proficient in text understanding but are typically unaware of graph structure. In this work, we propose BiGTex (Bidirectional Graph Text), a novel architecture that tightly integrates GNNs and LLMs through stacked Graph-Text Fusion Units. Each unit allows for mutual attention between textual and structural representations, enabling information to flow in both directions, text influencing structure and structure guiding textual interpretation. The proposed architecture is trained using parameter-efficient fine-tuning (LoRA), keeping the LLM frozen while adapting to task-specific signals. Extensive experiments on five benchmark datasets demonstrate that BiGTex achieves state-of-the-art performance in node classification and generalizes effectively to link prediction. An ablation study further highlights the importance of soft prompting and bi-directional attention in the model’s success.

Method: Introduces two metrics to estimate question difficulty, adaptive length-based rewards, and difficulty-aware KL divergence integrated into GRPO algorithm to guide fast or slow thinking.

Result: Achieves state-of-the-art accuracy with over 10% relative improvement compared to base model, while significantly reducing token usage across seven reasoning benchmarks.

Conclusion: FAST-GRPO effectively balances reasoning length and accuracy by dynamically adapting reasoning depth based on question characteristics.

Abstract: When applying reinforcement learning–typically through GRPO–to large vision-language model reasoning struggles to effectively scale reasoning length or generates verbose outputs across all tasks with only marginal gains in accuracy. To address this issue, we present FAST-GRPO, a variant of GRPO that dynamically adapts reasoning depth based on question characteristics. Through empirical analysis, we establish the feasibility of fast-slow thinking in LVLMs by investigating how response length and data distribution affect performance. Inspired by these observations, we introduce two complementary metrics to estimate the difficulty of the questions, guiding the model to determine when fast or slow thinking is more appropriate. Next, we incorporate adaptive length-based rewards and difficulty-aware KL divergence into the GRPO algorithm. Experiments across seven reasoning benchmarks demonstrate that FAST achieves state-of-the-art accuracy with over 10% relative improvement compared to the base model, while reducing token usage by 32.7-67.3% compared to previous slow-thinking approaches, effectively balancing reasoning length and accuracy.

Method: Leverages large language models and advanced multilingual embeddings to compute semantic distance as a language-agnostic proxy for divergent thinking.

Result: Evaluated across 11 diverse languages showing robust and consistent scoring, with convergent validity with other DT measures and correct discriminant validity with convergent thinking.

Conclusion: S-DAT enables more inclusive, global-scale creativity research and provides fairer evaluation of cognitive flexibility across diverse populations.

Abstract: This paper introduces S-DAT (Synthetic-Divergent Association Task), a scalable, multilingual framework for automated assessment of divergent thinking (DT) -a core component of human creativity. Traditional creativity assessments are often labor-intensive, language-specific, and reliant on subjective human ratings, limiting their scalability and cross-cultural applicability. In contrast, S-DAT leverages large language models and advanced multilingual embeddings to compute semantic distance – a language-agnostic proxy for DT. We evaluate S-DAT across eleven diverse languages, including English, Spanish, German, Russian, Hindi, and Japanese (Kanji, Hiragana, Katakana), demonstrating robust and consistent scoring across linguistic contexts. Unlike prior DAT approaches, the S-DAT shows convergent validity with other DT measures and correct discriminant validity with convergent thinking. This cross-linguistic flexibility allows for more inclusive, global-scale creativity research, addressing key limitations of earlier approaches. S-DAT provides a powerful tool for fairer, more comprehensive evaluation of cognitive flexibility in diverse populations and can be freely assessed online: https://sdat.iol.zib.de/.

Method: Proposed a human-AI collaboration framework with criteria-guided intent alignment and context-aware modeling, curated a dataset of 7,000 research papers with 140,000 instruction-response pairs, and developed XtraGPT as open-source LLMs (1.5B to 14B parameters) for context-aware, instruction-guided writing assistance.

Result: XtraGPT significantly outperforms same-scale baselines and approaches the quality of proprietary systems in both automated preference assessments and human evaluations, effectively improving scientific drafts.

Conclusion: The proposed framework and XtraGPT system successfully address the limitations of current LLM systems for academic writing by supporting context-aware, instruction-guided revisions that improve scientific draft quality.

Abstract: Despite the growing adoption of large language models (LLMs) in academic workflows, their capabilities remain limited to support high-quality scientific writing. Most existing systems are designed for general-purpose scientific text generation and fail to meet the sophisticated demands of research communication beyond surface-level polishing, such as conceptual coherence across sections. Furthermore, academic writing is inherently iterative and revision-driven, a process not well supported by direct prompting-based paradigms. To address these scenarios, we propose a human-AI collaboration framework for academic paper revision centered on criteria-guided intent alignment and context-aware modeling. To validate the framework, we curate a dataset of 7,000 research papers from top-tier venues annotated with 140,000 instruction-response pairs that reflect realistic, section-level scientific revisions. We instantiate the framework in XtraGPT, the first suite of open-source LLMs (1.5B to 14B parameters) for context-aware, instruction-guided writing assistance. Extensive experiments validate that XtraGPT significantly outperforms same-scale baselines and approaches the quality of proprietary systems. Both automated preference assessments and human evaluations confirm the effectiveness of XtraGPT in improving scientific drafts.

Method: Mined 140k KG-paths from open-domain KGs, curated a high-quality subset of 25.9k paths, and developed a KG-based multilingual, multihop benchmark called MultiHal for generative text evaluation.

Result: Baseline evaluation showed absolute improvements: 0.12-0.36 points for semantic similarity, 0.16-0.36 for NLI entailment, and 0.29-0.42 for hallucination detection in KG-RAG over vanilla QA across multiple languages and models.

Conclusion: MultiHal benchmark demonstrates the potential of KG integration for hallucination mitigation and is expected to foster future research in graph-based fact-checking tasks.

Abstract: Large Language Models (LLMs) have inherent limitations of faithfulness and factuality, commonly referred to as hallucinations. Several benchmarks have been developed that provide a test bed for factuality evaluation within the context of English-centric datasets, while relying on supplementary informative context like web links or text passages but ignoring the available structured factual resources. To this end, Knowledge Graphs (KGs) have been identified as a useful aid for hallucination mitigation, as they provide a structured way to represent the facts about entities and their relations with minimal linguistic overhead. We bridge the lack of KG paths and multilinguality for factual language modeling within the existing hallucination evaluation benchmarks and propose a KG-based multilingual, multihop benchmark called MultiHal framed for generative text evaluation. As part of our data collection pipeline, we mined 140k KG-paths from open-domain KGs, from which we pruned noisy KG-paths, curating a high-quality subset of 25.9k. Our baseline evaluation shows an absolute scale improvement by approximately 0.12 to 0.36 points for the semantic similarity score, 0.16 to 0.36 for NLI entailment and 0.29 to 0.42 for hallucination detection in KG-RAG over vanilla QA across multiple languages and multiple models, demonstrating the potential of KG integration. We anticipate MultiHal will foster future research towards several graph-based hallucination mitigation and fact-checking tasks.

Method: MoMoE orchestrates four operators (Allocate, Predict, Aggregate, Explain) with community-specialized experts (MoMoE-Community) and norm-violation experts (MoMoE-NormVio).

Result: On 30 unseen subreddits, best variants achieved Micro-F1 scores of 0.72 and 0.67, matching or surpassing fine-tuned baselines while producing concise explanations.

Conclusion: MoMoE enables scalable, transparent moderation without per-community fine-tuning, suggesting lightweight expert ensembles can guide future trustworthy AI governance research.

Abstract: Large language models (LLMs) have shown great potential in flagging harmful content in online communities. Yet, existing approaches for moderation require a separate model for every community and are opaque in their decision-making, limiting real-world adoption. We introduce Mixture of Moderation Experts (MoMoE), a modular, cross-community framework that adds post-hoc explanations to scalable content moderation. MoMoE orchestrates four operators – Allocate, Predict, Aggregate, Explain – and is instantiated as seven community-specialized experts (MoMoE-Community) and five norm-violation experts (MoMoE-NormVio). On 30 unseen subreddits, the best variants obtain Micro-F1 scores of 0.72 and 0.67, respectively, matching or surpassing strong fine-tuned baselines while consistently producing concise and reliable explanations. Although community-specialized experts deliver the highest peak accuracy, norm-violation experts provide steadier performance across domains. These findings show that MoMoE yields scalable, transparent moderation without needing per-community fine-tuning. More broadly, they suggest that lightweight, explainable expert ensembles can guide future NLP and HCI research on trustworthy human-AI governance of online communities.

Method: Leverage sparse autoencoders to identify toxicity-related directions in model residual streams, then perform targeted activation steering using decoder vectors with three tiers of steering aggressiveness.

Result: At stronger steering strengths, the method reduces toxicity by up to 20% compared to baselines, though fluency degrades on GPT-2 Small. Standard NLP benchmarks remain stable, indicating preserved knowledge and abilities.

Conclusion: SAE-based causal interventions show promise for LLM detoxification but have current limitations. Feature-splitting in wider SAEs hampers safety interventions, highlighting the importance of disentangled feature learning.

Abstract: Large language models (LLMs) are now ubiquitous in user-facing applications, yet they still generate undesirable toxic outputs, including profanity, vulgarity, and derogatory remarks. Although numerous detoxification methods exist, most apply broad, surface-level fixes and can therefore easily be circumvented by jailbreak attacks. In this paper we leverage sparse autoencoders (SAEs) to identify toxicity-related directions in the residual stream of models and perform targeted activation steering using the corresponding decoder vectors. We introduce three tiers of steering aggressiveness and evaluate them on GPT-2 Small and Gemma-2-2B, revealing trade-offs between toxicity reduction and language fluency. At stronger steering strengths, these causal interventions surpass competitive baselines in reducing toxicity by up to 20%, though fluency can degrade noticeably on GPT-2 Small depending on the aggressiveness. Crucially, standard NLP benchmark scores upon steering remain stable, indicating that the model’s knowledge and general abilities are preserved. We further show that feature-splitting in wider SAEs hampers safety interventions, underscoring the importance of disentangled feature learning. Our findings highlight both the promise and the current limitations of SAE-based causal interventions for LLM detoxification, further suggesting practical guidelines for safer language-model deployment.

Method: Used causal mediation and abstraction analysis on a custom dataset (CausalToM) with simple stories where characters independently change object states. Identified a lookback mechanism that binds character-object-state triples using Ordering IDs in low-rank subspaces.

Result: Discovered a pervasive algorithmic pattern where LMs use lookback mechanisms to retrieve correct state information when reasoning about beliefs. When visibility is specified, the model generates visibility IDs to update character beliefs accordingly.

Conclusion: The work provides insights into belief tracking mechanisms in language models and takes a step toward reverse-engineering Theory of Mind reasoning capabilities in LMs.

Abstract: How do language models (LMs) represent characters’ beliefs, especially when those beliefs may differ from reality? This question lies at the heart of understanding the Theory of Mind (ToM) capabilities of LMs. We analyze LMs’ ability to reason about characters’ beliefs using causal mediation and abstraction. We construct a dataset, CausalToM, consisting of simple stories where two characters independently change the state of two objects, potentially unaware of each other’s actions. Our investigation uncovers a pervasive algorithmic pattern that we call a lookback mechanism, which enables the LM to recall important information when it becomes necessary. The LM binds each character-object-state triple together by co-locating their reference information, represented as Ordering IDs (OIs), in low-rank subspaces of the state token’s residual stream. When asked about a character’s beliefs regarding the state of an object, the binding lookback retrieves the correct state OI and then the answer lookback retrieves the corresponding state token. When we introduce text specifying that one character is (not) visible to the other, we find that the LM first generates a visibility ID encoding the relation between the observing and the observed character OIs. In a visibility lookback, this ID is used to retrieve information about the observed character and update the observing character’s beliefs. Our work provides insights into belief tracking mechanisms, taking a step toward reverse-engineering ToM reasoning in LMs.

Method: After generating a token, MoI constructs new input by blending the discrete token with the previously discarded token distribution using Bayesian estimation - treating distribution as prior, sampled token as observation, and using posterior expectation as input.

Result: MoI consistently improves performance on mathematical reasoning, code generation, and PhD-level QA tasks across multiple models (QwQ-32B, Nemotron-Super-49B, Gemma-3-27B, DAPO-Qwen-32B) with no additional training and negligible computational overhead.

Conclusion: Preserving token distribution information through continuous posterior expectations significantly enhances autoregressive generation quality and reasoning capabilities without requiring model retraining.

Abstract: In standard autoregressive generation, an LLM predicts the next-token distribution, samples a discrete token, and then discards the distribution, passing only the sampled token as new input. To preserve this distribution’s rich information, we propose Mixture of Inputs (MoI), a training-free method for autoregressive generation. After generating a token following the standard paradigm, we construct a new input that blends the generated discrete token with the previously discarded token distribution. Specifically, we employ a Bayesian estimation method that treats the token distribution as the prior, the sampled token as the observation, and replaces the conventional one-hot vector with the continuous posterior expectation as the new model input. MoI allows the model to maintain a richer internal representation throughout the generation process, resulting in improved text quality and reasoning capabilities. On mathematical reasoning, code generation, and PhD-level QA tasks, MoI consistently improves performance across multiple models including QwQ-32B, Nemotron-Super-49B, Gemma-3-27B, and DAPO-Qwen-32B, with no additional training and negligible computational overhead.

Method: Created MEMENTO, a two-stage evaluation framework with single-memory and joint-memory tasks. Developed a hierarchical knowledge graph-based user-profile memory module that separately manages personalized knowledge to address identified bottlenecks.

Result: Current agents can recall simple object semantics but fail to apply sequential user patterns to planning. The proposed memory architecture achieved substantial improvements on both single and joint-memory tasks.

Conclusion: Episodic memory provides both personalized knowledge and in-context learning benefits. The hierarchical knowledge graph-based approach effectively addresses information overload and coordination failures in memory utilization for personalized assistance tasks.

Abstract: LLM-powered embodied agents have shown success on conventional object-rearrangement tasks, but providing personalized assistance that leverages user-specific knowledge from past interactions presents new challenges. We investigate these challenges through the lens of agents’ memory utilization along two critical dimensions: object semantics (identifying objects based on personal meaning) and user patterns (recalling sequences from behavioral routines). To assess these capabilities, we construct MEMENTO, an end-to-end two-stage evaluation framework comprising single-memory and joint-memory tasks. Our experiments reveal that current agents can recall simple object semantics but struggle to apply sequential user patterns to planning. Through in-depth analysis, we identify two critical bottlenecks: information overload and coordination failures when handling multiple memories. Based on these findings, we explore memory architectural approaches to address these challenges. Given our observation that episodic memory provides both personalized knowledge and in-context learning benefits, we design a hierarchical knowledge graph-based user-profile memory module that separately manages personalized knowledge, achieving substantial improvements on both single and joint-memory tasks. Project website: https://connoriginal.github.io/MEMENTO

Method: Analyzed cross-lingual detoxification through 392 extensive settings to evaluate toxicity reduction in cross-distribution settings with limited data.

Result: Investigated how mitigation impacts model performance on non-toxic tasks, revealing trade-offs between safety and knowledge preservation.

Conclusion: Cross-lingual detoxification is an effective paradigm for transferring detoxification capabilities between languages, though it involves balancing safety improvements with potential knowledge loss.

Abstract: As large language models (LLMs) become increasingly prevalent in global applications, ensuring that they are toxicity-free across diverse linguistic contexts remains a critical challenge. We explore “Cross-lingual Detoxification”, a cross-lingual paradigm that mitigates toxicity, enabling detoxification capabilities to transfer between high and low-resource languages across different script families. We analyze cross-lingual detoxification’s effectiveness through 392 extensive settings to evaluate toxicity reduction in cross-distribution settings with limited data and investigate how mitigation impacts model performance on non-toxic tasks, revealing trade-offs between safety and knowledge preservation. Our code and dataset are publicly available at https://github.com/himanshubeniwal/Breaking-mBad.

Method: HRPO integrates prior hidden states into sampled tokens with learnable gating, and progressively incorporates more hidden features while maintaining token embeddings. It introduces stochasticity via token sampling for RL optimization without CoT trajectories.

Result: HRPO outperforms prior methods in knowledge- and reasoning-intensive tasks across diverse benchmarks, while maintaining interpretability and exhibiting cross-lingual patterns and shorter completion lengths.

Conclusion: HRPO demonstrates the potential of RL-based approaches for latent reasoning, offering insights for future work in combining discrete and continuous representations in LLMs.

Abstract: Recent advances in large language models (LLMs) have introduced latent reasoning as a promising alternative to autoregressive reasoning. By performing internal computation with hidden states from previous steps, latent reasoning benefit from more informative features rather than sampling a discrete chain-of-thought (CoT) path. Yet latent reasoning approaches are often incompatible with LLMs, as their continuous paradigm conflicts with the discrete nature of autoregressive generation. Moreover, these methods rely on CoT traces for training and thus fail to exploit the inherent reasoning patterns of LLMs. In this work, we explore latent reasoning by leveraging the intrinsic capabilities of LLMs via reinforcement learning (RL). To this end, we introduce hybrid reasoning policy optimization (HRPO), an RL-based hybrid latent reasoning approach that (1) integrates prior hidden states into sampled tokens with a learnable gating mechanism, and (2) initializes training with predominantly token embeddings while progressively incorporating more hidden features. This design maintains LLMs’ generative capabilities and incentivizes hybrid reasoning using both discrete and continuous representations. In addition, the hybrid HRPO introduces stochasticity into latent reasoning via token sampling, thereby enabling RL-based optimization without requiring CoT trajectories. Extensive evaluations across diverse benchmarks show that HRPO outperforms prior methods in both knowledge- and reasoning-intensive tasks. Furthermore, HRPO-trained LLMs remain interpretable and exhibit intriguing behaviors like cross-lingual patterns and shorter completion lengths, highlighting the potential of our RL-based approach and offer insights for future work in latent reasoning.

Method: Introduces a theoretical generative model where words are defined by binary semantic attributes and co-occurrence probabilities are derived from attribute-based interactions.

Result: The model analytically reproduces the emergence of linear analogy structure and accounts for key empirical properties including emergence in top eigenvectors, dimensional saturation, logarithmic enhancement, and robustness to corpus removal.

Conclusion: The attribute-based generative model successfully explains the linear analogy phenomenon in word embeddings and provides fine-grained understanding of embedding dimensions, showing robustness to noise and agreement with real data.

Abstract: Models such as Word2Vec and GloVe construct word embeddings based on the co-occurrence probability $P(i,j)$ of words $i$ and $j$ in text corpora. The resulting vectors $W_i$ not only group semantically similar words but also exhibit a striking linear analogy structure – for example, $W_{\text{king}} - W_{\text{man}} + W_{\text{woman}} \approx W_{\text{queen}}$ – whose theoretical origin remains unclear. Previous observations indicate that this analogy structure: (i) already emerges in the top eigenvectors of the matrix $M(i,j) = P(i,j)/P(i)P(j)$, (ii) strengthens and then saturates as more eigenvectors of $M (i, j)$, which controls the dimension of the embeddings, are included, (iii) is enhanced when using $\log M(i,j)$ rather than $M(i,j)$, and (iv) persists even when all word pairs involved in a specific analogy relation (e.g., king-queen, man-woman) are removed from the corpus. To explain these phenomena, we introduce a theoretical generative model in which words are defined by binary semantic attributes, and co-occurrence probabilities are derived from attribute-based interactions. This model analytically reproduces the emergence of linear analogy structure and naturally accounts for properties (i)-(iv). It can be viewed as giving fine-grained resolution into the role of each additional embedding dimension. It is robust to various forms of noise and agrees well with co-occurrence statistics measured on Wikipedia and the analogy benchmark introduced by Mikolov et al.

Method: Combines three innovations: cross-attention-based mechanism for feature injection, adaptive intermediate feature selection using attention entropy, and visual token compression to reduce draft model latency.

Result: Achieves up to 3.6x speedup over conventional decoding and significantly outperforms prior speculative decoding baselines in both inference throughput and speculative draft acceptance length across multiple VLMs including LLaVA, Pixtral, SmolVLM and Gemma3.

Conclusion: DREAM enables efficient, accurate, and parallel multimodal decoding with substantial throughput improvements, making it a promising solution for accelerating vision-language model inference.

Abstract: Speculative decoding (SD) has emerged as a powerful method for accelerating autoregressive generation in large language models (LLMs), yet its integration into vision-language models (VLMs) remains underexplored. We introduce DREAM, a novel speculative decoding framework tailored for VLMs that combines three key innovations: (1) a cross-attention-based mechanism to inject intermediate features from the target model into the draft model for improved alignment, (2) adaptive intermediate feature selection based on attention entropy to guide efficient draft model training, and (3) visual token compression to reduce draft model latency. DREAM enables efficient, accurate, and parallel multimodal decoding with significant throughput improvement. Experiments across a diverse set of recent popular VLMs, including LLaVA, Pixtral, SmolVLM and Gemma3, demonstrate up to 3.6x speedup over conventional decoding and significantly outperform prior SD baselines in both inference throughput and speculative draft acceptance length across a broad range of multimodal benchmarks. The code is publicly available at: https://github.com/SAI-Lab-NYU/DREAM.git

Method: Collected live-streamed consultations from short-video platforms, annotated by legal experts, and proposed an evaluation framework with 12 metrics across clarification capability and professional advice quality.

Result: Even state-of-the-art models like GPT-4 achieved only 39.8% recall for clarification and 59% overall score for advice quality, showing significant challenges in professional consultation scenarios.

Conclusion: The benchmark advances legal domain dialogue systems research and enables simulation of more realistic user-expert interactions, with exploration of strategies to enhance LLMs’ legal consultation abilities.

Abstract: Legal consultation is essential for safeguarding individual rights and ensuring access to justice, yet remains costly and inaccessible to many individuals due to the shortage of professionals. While recent advances in Large Language Models (LLMs) offer a promising path toward scalable, low-cost legal assistance, current systems fall short in handling the interactive and knowledge-intensive nature of real-world consultations. To address these challenges, we introduce LeCoDe, a real-world multi-turn benchmark dataset comprising 3,696 legal consultation dialogues with 110,008 dialogue turns, designed to evaluate and improve LLMs’ legal consultation capability. With LeCoDe, we innovatively collect live-streamed consultations from short-video platforms, providing authentic multi-turn legal consultation dialogues. The rigorous annotation by legal experts further enhances the dataset with professional insights and expertise. Furthermore, we propose a comprehensive evaluation framework that assesses LLMs’ consultation capabilities in terms of (1) clarification capability and (2) professional advice quality. This unified framework incorporates 12 metrics across two dimensions. Through extensive experiments on various general and domain-specific LLMs, our results reveal significant challenges in this task, with even state-of-the-art models like GPT-4 achieving only 39.8% recall for clarification and 59% overall score for advice quality, highlighting the complexity of professional consultation scenarios. Based on these findings, we further explore several strategies to enhance LLMs’ legal consultation abilities. Our benchmark contributes to advancing research in legal domain dialogue systems, particularly in simulating more real-world user-expert interactions.

Method: Proposes C-PMI calibrated decoding that jointly models visual and textual token contributions, formulated as bi-level optimization with token purification mechanism to dynamically regulate decoding by sampling text tokens relevant to images and refining image tokens relevant to generated text.

Result: Extensive experiments show significant reduction in hallucinations across various benchmarks while maintaining decoding efficiency.

Conclusion: The C-PMI approach effectively mitigates hallucinations in LVLMs by enhancing mutual dependency between visual and textual information through adaptive decoding optimization.

Abstract: Large Vision-Language Models (LVLMs) are susceptible to hallucinations, where generated responses seem semantically plausible yet exhibit little or no relevance to the input image. Previous studies reveal that this issue primarily stems from LVLMs’ over-reliance on language priors while disregarding the visual information during decoding. To alleviate this issue, we introduce a novel Conditional Pointwise Mutual Information (C-PMI) calibrated decoding strategy, which adaptively strengthens the mutual dependency between generated texts and input images to mitigate hallucinations. Unlike existing methods solely focusing on text token sampling, we propose to jointly model the contributions of visual and textual tokens to C-PMI, formulating hallucination mitigation as a bi-level optimization problem aimed at maximizing mutual information. To solve it, we design a token purification mechanism that dynamically regulates the decoding process by sampling text tokens remaining maximally relevant to the given image, while simultaneously refining image tokens most pertinent to the generated response. Extensive experiments across various benchmarks reveal that the proposed method significantly reduces hallucinations in LVLMs while preserving decoding efficiency.

Method: Train regression models to improve existing LLM judges’ scores using their rationales and scores, aligning them with human preferences in specific domains. Presents four quantitative judges for different types of absolute and relative feedback.

Result: Quantitative judges improve predictive power of existing judges through post-hoc modeling, validated empirically on four datasets using two base judges. More computationally efficient than supervised fine-tuning and more statistically efficient with limited human feedback.

Conclusion: The framework successfully aligns LLM evaluation scores with human preferences, offering an efficient and versatile approach to improve LLM-as-a-judge systems through quantitative post-hoc modeling.

Abstract: LLM-as-a-judge is a framework where a large language model (LLM) evaluates the output of another LLM. While LLMs excel at producing qualitative textual evaluations, they often struggle to predict human preferences and numeric scores. We propose quantitative LLM judges, which align evaluation scores of existing LLM judges to humans in a given domain using regression models. The models are trained to improve the score of the original judge using its rationale and score. We present four quantitative judges for different types of absolute and relative feedback, which showcases the generality and versatility of our framework. Our framework is more computationally efficient than supervised fine-tuning and can be more statistically efficient when human feedback is limited, which is expected in practice. We validate these claims empirically on four datasets using two base judges. Our experiments show that quantitative judges can improve the predictive power of existing judges through post-hoc modeling.

Method: Used Domain Adaptive Continual Pretraining (DACP) recipe across diverse foundation models and service domains to create ixi-GEN models.

Result: ixi-GEN models achieved substantial target-domain performance gains while preserving general capabilities.

Conclusion: DACP-based approach provides cost-efficient and scalable solution for enterprise-level deployment of domain-adapted small LLMs.

Abstract: The emergence of open-source large language models (LLMs) has expanded opportunities for enterprise applications; however, many organizations still lack the infrastructure to deploy and maintain large-scale models. As a result, small LLMs (sLLMs) have become a practical alternative despite inherent performance limitations. While Domain Adaptive Continual Pretraining (DACP) has been explored for domain adaptation, its utility in commercial settings remains under-examined. In this study, we validate the effectiveness of a DACP-based recipe across diverse foundation models and service domains, producing DACP-applied sLLMs (ixi-GEN). Through extensive experiments and real-world evaluations, we demonstrate that ixi-GEN models achieve substantial gains in target-domain performance while preserving general capabilities, offering a cost-efficient and scalable solution for enterprise-level deployment.

Method: Created RMTBench with custom characters (detailed backgrounds) and abstract characters (simple traits), constructed dialogues based on explicit user motivations, and implemented multi-turn dialogue simulation with LLM-based scoring.

Result: Developed a comprehensive benchmark that shifts focus from character background to user intention fulfillment, capturing complex conversation intentions between users and characters.

Conclusion: RMTBench bridges the gap between academic evaluation and practical deployment requirements, offering a more effective framework for assessing role-playing capabilities in LLMs.

Abstract: Recent advancements in Large Language Models (LLMs) have shown outstanding potential for role-playing applications. Evaluating these capabilities is becoming crucial yet remains challenging. Existing benchmarks mostly adopt a \textbf{character-centric} approach, simplify user-character interactions to isolated Q&A tasks, and fail to reflect real-world applications. To address this limitation, we introduce RMTBench, a comprehensive \textbf{user-centric} bilingual role-playing benchmark featuring 80 diverse characters and over 8,000 dialogue rounds. RMTBench includes custom characters with detailed backgrounds and abstract characters defined by simple traits, enabling evaluation across various user scenarios. Our benchmark constructs dialogues based on explicit user motivations rather than character descriptions, ensuring alignment with practical user applications. Furthermore, we construct an authentic multi-turn dialogue simulation mechanism. With carefully selected evaluation dimensions and LLM-based scoring, this mechanism captures the complex intention of conversations between the user and the character. By shifting focus from character background to user intention fulfillment, RMTBench bridges the gap between academic evaluation and practical deployment requirements, offering a more effective framework for assessing role-playing capabilities in LLMs. All code and datasets will be released soon. We release the datasets at https://huggingface.co/datasets/xiangh/RMTBENCH.

Method: Pretrain an MLP to imitate kNN retriever behavior on the entire pretraining dataset, then integrate it with Transformer decoders through probability interpolation.

Result: Achieved 17.5% and 24.1% scaling gains on WikiText-103 and Web datasets, 12.3% improvement on QA benchmarks, 5.2 points gain on general NLP tasks, reduced hallucinations by 10 points, and 2.5x faster inference than RAG.

Conclusion: Learning retrieval patterns parametrically effectively bridges efficient inference and knowledge access, offering a practical alternative to both RAG and fine-tuning approaches.

Abstract: Modern approaches to enhancing Large Language Models’ factual accuracy and knowledge utilization face a fundamental trade-off: non-parametric retrieval-augmented generation (RAG) provides flexible access to external knowledge but suffers from high inference latency and shallow integration, while parametric fine-tuning methods like LoRA risk catastrophic forgetting and degraded general capabilities. In this work, we propose MLP Memory, a lightweight parametric module that learns to internalize retrieval patterns without explicit document access. By pretraining an MLP to imitate a $k$NN retriever’s behavior on the entire pretraining dataset, we create a differentiable memory component that captures the benefits of retrieval-based knowledge access in a fully parametric form. Our architecture integrates this pretrained MLP Memory with Transformer decoders through simple probability interpolation, yielding 17.5% and 24.1% scaling gains on WikiText-103 and Web datasets, respectively. It further achieves 12.3% relative improvement on five question-answering benchmarks and 5.2 points absolute gain across nine general NLP tasks, while reducing hallucinations by up to 10 points on HaluEval. Moreover, MLP Memory delivers 2.5$\times$ faster inference than RAG with superior accuracy. Our findings show that learning retrieval patterns parametrically bridges the gap between efficient inference and effective knowledge access, offering a practical alternative to both RAG and fine-tuning approaches.

Method: Memory Decoder employs a small transformer decoder that learns to imitate the behavior of an external non-parametric retriever, then can be seamlessly integrated with any pretrained language model sharing the same tokenizer.

Result: Experimental results show Memory Decoder effectively adapts various Qwen and Llama models to three specialized domains (biomedicine, finance, law), reducing perplexity by an average of 6.17 points.

Conclusion: Memory Decoder introduces a novel paradigm centered on a specially pretrained memory component for domain-specific adaptation that can be integrated plug-and-play, consistently enhancing performance across multiple models.

Abstract: Large Language Models (LLMs) have shown strong abilities in general language tasks, yet adapting them to specific domains remains a challenge. Current method like Domain Adaptive Pretraining (DAPT) requires costly full-parameter training and suffers from catastrophic forgetting. Meanwhile, Retrieval-Augmented Generation (RAG) introduces substantial inference latency due to expensive nearest-neighbor searches and longer context. This paper introduces Memory Decoder, a plug-and-play pretrained memory that enables efficient domain adaptation without changing the original model’s parameters. Memory Decoder employs a small transformer decoder that learns to imitate the behavior of an external non-parametric retriever. Once trained, Memory Decoder can be seamlessly integrated with any pretrained language model that shares the same tokenizer, requiring no model-specific modifications. Experimental results demonstrate that Memory Decoder enables effective adaptation of various Qwen and Llama models to three distinct specialized domains: biomedicine, finance, and law, reducing perplexity by an average of 6.17 points. Overall, Memory Decoder introduces a novel paradigm centered on a specially pretrained memory component designed for domain-specific adaptation. This memory architecture can be integrated in a plug-and-play manner, consistently enhancing performance across multiple models within the target domain.

Method: Partitions responses into fixed-size blocks, selects one active block per step for stochastic masking, freezes preceding tokens, hides future ones, and computes loss only over active block to mirror blockwise decoding.

Result: Experiments on GSM8K, MATH, and MetaMathQA show consistent gains over classical SFT under equal compute or token budgets, with improvements confirmed to stem from training-inference alignment rather than masking effects.

Conclusion: Matching supervision granularity to decoding procedure is crucial for diffusion-based language models, with Blockwise SFT providing faithful training-inference alignment.

Abstract: Discrete diffusion language models have shown strong potential for text generation, yet standard supervised fine-tuning (SFT) misaligns with their semi-autoregressive inference: training randomly masks tokens across the entire response, while inference generates fixed-size blocks sequentially. This mismatch introduces noisy prefixes and leaky suffixes, biasing gradients away from the desired blockwise likelihood. We propose Blockwise SFT, which partitions responses into fixed-size blocks, selects one active block per step for stochastic masking, freezes all preceding tokens, and fully hides future ones. Loss is computed only over the active block, directly mirroring the blockwise decoding process. Experiments on GSM8K, MATH, and MetaMathQA show consistent gains over classical SFT under equal compute or token budgets. Block size consistency studies and ablations confirm that improvements stem from faithful training-inference alignment rather than incidental masking effects. Our results highlight the importance of matching supervision granularity to the decoding procedure in diffusion-based language models.

Method: Proposed Layer Fused Decoding (LFD) that combines representations from an intermediate layer with final-layer outputs. Used internal knowledge score (IKS) criterion to identify the optimal intermediate layer with lowest IKS value in latter half of layers.

Result: Experimental results across multiple benchmarks show LFD helps RAG systems more effectively surface retrieved context knowledge with minimal cost.

Conclusion: The layer-specific functional demarcation in LLMs enables effective external knowledge integration through LFD, improving RAG performance while maintaining efficiency.

Abstract: Retrieval-augmented generation (RAG) incorporates external knowledge into large language models (LLMs), improving their adaptability to downstream tasks and enabling information updates. Surprisingly, recent empirical evidence demonstrates that injecting noise into retrieved relevant documents paradoxically facilitates exploitation of external knowledge and improves generation quality. Although counterintuitive and challenging to apply in practice, this phenomenon enables granular control and rigorous analysis of how LLMs integrate external knowledge. Therefore, in this paper, we intervene on noise injection and establish a layer-specific functional demarcation within the LLM: shallow layers specialize in local context modeling, intermediate layers focus on integrating long-range external factual knowledge, and deeper layers primarily rely on parametric internal knowledge. Building on this insight, we propose Layer Fused Decoding (LFD), a simple decoding strategy that directly combines representations from an intermediate layer with final-layer decoding outputs to fully exploit the external factual knowledge. To identify the optimal intermediate layer, we introduce an internal knowledge score (IKS) criterion that selects the layer with the lowest IKS value in the latter half of layers. Experimental results across multiple benchmarks demonstrate that LFD helps RAG systems more effectively surface retrieved context knowledge with minimal cost.

Method: Automated framework that parses behavioral statements, generates targeted prompts, and uses models to judge adherence, focusing on three-way consistency between provider specifications, model outputs, and the provider’s own models as judges.

Result: Applied to 16 models from six developers across 100+ behavioral statements, finding systematic inconsistencies including compliance gaps of up to 20% across providers.

Conclusion: Foundation models show significant gaps in consistently satisfying their own developers’ behavioral specifications when judged by the developers’ own evaluator models.

Abstract: Companies that develop foundation models publish behavioral guidelines they pledge their models will follow, but it remains unclear if models actually do so. While providers such as OpenAI, Anthropic, and Google have published detailed specifications describing both desired safety constraints and qualitative traits for their models, there has been no systematic audit of adherence to these guidelines. We introduce an automated framework that audits models against their providers specifications by parsing behavioral statements, generating targeted prompts, and using models to judge adherence. Our central focus is on three way consistency between a provider specification, its model outputs, and its own models as judges; an extension of prior two way generator validator consistency. This establishes a necessary baseline: at minimum, a foundation model should consistently satisfy the developer behavioral specifications when judged by the developer evaluator models. We apply our framework to 16 models from six developers across more than 100 behavioral statements, finding systematic inconsistencies including compliance gaps of up to 20 percent across providers.

Method: Used unified benchmark with zero-, one-, and five-shot prompting across five core biomedical NLP tasks and nine expanded QA datasets, employing standardized prompts, fixed decoding parameters, and consistent inference pipelines.

Result: GPT-5 consistently outperformed GPT-4o with largest improvements on reasoning-intensive datasets (MedXpertQA, DiagnosisArena) and multimodal QA. Achieved better chemical NER and ChemProt scores but remained below domain-tuned baselines for disease NER and summarization.

Conclusion: GPT-5 approaches deployment-ready performance for biomedical QA with favorable accuracy-cost balance. Recommends tiered prompting strategy: direct prompting for cost-sensitive applications and chain-of-thought for complex scenarios, highlighting need for hybrid solutions where precision is critical.

Abstract: Biomedical literature and clinical narratives pose multifaceted challenges for natural language understanding, from precise entity extraction and document synthesis to multi-step diagnostic reasoning. This study extends a unified benchmark to evaluate GPT-5 and GPT-4o under zero-, one-, and five-shot prompting across five core biomedical NLP tasks: named entity recognition, relation extraction, multi-label document classification, summarization, and simplification, and nine expanded biomedical QA datasets covering factual knowledge, clinical reasoning, and multimodal visual understanding. Using standardized prompts, fixed decoding parameters, and consistent inference pipelines, we assessed model performance, latency, and token-normalized cost under official pricing. GPT-5 consistently outperformed GPT-4o, with the largest gains on reasoning-intensive datasets such as MedXpertQA and DiagnosisArena and stable improvements in multimodal QA. In core tasks, GPT-5 achieved better chemical NER and ChemProt scores but remained below domain-tuned baselines for disease NER and summarization. Despite producing longer outputs, GPT-5 showed comparable latency and 30 to 50 percent lower effective cost per correct prediction. Fine-grained analyses revealed improvements in diagnosis, treatment, and reasoning subtypes, whereas boundary-sensitive extraction and evidence-dense summarization remain challenging. Overall, GPT-5 approaches deployment-ready performance for biomedical QA while offering a favorable balance of accuracy, interpretability, and economic efficiency. The results support a tiered prompting strategy: direct prompting for large-scale or cost-sensitive applications, and chain-of-thought scaffolds for analytically complex or high-stakes scenarios, highlighting the continued need for hybrid solutions where precision and factual fidelity are critical.

Method: Text2Mem defines a compact operation set with JSON-based schema instances, includes parser, validator, and adapters for SQL backend or real frameworks, with unified execution contract.

Result: The design ensures safety, determinism, and portability across heterogeneous backends, establishing standardized foundation for memory control in agents.

Conclusion: Text2Mem provides the first standardized foundation for memory control in agents, with planned Text2Mem Bench for systematic evaluation.

Abstract: Large language model agents increasingly depend on memory to sustain long horizon interaction, but existing frameworks remain limited. Most expose only a few basic primitives such as encode, retrieve, and delete, while higher order operations like merge, promote, demote, split, lock, and expire are missing or inconsistently supported. Moreover, there is no formal and executable specification for memory commands, leaving scope and lifecycle rules implicit and causing unpredictable behavior across systems. We introduce Text2Mem, a unified memory operation language that provides a standardized pathway from natural language to reliable execution. Text2Mem defines a compact yet expressive operation set aligned with encoding, storage, and retrieval. Each instruction is represented as a JSON based schema instance with required fields and semantic invariants, which a parser transforms into typed operation objects with normalized parameters. A validator ensures correctness before execution, while adapters map typed objects either to a SQL prototype backend or to real memory frameworks. Model based services such as embeddings or summarization are integrated when required. All results are returned through a unified execution contract. This design ensures safety, determinism, and portability across heterogeneous backends. We also outline Text2Mem Bench, a planned benchmark that separates schema generation from backend execution to enable systematic evaluation. Together, these components establish the first standardized foundation for memory control in agents.

Method: Systematically evaluated six multilingual LLMs (0.5B-9B parameters) on MedEV dataset using zero-shot, few-shot, and dictionary-augmented prompting with Meddict lexicon, including terminology-aware cues and embedding-based example retrieval.

Result: Model scale is the primary performance driver - larger LLMs achieve strong zero-shot results, while few-shot prompting yields only marginal improvements. Terminology-aware cues and embedding-based retrieval consistently improve domain-specific translation.

Conclusion: Findings underscore both the promise and current limitations of multilingual LLMs for medical English-Vietnamese translation, highlighting the importance of model scale and domain-specific adaptation strategies.

Abstract: Medical English-Vietnamese machine translation (En-Vi MT) is essential for healthcare access and communication in Vietnam, yet Vietnamese remains a low-resource and under-studied language. We systematically evaluate prompting strategies for six multilingual LLMs (0.5B-9B parameters) on the MedEV dataset, comparing zero-shot, few-shot, and dictionary-augmented prompting with Meddict, an English-Vietnamese medical lexicon. Results show that model scale is the primary driver of performance: larger LLMs achieve strong zero-shot results, while few-shot prompting yields only marginal improvements. In contrast, terminology-aware cues and embedding-based example retrieval consistently improve domain-specific translation. These findings underscore both the promise and the current limitations of multilingual LLMs for medical En-Vi MT.

Method: Developed PersonaMatrix framework with six personas (legal and non-legal users), created a controlled dimension-shifted dataset of U.S. civil rights case summaries varying in depth, accessibility, and procedural detail, and introduced Diversity-Coverage Index (DCI).

Result: The framework reveals divergent optimization goals between persona-aware and persona-agnostic evaluation approaches, showing that different user groups require different summary characteristics.

Conclusion: PersonaMatrix enables refinement of legal AI summarization systems for both expert and non-expert users, potentially increasing access to legal knowledge through more targeted and effective summarization.

Abstract: Legal documents are often long, dense, and difficult to comprehend, not only for laypeople but also for legal experts. While automated document summarization has great potential to improve access to legal knowledge, prevailing task-based evaluators overlook divergent user and stakeholder needs. Tool development is needed to encompass the technicality of a case summary for a litigator yet be accessible for a self-help public researching for their lawsuit. We introduce PersonaMatrix, a persona-by-criterion evaluation framework that scores summaries through the lens of six personas, including legal and non-legal users. We also introduce a controlled dimension-shifted pilot dataset of U.S. civil rights case summaries that varies along depth, accessibility, and procedural detail as well as Diversity-Coverage Index (DCI) to expose divergent optima of legal summary between persona-aware and persona-agnostic judges. This work enables refinement of legal AI summarization systems for both expert and non-expert users, with the potential to increase access to legal knowledge. The code base and data are publicly available in GitHub.

Method: Created WolBanking77 dataset containing 9,791 text sentences and over 4 hours of spoken sentences in the banking domain. Conducted experiments with various text and voice state-of-the-art models as baselines.

Result: Promising results on the dataset with reported baseline F1-scores and word error rates for NLP and ASR models trained on WolBanking77, along with model comparisons.

Conclusion: The paper presents a valuable resource for intent classification research in low-resource languages and demonstrates the feasibility of such systems for spoken languages in regions with high illiteracy rates.

Abstract: Intent classification models have made a significant progress in recent years. However, previous studies primarily focus on high-resource language datasets, which results in a gap for low-resource languages and for regions with high rates of illiteracy, where languages are more spoken than read or written. This is the case in Senegal, for example, where Wolof is spoken by around 90% of the population, while the national illiteracy rate remains at of 42%. Wolof is actually spoken by more than 10 million people in West African region. To address these limitations, we introduce the Wolof Banking Speech Intent Classification Dataset (WolBanking77), for academic research in intent classification. WolBanking77 currently contains 9,791 text sentences in the banking domain and more than 4 hours of spoken sentences. Experiments on various baselines are conducted in this work, including text and voice state-of-the-art models. The results are very promising on this current dataset. In addition, this paper presents an in-depth examination of the dataset’s contents. We report baseline F1-scores and word error rates metrics respectively on NLP and ASR models trained on WolBanking77 dataset and also comparisons between models. Dataset and code available at: \href{https://github.com/abdoukarim/wolbanking77}{wolbanking77}.

Method: Built a large-scale TCM corpus (0.97GB unsupervised data + 611,312 QA pairs) and employed two-stage training with QLoRA, DeepSpeed Stage 2, and Flash Attention 2. Optimal configuration: LoRA rank=128, alpha=256, epoch=4, dropout=0.2, max length=2048.

Result: Ranked top-three in all metrics for six datasets (APQ, TCMCD, HFR, HCCA, DHPE, TLAW) and achieved top results in the other six (TCMEE, APR, GCPMI, TCMKQA, TCMRC, ADTG) across 12 benchmarks.

Conclusion: TianHui enables systematic preservation and scalable application of TCM knowledge, with all resources open-sourced for community use.

Abstract: Domain-specific LLMs in TCM face limitations in research settings due to constrained adaptability, insufficient evaluation datasets, and limited computational resources. This study presents TianHui, a specialized TCM LLM built through contextual data integration and domain knowledge fusion. We constructed a large-scale TCM corpus (0.97GB unsupervised data + 611,312 QA pairs) and employed a two-stage training strategy with QLoRA, DeepSpeed Stage 2, and Flash Attention 2. Evaluation on 12 benchmarks showed TianHui ranked top-three in all metrics for six datasets (APQ, TCMCD, HFR, HCCA, DHPE, TLAW) and achieved top results in the other six (TCMEE, APR, GCPMI, TCMKQA, TCMRC, ADTG). Optimal configuration was identified as LoRA rank=128, alpha=256, epoch=4, dropout=0.2, max length=2048. TianHui enables systematic preservation and scalable application of TCM knowledge. All resources are open-sourced.

Method: Introduced CSP-Aware Entropy (computing information gain after constraint propagation) and Probabilistic CSP framework integrating Bayesian word-frequency priors with logical constraints.

Result: CSP-Aware Entropy achieved 3.54 average guesses with 99.9% success rate, 1.7% improvement over Forward Checking with 46% faster runtime. Maintained advantages under noise and achieved 100% success across all noise levels with Probabilistic CSP.

Conclusion: Principled constraint satisfaction techniques outperform classical information-theoretic and learning-based approaches for structured puzzle-solving domains, demonstrating transferability across languages without language-specific tuning.

Abstract: Wordle presents an algorithmically rich testbed for constraint satisfaction problem (CSP) solving. While existing solvers rely on information-theoretic entropy maximization or frequency-based heuristics without formal constraint treatment, we present the first comprehensive CSP formulation of Wordle with novel constraint-aware solving strategies. We introduce CSP-Aware Entropy, computing information gain after constraint propagation rather than on raw candidate sets, and a Probabilistic CSP framework integrating Bayesian word-frequency priors with logical constraints. Through evaluation on 2,315 English words, CSP-Aware Entropy achieves 3.54 average guesses with 99.9% success rate, a statistically significant 1.7% improvement over Forward Checking (t=-4.82, p<0.001, Cohen’s d=0.07) with 46% faster runtime (12.9ms versus 23.7ms per guess). Under 10% noise, CSP-aware approaches maintain 5.3 percentage point advantages (29.0% versus 23.7%, p=0.041), while Probabilistic CSP achieves 100% success across all noise levels (0-20%) through constraint recovery mechanisms. Cross-lexicon validation on 500 Spanish words demonstrates 88% success with zero language-specific tuning, validating that core CSP principles transfer across languages despite an 11.2 percentage point gap from linguistic differences (p<0.001, Fisher’s exact test). Our open-source implementation with 34 unit tests achieving 91% code coverage provides reproducible infrastructure for CSP research. The combination of formal CSP treatment, constraint-aware heuristics, probabilistic-logical integration, robustness analysis, and cross-lexicon validation establishes new performance benchmarks demonstrating that principled constraint satisfaction techniques outperform classical information-theoretic and learning-based approaches for structured puzzle-solving domains.

Method: Stress-test model specifications by generating diverse value tradeoff scenarios where models must choose between pairs of legitimate but conflicting principles, then evaluate responses from twelve frontier LLMs using value classification scores.

Result: Identified over 70,000 cases exhibiting significant behavioral divergence across models, with high divergence strongly predicting underlying specification problems including direct contradictions and interpretive ambiguities.

Conclusion: The methodology effectively reveals specification flaws in current AI models, providing insights into value prioritization patterns and identifying both misalignment cases and false-positive refusals across all tested frontier models.

Abstract: Large language models (LLMs) are increasingly trained from AI constitutions and model specifications that establish behavioral guidelines and ethical principles. However, these specifications face critical challenges, including internal conflicts between principles and insufficient coverage of nuanced scenarios. We present a systematic methodology for stress-testing model character specifications, automatically identifying numerous cases of principle contradictions and interpretive ambiguities in current model specs. We stress test current model specs by generating scenarios that force explicit tradeoffs between competing value-based principles. Using a comprehensive taxonomy we generate diverse value tradeoff scenarios where models must choose between pairs of legitimate principles that cannot be simultaneously satisfied. We evaluate responses from twelve frontier LLMs across major providers (Anthropic, OpenAI, Google, xAI) and measure behavioral disagreement through value classification scores. Among these scenarios, we identify over 70,000 cases exhibiting significant behavioral divergence. Empirically, we show this high divergence in model behavior strongly predicts underlying problems in model specifications. Through qualitative analysis, we provide numerous example issues in current model specs such as direct contradiction and interpretive ambiguities of several principles. Additionally, our generated dataset also reveals both clear misalignment cases and false-positive refusals across all of the frontier models we study. Lastly, we also provide value prioritization patterns and differences of these models.

Method: Introduces Rubric Reward Model (RRM) - a process-oriented reward function that evaluates entire reasoning trajectories against problem-specific rubrics, providing fine-grained calibrated rewards that penalize logical flaws and encourage rigorous deduction.

Result: RRM-based training consistently outperforms outcome-only supervision across four math benchmarks, boosting Verified Pass@1024 on AIME2024 from 26.7% to 62.6% and reducing Miracle Steps by 71%.

Conclusion: Rewarding the solution process is crucial for building models that are not only more accurate but also more reliable, demonstrating that process-oriented evaluation addresses fundamental limitations of outcome-only supervision.

Abstract: Large language models for mathematical reasoning are typically trained with outcome-based rewards, which credit only the final answer. In our experiments, we observe that this paradigm is highly susceptible to reward hacking, leading to a substantial overestimation of a model’s reasoning ability. This is evidenced by a high incidence of false positives - solutions that reach the correct final answer through an unsound reasoning process. Through a systematic analysis with human verification, we establish a taxonomy of these failure modes, identifying patterns like Miracle Steps - abrupt jumps to a correct output without a valid preceding derivation. Probing experiments suggest a strong association between these Miracle Steps and memorization, where the model appears to recall the answer directly rather than deriving it. To mitigate this systemic issue, we introduce the Rubric Reward Model (RRM), a process-oriented reward function that evaluates the entire reasoning trajectory against problem-specific rubrics. The generative RRM provides fine-grained, calibrated rewards (0-1) that explicitly penalize logical flaws and encourage rigorous deduction. When integrated into a reinforcement learning pipeline, RRM-based training consistently outperforms outcome-only supervision across four math benchmarks. Notably, it boosts Verified Pass@1024 on AIME2024 from 26.7% to 62.6% and reduces the incidence of Miracle Steps by 71%. Our work demonstrates that rewarding the solution process is crucial for building models that are not only more accurate but also more reliable.

Method: Three synergistic components: adaptive execution scheduling (dynamic compute graph balancing), context-aligned drafting (lightweight online calibration), and hardware-efficient draft extension (reusing and expanding intermediate sequences).

Result: Experiments show consistent improvements of up to 3.8x in generation speed and 4.7x in energy efficiency compared to existing mobile inference solutions.

Conclusion: The framework effectively accelerates context-aware text generation on mobile devices through speculative decoding and dynamic hardware scheduling, with component-level analysis validating each optimization’s contribution.

Abstract: Enhancing on-device large language models (LLMs) with contextual information from local data enables personalized and task-aware generation, powering use cases such as intelligent assistants and UI agents. While recent developments in neural processors have substantially improved the efficiency of prefill on mobile devices, the token-by-token generation process still suffers from high latency and limited hardware utilization due to its inherently memory-bound characteristics. This work presents sd.npu, a mobile inference framework that integrates speculative decoding with dynamic hardware scheduling to accelerate context-aware text generation on mobile devices. The framework introduces three synergistic components: (1) adaptive execution scheduling, which dynamically balances compute graphs between prefill and decoding phases; (2) context-aligned drafting, which improves speculative efficiency through lightweight online calibration to current tasks; and (3) hardware-efficient draft extension, which reuses and expands intermediate sequences to improve processing parallelism and reduce verification cost. Experiments on multiple smartphones and representative workloads show consistent improvements of up to 3.8x in generation speed and 4.7x in energy efficiency compared with existing mobile inference solutions. Component-level analysis further validates the contribution of each optimization.

Method: Used incremental learning setup as diagnostic tool to measure semantic drift, analyzed representation overlap between entities, and investigated ‘O’ tag plasticity by unfreezing the background class classifier.

Result: Found LOC entity is vulnerable due to pattern-like feature overlap with PII (e.g., postal codes), and identified “reverse O-tag representation drift” where frozen ‘O’ tag blocks new learning until unfrozen.

Conclusion: NER model adaptation involves feature independence mechanisms, representation overlap vulnerabilities, and requires ‘O’ tag plasticity for successful learning of new pattern-based entities.

Abstract: Extending Named Entity Recognition (NER) models to new PII entities in noisy spoken-language data is a common need. We find that jointly fine-tuning a BERT model on standard semantic entities (PER, LOC, ORG) and new pattern-based PII (EMAIL, PHONE) results in minimal degradation for original classes. We investigate this “peaceful coexistence,” hypothesizing that the model uses independent semantic vs. morphological feature mechanisms. Using an incremental learning setup as a diagnostic tool, we measure semantic drift and find two key insights. First, the LOC (location) entity is uniquely vulnerable due to a representation overlap with new PII, as it shares pattern-like features (e.g., postal codes). Second, we identify a “reverse O-tag representation drift.” The model, initially trained to map PII patterns to ‘O’, blocks new learning. This is resolved only by unfreezing the ‘O’ tag’s classifier, allowing the background class to adapt and “release” these patterns. This work provides a mechanistic diagnosis of NER model adaptation, highlighting feature independence, representation overlap, and ‘O’ tag plasticity. Work done based on data gathered by https://www.papernest.com

Method: Multi-stage curriculum: Mid-Term Training for reasoning/planning/reflection, SFT with balanced programming languages/contexts/tasks, RFT with multi-ground-truth reward formulation, and Reinforcement-to-Deployment Adaptation with Error-Masked SFT and Tree-Structured Trajectory Training.

Result: KAT-Coder achieves robust tool-use reliability, instruction alignment, and long-context reasoning, forming a deployable foundation for real-world intelligent coding agents. The KAT-Dev 32B model has been open-sourced.

Conclusion: The multi-stage training curriculum enables KAT-Coder to effectively bridge the gap between static training and dynamic agentic execution, creating a foundation for deployable intelligent coding agents.

Abstract: Recent advances in large language models (LLMs) have enabled progress in agentic coding, where models autonomously reason, plan, and act within interactive software development workflows. However, bridging the gap between static text-based training and dynamic real-world agentic execution remains a core challenge. In this technical report, we present KAT-Coder, a large-scale agentic code model trained through a multi-stage curriculum encompassing Mid-Term Training, Supervised Fine-Tuning (SFT), Reinforcement Fine-Tuning (RFT), and Reinforcement-to-Deployment Adaptation. The Mid-Term stage enhances reasoning, planning, and reflection capabilities through a corpus of real software engineering data and synthetic agentic interactions. The SFT stage constructs a million-sample dataset balancing twenty programming languages, ten development contexts, and ten task archetypes. The RFT stage introduces a novel multi-ground-truth reward formulation for stable and sample-efficient policy optimization. Finally, the Reinforcement-to-Deployment phase adapts the model to production-grade IDE environments using Error-Masked SFT and Tree-Structured Trajectory Training. In summary, these stages enable KAT-Coder to achieve robust tool-use reliability, instruction alignment, and long-context reasoning, forming a deployable foundation for real-world intelligent coding agents. Our KAT series 32B model, KAT-Dev, has been open-sourced on https://huggingface.co/Kwaipilot/KAT-Dev.

Method: Comprehensive examination of a widely used professional assessment questionnaire using state-of-the-art LLMs to generate responses, followed by performance evaluation against company-referenced solutions.

Result: Significant inconsistency between model-generated answers and company-referenced solutions; all evaluated LLMs failed to pass the hiring evaluation.

Conclusion: Despite their coding capabilities, current LLMs cannot successfully pass professional hiring evaluations for software engineering positions.

Abstract: With the proliferation of the internet and the rapid advancement of Artificial Intelligence, leading technology companies face an urgent annual demand for a considerable number of software and algorithm engineers. To efficiently and effectively identify high-potential candidates from thousands of applicants, these firms have established a multi-stage selection process, which crucially includes a standardized hiring evaluation designed to assess job-specific competencies. Motivated by the demonstrated prowess of Large Language Models (LLMs) in coding and reasoning tasks, this paper investigates a critical question: Can LLMs successfully pass these hiring evaluations? To this end, we conduct a comprehensive examination of a widely used professional assessment questionnaire. We employ state-of-the-art LLMs to generate responses and subsequently evaluate their performance. Contrary to any prior expectation of LLMs being ideal engineers, our analysis reveals a significant inconsistency between the model-generated answers and the company-referenced solutions. Our empirical findings lead to a striking conclusion: All evaluated LLMs fails to pass the hiring evaluation.

Method: Proposed Zhyper framework uses factorized hypernetworks to generate context-aware LoRA adapters from textual descriptions, enabling parameter-efficient conditioning.

Result: Zhyper achieves competitive performance on multiple benchmarks with up to 26x fewer parameters than state-of-the-art baselines, and shows improved generalization to out-of-domain settings in cultural alignment tasks.

Conclusion: Zhyper provides an effective parameter-efficient solution for LLM conditioning that captures fine-grained contextual values while maintaining strong generalization capabilities.

Abstract: Large Language Model (LLM) conditioning refers to instructing an LLM to generate content in accordance with the norms and values of a specific culture, beliefs of a particular political orientation, or any desired text-specified semantic conditioning. Unfortunately, prompt engineering does not ensure that LLMs behave in accordance with a desired conditioning due to the inductive bias of the pre-training and alignment datasets. Prior works have focused on fine-tuning LLMs by directly conditioning the LoRA weights; however, such methods introduce a large number of parameters. As a remedy, we propose Zhyper, a parameter-efficient factorized hypernetwork framework that generates context-aware LoRA adapters from textual descriptions. Experiments on multiple benchmarks show that Zhyper achieves competitive performance with up to 26x fewer parameters than the state-of-the-art baselines. Furthermore, we extend Zhyper to cultural alignment, demonstrating improved generalization to out-of-domain settings and a better capturing of fine-grained contextual values.

Method: Three-step approach: (1) continued pre-training on unlabeled code-mixed text, (2) merging the adapted checkpoint with the base multilingual model, (3) fine-tuning on downstream task data. Evaluated on English-Hindi and English-Spanish sentence classification using XLM-R and Llama-3.2-1B models.

Result: Merged models consistently outperform full fine-tuning (2-5 F1 points gain) and CPT->FT (~1-2 points gain). Better cross-pair transfer: merged checkpoints achieve 0.65-0.68 F1 vs 0.61-0.63 for full fine-tuning on En-Ta and En-Ml. Zero/few-shot prompting with larger LLMs lags behind fine-tuned models.

Conclusion: Model merging effectively leverages unlabeled code-mixed data and provides more reliable knowledge for low-resource language pairs. The paper provides adaptation recipes for different data regimes and discusses scaling considerations.

Abstract: We study model merging as a practical alternative to conventional adaptation strategies for code-mixed NLP. Starting from a multilingual base model, we: (i) perform continued pre-training (CPT) on unlabeled code-mixed text to obtain an adapted checkpoint, (ii) merge checkpoint with the base model, and (iii) fine-tune (FT) on the downstream task data. We evaluate our approach for sentence classification (sentiment and hate speech) task in English-Hindi (En-Hi) and English-Spanish (En-Es) using XLM-R and Llama-3.2-1B models. Our results show that merged models consistently outperform full fine-tuning and CPT->FT. We observe gains of 2–5 points in F1 over full fine-tuning and ~1-2 points over CPT->FT, indicating that unlabeled data is leveraged more effectively via merging than via CPT alone. Zero-/few-shot prompting with larger LLMs (e.g., Llama-3.3-70B) lags behind fine-tuned and merged checkpoints, underscoring limits of in-context learning for code-mixed inputs. We further test cross-pair transfer by training on En-Hi and evaluating on En-Ta and En-Ml: merged checkpoints transfer more strongly than monolingual-English baselines (e.g., TV/TIES variants reaching 0.65-0.68 F1 vs 0.61-0.63 for full fine-tuning), suggesting that code-mixed knowledge is a more reliable substrate for low-resource pairs. We conclude with adaptation recipes matched to common data regimes (labeled only; labeled+unlabeled; transfer-only) and discuss limitations and scaling considerations for broader tasks and larger models.

Method: Apply 2D DFT to transform images into Fourier domain to detect subtle periodic artifacts, then train ResNet50 neural network on these transformed images to differentiate real vs synthetic images.

Result: Frequency-domain model achieves 92.8% accuracy and 0.95 AUC, significantly outperforming equivalent model trained on raw spatial-domain images.

Conclusion: GAN-generated images have unique frequency-domain signatures; combining signal processing with deep learning enhances digital forensics and strengthens industrial AI system trustworthiness.

Abstract: The rapid rise of photorealistic images produced from Generative Adversarial Networks (GANs) poses a serious challenge for image forensics and industrial systems requiring reliable content authenticity. This paper uses frequency-domain analysis combined with deep learning to solve the problem of distinguishing StyleGAN-generated images from real ones. Specifically, a two-dimensional Discrete Fourier Transform (2D DFT) was applied to transform images into the Fourier domain, where subtle periodic artifacts become detectable. A ResNet50 neural network is trained on these transformed images to differentiate between real and synthetic ones. The experiments demonstrate that the frequency-domain model achieves a 92.8 percent and an AUC of 0.95, significantly outperforming the equivalent model trained on raw spatial-domain images. These results indicate that the GAN-generated images have unique frequency-domain signatures or “fingerprints”. The method proposed highlights the industrial potential of combining signal processing techniques and deep learning to enhance digital forensics and strengthen the trustworthiness of industrial AI systems.

Method: Proposes CMC with Pseudo Label Generation Module for self-supervised unimodal pretraining, Parameter-free Fusion Module and Multimodal Consensus Router for multimodal finetuning.

Result: Achieves state-of-the-art performance on CH-SIMS, CH-SIMS v2, CMU-MOSI, and CMU-MOSEI datasets, with notable advantages in semantic inconsistency scenarios.

Conclusion: CMC effectively mitigates text dominance and guides fusion toward reliable consensus, improving multimodal emotion recognition performance.

Abstract: In recent years, Multimodal Emotion Recognition (MER) has made substantial progress. Nevertheless, most existing approaches neglect the semantic inconsistencies that may arise across modalities, such as conflicting emotional cues between text and visual inputs. Besides, current methods are often dominated by the text modality due to its strong representational capacity, which can compromise recognition accuracy. To address these challenges, we propose a model termed Calibrated Multimodal Consensus (CMC). CMC introduces a Pseudo Label Generation Module (PLGM) to produce pseudo unimodal labels, enabling unimodal pretraining in a self-supervised fashion. It then employs a Parameter-free Fusion Module (PFM) and a Multimodal Consensus Router (MCR) for multimodal finetuning, thereby mitigating text dominance and guiding the fusion process toward a more reliable consensus. Experimental results demonstrate that CMC achieves performance on par with or superior to state-of-the-art methods across four datasets, CH-SIMS, CH-SIMS v2, CMU-MOSI, and CMU-MOSEI, and exhibits notable advantages in scenarios with semantic inconsistencies on CH-SIMS and CH-SIMS v2. The implementation of this work is publicly accessible at https://github.com/gw-zhong/CMC.

Method: Combine Vision Transformers (ViTs) with contrastive learning objectives (both supervised and self-supervised) for multi-view 3D analysis.

Result: Achieved 90.6% accuracy on ModelNet10, demonstrating ViTs capture global shape semantics while contrastive learning refines local features.

Conclusion: ViTs paired with contrastive learning effectively unify 3D shape understanding pipelines, overcoming CNN limitations and reducing dependency on labeled data through empirical validation.

Abstract: This paper addresses the challenges in representation learning of 3D shape features by investigating state-of-the-art backbones paired with both contrastive supervised and self-supervised learning objectives. Computer vision methods struggle with recognizing 3D objects from 2D images, often requiring extensive labeled data and relying on Convolutional Neural Networks (CNNs) that may overlook crucial shape relationships. Our work demonstrates that Vision Transformers (ViTs) based architectures, when paired with modern contrastive objectives, achieve promising results in multi-view 3D analysis on our downstream tasks, unifying contrastive and 3D shape understanding pipelines. For example, supervised contrastive losses reached about 90.6% accuracy on ModelNet10. The use of ViTs and contrastive learning, leveraging ViTs’ ability to understand overall shapes and contrastive learning’s effectiveness, overcomes the need for extensive labeled data and the limitations of CNNs in capturing crucial shape relationships. The success stems from capturing global shape semantics via ViTs and refining local discriminative features through contrastive optimization. Importantly, our approach is empirical, as it is grounded on extensive experimental evaluation to validate the effectiveness of combining ViTs with contrastive objectives for 3D representation learning.

Method: Generate fine-grained multi-modal features by summarizing video content and using text captions to enhance visual features of highlighted portions, reducing feature redundancy.

Result: Achieves state-of-the-art performance on four major datasets, with ablation experiments confirming improvements come from reduced redundant information.

Conclusion: Leveraging multi-modal diversity through fine-grained feature refinement effectively reduces redundancy and enhances video anomaly detection performance.

Abstract: Video anomaly detection (VAD) is a challenging task that detects anomalous frames in continuous surveillance videos. Most previous work utilizes the spatio-temporal correlation of visual features to distinguish whether there are abnormalities in video snippets. Recently, some works attempt to introduce multi-modal information, like text feature, to enhance the results of video anomaly detection. However, these works merely incorporate text features into video snippets in a coarse manner, overlooking the significant amount of redundant information that may exist within the video snippets. Therefore, we propose to leverage the diversity among multi-modal information to further refine the extracted features, reducing the redundancy in visual features, and we propose Grained Multi-modal Feature for Video Anomaly Detection (GMFVAD). Specifically, we generate more grained multi-modal feature based on the video snippet, which summarizes the main content, and text features based on the captions of original video will be introduced to further enhance the visual features of highlighted portions. Experiments show that the proposed GMFVAD achieves state-of-the-art performance on four mainly datasets. Ablation experiments also validate that the improvement of GMFVAD is due to the reduction of redundant information.

Method: Uses a two-stage transformer refinement and tracking pipeline with multimodal BEV fusion features. Includes temporal smoother for trajectory refinement and fusion tracker that integrates bounding boxes with multimodal features for identity propagation.

Result: Achieves 74.7 aMOTA on nuScenes test set, demonstrating strong performance on 3D MOT benchmarks with reduced identity switches while maintaining competitive accuracy. Outperforms previous single-sensor approaches.

Conclusion: Transformer-based tracking methods significantly benefit from multimodal sensor features. The framework provides efficient improvement for transformer-based trackers to compete with neural-network-based methods even with limited data and no pretraining.

Abstract: We propose FutrTrack, a modular camera-LiDAR multi-object tracking framework that builds on existing 3D detectors by introducing a transformer-based smoother and a fusion-driven tracker. Inspired by query-based tracking frameworks, FutrTrack employs a multimodal two-stage transformer refinement and tracking pipeline. Our fusion tracker integrates bounding boxes with multimodal bird’s-eye-view (BEV) fusion features from multiple cameras and LiDAR without the need for an explicit motion model. The tracker assigns and propagates identities across frames, leveraging both geometric and semantic cues for robust re-identification under occlusion and viewpoint changes. Prior to tracking, we refine sequences of bounding boxes with a temporal smoother over a moving window to refine trajectories, reduce jitter, and improve spatial consistency. Evaluated on nuScenes and KITTI, FutrTrack demonstrates that query-based transformer tracking methods benefit significantly from multimodal sensor features compared with previous single-sensor approaches. With an aMOTA of 74.7 on the nuScenes test set, FutrTrack achieves strong performance on 3D MOT benchmarks, reducing identity switches while maintaining competitive accuracy. Our approach provides an efficient framework for improving transformer-based trackers to compete with other neural-network-based methods even with limited data and without pretraining.

Method: Introduced VCR-OOD datasets (VCR-OOD-QA and VCR-OOD-VA) for evaluation, analyzed causal graphs and prediction shortcuts, and adopted backdoor adjustment method with a dictionary based on correct answers to remove bias.

Result: Experiments demonstrate the effectiveness of the debiasing method across different datasets.

Conclusion: The proposed approach successfully addresses bias in VCR datasets and improves model generalization through systematic debiasing.

Abstract: Visual Commonsense Reasoning (VCR) refers to answering questions and providing explanations based on images. While existing methods achieve high prediction accuracy, they often overlook bias in datasets and lack debiasing strategies. In this paper, our analysis reveals co-occurrence and statistical biases in both textual and visual data. We introduce the VCR-OOD datasets, comprising VCR-OOD-QA and VCR-OOD-VA subsets, which are designed to evaluate the generalization capabilities of models across two modalities. Furthermore, we analyze the causal graphs and prediction shortcuts in VCR and adopt a backdoor adjustment method to remove bias. Specifically, we create a dictionary based on the set of correct answers to eliminate prediction shortcuts. Experiments demonstrate the effectiveness of our debiasing method across different datasets.

Method: Used Online Label Smoothing (OLS) - a dynamic approach that adjusts soft labels throughout training based on model’s prediction patterns. Evaluated on RadImageNet dataset using ResNet-50, MobileNetV2, and VGG-19 architectures.

Result: OLS consistently improved both Top-1 and Top-5 classification accuracy compared to standard methods. Also led to more compact and well-separated feature embeddings, indicating improved representation learning.

Conclusion: OLS strengthens predictive performance and enhances calibration, making it a practical solution for developing trustworthy AI systems in medical imaging.

Abstract: Deep learning models, especially convolutional neural networks, have achieved impressive results in medical image classification. However, these models often produce overconfident predictions, which can undermine their reliability in critical healthcare settings. While traditional label smoothing offers a simple way to reduce such overconfidence, it fails to consider relationships between classes by treating all non-target classes equally. In this study, we explore the use of Online Label Smoothing (OLS), a dynamic approach that adjusts soft labels throughout training based on the model’s own prediction patterns. We evaluate OLS on the large-scale RadImageNet dataset using three widely used architectures: ResNet-50, MobileNetV2, and VGG-19. Our results show that OLS consistently improves both Top-1 and Top-5 classification accuracy compared to standard training methods, including hard labels, conventional label smoothing, and teacher-free knowledge distillation. In addition to accuracy gains, OLS leads to more compact and well-separated feature embeddings, indicating improved representation learning. These findings suggest that OLS not only strengthens predictive performance but also enhances calibration, making it a practical and effective solution for developing trustworthy AI systems in the medical imaging domain.

Method: Proposes DMC³ framework with: 1) Counterfactual sample construction via event description paraphrasing (text) and core interaction mining (visual), 2) Baseline model processing original and counterfactual samples, 3) Contrastive optimization to minimize distance to positive samples and maximize distance to negative samples.

Result: Achieves state-of-the-art performance: 52.51% on EgoTaskQA normal split, 46.04% on EgoTaskQA indirect split, and 13.2% on QAEGO4D.

Conclusion: The proposed DMC³ framework effectively addresses first-person perspective challenges in Egocentric VideoQA through counterfactual contrastive learning, demonstrating superior performance on benchmark datasets.

Abstract: Egocentric Video Question Answering (Egocentric VideoQA) plays an important role in egocentric video understanding, which refers to answering questions based on first-person videos. Although existing methods have made progress through the paradigm of pre-training and fine-tuning, they ignore the unique challenges posed by the first-person perspective, such as understanding multiple events and recognizing hand-object interactions. To deal with these challenges, we propose a Dual-Modal Counterfactual Contrastive Construction (DMC$^3$) framework, which contains an egocentric videoqa baseline, a counterfactual sample construction module and a counterfactual sample-involved contrastive optimization. Specifically, We first develop a counterfactual sample construction module to generate positive and negative samples for textual and visual modalities through event description paraphrasing and core interaction mining, respectively. Then, We feed these samples together with the original samples into the baseline. Finally, in the counterfactual sample-involved contrastive optimization module, we apply contrastive loss to minimize the distance between the original sample features and the positive sample features, while maximizing the distance from the negative samples. Experiments show that our method achieve 52.51% and 46.04% on the \textit{normal} and \textit{indirect} splits of EgoTaskQA, and 13.2% on QAEGO4D, both reaching the state-of-the-art performance.

Method: Uses a simple pre and post processing pipeline to enhance detection consistency across distorted regions, trains multiple state-of-the-art detection models on fisheye traffic imagery, and combines their outputs through ensemble strategy.

Result: Achieved F1 score of 0.6366 on the 2025 AI City Challenge Track 4, ranking 8th out of 62 teams, demonstrating robust performance in fisheye imagery conditions.

Conclusion: The framework effectively addresses fisheye-specific challenges and provides a practical solution for wide-area traffic surveillance using fisheye cameras.

Abstract: Fisheye cameras offer an efficient solution for wide-area traffic surveillance by capturing large fields of view from a single vantage point. However, the strong radial distortion and nonuniform resolution inherent in fisheye imagery introduce substantial challenges for standard object detectors, particularly near image boundaries where object appearance is severely degraded. In this work, we present a detection framework designed to operate robustly under these conditions. Our approach employs a simple yet effective pre and post processing pipeline that enhances detection consistency across the image, especially in regions affected by severe distortion. We train several state-of-the-art detection models on the fisheye traffic imagery and combine their outputs through an ensemble strategy to improve overall detection accuracy. Our method achieves an F1 score of0.6366 on the 2025 AI City Challenge Track 4, placing 8thoverall out of 62 teams. These results demonstrate the effectiveness of our framework in addressing issues inherent to fisheye imagery.

Method: Novel causal approach that predicts culinary elements overlooked in images and explicitly injects these elements into cross-modal representation learning to mitigate biases in learning.

Result: Achieves impressive retrieval performance on both Recipe1M dataset and new multilingual multicultural cuisine dataset, capable of uncovering subtle ingredients and cooking actions.

Conclusion: Causal representation learning effectively mitigates biases in cross-modal learning by capturing non-visual recipe details, improving retrieval accuracy for similar recipes with subtle differences.

Abstract: Existing approaches for image-to-recipe retrieval have the implicit assumption that a food image can fully capture the details textually documented in its recipe. However, a food image only reflects the visual outcome of a cooked dish and not the underlying cooking process. Consequently, learning cross-modal representations to bridge the modality gap between images and recipes tends to ignore subtle, recipe-specific details that are not visually apparent but are crucial for recipe retrieval. Specifically, the representations are biased to capture the dominant visual elements, resulting in difficulty in ranking similar recipes with subtle differences in use of ingredients and cooking methods. The bias in representation learning is expected to be more severe when the training data is mixed of images and recipes sourced from different cuisines. This paper proposes a novel causal approach that predicts the culinary elements potentially overlooked in images, while explicitly injecting these elements into cross-modal representation learning to mitigate biases. Experiments are conducted on the standard monolingual Recipe1M dataset and a newly curated multilingual multicultural cuisine dataset. The results indicate that the proposed causal representation learning is capable of uncovering subtle ingredients and cooking actions and achieves impressive retrieval performance on both monolingual and multilingual multicultural datasets.

Method: Proposes a render-aware filtering method using sensitivity scores from intermediate gradients, targeting instabilities from anisotropic orientations rather than isotropic variance.

Result: Substantially improves visual quality, realism, and consistency compared to existing NeRF-based approaches like BayesRays, while maintaining real-time performance.

Conclusion: The filter effectively addresses generative uncertainty in 3DGS, enabling high visual fidelity during free navigation outside original training viewpoints without requiring retraining or fine-tuning.

Abstract: When viewing a 3D Gaussian Splatting (3DGS) model from camera positions significantly outside the training data distribution, substantial visual noise commonly occurs. These artifacts result from the lack of training data in these extrapolated regions, leading to uncertain density, color, and geometry predictions from the model. To address this issue, we propose a novel real-time render-aware filtering method. Our approach leverages sensitivity scores derived from intermediate gradients, explicitly targeting instabilities caused by anisotropic orientations rather than isotropic variance. This filtering method directly addresses the core issue of generative uncertainty, allowing 3D reconstruction systems to maintain high visual fidelity even when users freely navigate outside the original training viewpoints. Experimental evaluation demonstrates that our method substantially improves visual quality, realism, and consistency compared to existing Neural Radiance Field (NeRF)-based approaches such as BayesRays. Critically, our filter seamlessly integrates into existing 3DGS rendering pipelines in real-time, unlike methods that require extensive post-hoc retraining or fine-tuning. Code and results at https://damian-bowness.github.io/EV3DGS

Method: Curated two high-quality datasets (STGR-CoT-30k for SFT and STGR-RL-36k for RL) with unified spatio-temporal annotations, and adopted a cold-start reinforcement learning strategy with multiple specially designed rewards that jointly encourage answer accuracy, temporal alignment, and spatial precision.

Result: Achieved state-of-the-art performance on V-STAR benchmark, raising mAM by 14.4% and mLGM by 24.2% on the Qwen2.5-VL baseline. Consistent improvements observed on VideoMME, WorldSense, VideoMMMU, and TVGBench. Reasoning traces also provide valuable signals for test-time scaling and confidence-aware verification.

Conclusion: Open-o3 Video successfully integrates explicit spatio-temporal evidence into video reasoning, demonstrating superior performance across multiple benchmarks while providing interpretable reasoning traces that improve answer reliability and enable confidence-aware verification.

Abstract: Most video reasoning models only generate textual reasoning traces without indicating when and where key evidence appears. Recent models such as OpenAI-o3 have sparked wide interest in evidence-centered reasoning for images, yet extending this ability to videos is more challenging, as it requires joint temporal tracking and spatial localization across dynamic scenes. We introduce Open-o3 Video, a non-agent framework that integrates explicit spatio-temporal evidence into video reasoning, and carefully collect training data and design training strategies to address the aforementioned challenges. The model highlights key timestamps, objects, and bounding boxes alongside its answers, allowing reasoning to be grounded in concrete visual observations. To enable this functionality, we first curate and build two high-quality datasets, STGR-CoT-30k for SFT and STGR-RL-36k for RL, with carefully constructed temporal and spatial annotations, since most existing datasets offer either temporal spans for videos or spatial boxes on images, lacking unified spatio-temporal supervision and reasoning traces. Then, we adopt a cold-start reinforcement learning strategy with multiple specially designed rewards that jointly encourage answer accuracy, temporal alignment, and spatial precision. On V-STAR benchmark, Open-o3 Video achieves state-of-the-art performance, raising mAM by 14.4% and mLGM by 24.2% on the Qwen2.5-VL baseline. Consistent improvements are also observed on a broad range of video understanding benchmarks, including VideoMME, WorldSense, VideoMMMU, and TVGBench. Beyond accuracy, the reasoning traces produced by Open-o3 Video also provide valuable signals for test-time scaling, enabling confidence-aware verification and improving answer reliability.

Method: Two-stage hybrid framework: 1) Reverse time migration (time-reversal acoustics) applied to multi-angle acquisitions to produce migration fragments preserving structural details under low SNR; 2) Transformer-based super-resolution encoder-decoder with graph-based attention unit (GAU) fuses fragments into coherent SoS image. Uses partial-array acquisition with movable low-count transducer set.

Result: Experiments on two synthetic datasets show BrainPuzzle achieves superior speed-of-sound reconstruction accuracy and image completeness compared to existing methods.

Conclusion: BrainPuzzle demonstrates potential for advancing quantitative ultrasound brain imaging by effectively combining physical modeling with machine learning to overcome limitations of traditional approaches.

Abstract: Ultrasound brain imaging remains challenging due to the large difference in sound speed between the skull and brain tissues and the difficulty of coupling large probes to the skull. This work aims to achieve quantitative transcranial ultrasound by reconstructing an accurate speed-of-sound (SoS) map of the brain. Traditional physics-based full-waveform inversion (FWI) is limited by weak signals caused by skull-induced attenuation, mode conversion, and phase aberration, as well as incomplete spatial coverage since full-aperture arrays are clinically impractical. In contrast, purely data-driven methods that learn directly from raw ultrasound data often fail to model the complex nonlinear and nonlocal wave propagation through bone, leading to anatomically plausible but quantitatively biased SoS maps under low signal-to-noise and sparse-aperture conditions. To address these issues, we propose BrainPuzzle, a hybrid two-stage framework that combines physical modeling with machine learning. In the first stage, reverse time migration (time-reversal acoustics) is applied to multi-angle acquisitions to produce migration fragments that preserve structural details even under low SNR. In the second stage, a transformer-based super-resolution encoder-decoder with a graph-based attention unit (GAU) fuses these fragments into a coherent and quantitatively accurate SoS image. A partial-array acquisition strategy using a movable low-count transducer set improves feasibility and coupling, while the hybrid algorithm compensates for the missing aperture. Experiments on two synthetic datasets show that BrainPuzzle achieves superior SoS reconstruction accuracy and image completeness, demonstrating its potential for advancing quantitative ultrasound brain imaging.

Method: Shifts from single-frame to multi-frame training using Back Projection, optical flow estimation via TVNet, and replaces deterministic predictions with unbiased Monte Carlo estimation through Langevin dynamics.

Result: Outperforms Lucky Imaging, DIP, RVRT, and DiffIR2VR-Zero on synthetic datasets for SSIM, PSNR, LPIPS, and DISTS metrics. Requires fewer input images than Lucky Imaging and is less prone to overfitting or artifacts.

Conclusion: The method maintains high reconstruction quality on real-world astronomical data despite domain shifts, confirming practical robustness for astrophotography applications.

Abstract: Modern image restoration and super-resolution methods utilize deep learning due to its superior performance compared to traditional algorithms. However, deep learning typically requires large training datasets, which are rarely available in astrophotography. Deep Image Prior (DIP) bypasses this constraint by performing blind training on a single image. Although effective in some cases, DIP often suffers from overfitting, artifact generation, and instability. To overcome these issues and improve general performance, this work proposes DIPLI - a framework that shifts from single-frame to multi-frame training using the Back Projection technique, combined with optical flow estimation via the TVNet model, and replaces deterministic predictions with unbiased Monte Carlo estimation obtained through Langevin dynamics. A comprehensive evaluation compares the method against Lucky Imaging, a classical computer vision technique still widely used in astronomical image reconstruction, DIP, the transformer-based model RVRT, and the diffusion-based model DiffIR2VR-Zero. Experiments on synthetic datasets demonstrate consistent improvements, with the method outperforming baselines for SSIM, PSNR, LPIPS, and DISTS metrics in the majority of cases. In addition to superior reconstruction quality, the model also requires far fewer input images than Lucky Imaging and is less prone to overfitting or artifact generation. Evaluation on real-world astronomical data, where domain shifts typically hinder generalization, shows that the method maintains high reconstruction quality, confirming practical robustness.

Method: Proposes learnable pair-specific increments Δ_ij derived via multivariate first-order Taylor expansion of InfoNCE loss under trust-region constraint, coupled by a lightweight neural module conditioned on semantic gap, and regularized through variational information bottleneck.

Result: Experiments on four benchmarks show GARE consistently improves alignment accuracy and robustness, validating effectiveness of gap-aware tension mitigation.

Conclusion: GARE effectively addresses modality gap issues in text-video retrieval through gap-aware gradient redistribution and noise absorption, demonstrating improved performance across multiple benchmarks.

Abstract: Recent progress in text-video retrieval has been largely driven by contrastive learning. However, existing methods often overlook the effect of the modality gap, which causes anchor representations to undergo in-place optimization (i.e., optimization tension) that limits their alignment capacity. Moreover, noisy hard negatives further distort the semantics of anchors. To address these issues, we propose GARE, a Gap-Aware Retrieval framework that introduces a learnable, pair-specific increment $\Delta_{ij}$ between text $t_i$ and video $v_j$, redistributing gradients to relieve optimization tension and absorb noise. We derive $\Delta_{ij}$ via a multivariate first-order Taylor expansion of the InfoNCE loss under a trust-region constraint, showing that it guides updates along locally consistent descent directions. A lightweight neural module conditioned on the semantic gap couples increments across batches for structure-aware correction. Furthermore, we regularize $\Delta$ through a variational information bottleneck with relaxed compression, enhancing stability and semantic consistency. Experiments on four benchmarks demonstrate that GARE consistently improves alignment accuracy and robustness, validating the effectiveness of gap-aware tension mitigation. Code is available at https://github.com/musicman217/GARE-text-video-retrieval.

Method: A unified evaluation framework across six countries using an 8-category/36-subcategory schema with era-aware prompts, combining automatic metrics, culture-aware retrieval-augmented VQA, and expert human judgments from native reviewers with open models and fixed settings.

Result: Three key findings: (1) country-agnostic prompts default to Global-North modern depictions, (2) iterative I2I editing erodes cultural fidelity despite stable metrics, (3) I2I models apply superficial cues rather than era-consistent changes, often retaining source identity for Global-South targets.

Conclusion: Culture-sensitive edits remain unreliable in current generative image systems, and the study provides a reproducible, culture-centered benchmark for diagnosing and tracking cultural bias through standardized data, prompts, and evaluation protocols.

Abstract: Generative image models produce striking visuals yet often misrepresent culture. Prior work has examined cultural bias mainly in text-to-image (T2I) systems, leaving image-to-image (I2I) editors underexplored. We bridge this gap with a unified evaluation across six countries, an 8-category/36-subcategory schema, and era-aware prompts, auditing both T2I generation and I2I editing under a standardized protocol that yields comparable diagnostics. Using open models with fixed settings, we derive cross-country, cross-era, and cross-category evaluations. Our framework combines standard automatic metrics, a culture-aware retrieval-augmented VQA, and expert human judgments collected from native reviewers. To enable reproducibility, we release the complete image corpus, prompts, and configurations. Our study reveals three findings: (1) under country-agnostic prompts, models default to Global-North, modern-leaning depictions that flatten cross-country distinctions; (2) iterative I2I editing erodes cultural fidelity even when conventional metrics remain flat or improve; and (3) I2I models apply superficial cues (palette shifts, generic props) rather than era-consistent, context-aware changes, often retaining source identity for Global-South targets. These results highlight that culture-sensitive edits remain unreliable in current systems. By releasing standardized data, prompts, and human evaluation protocols, we provide a reproducible, culture-centered benchmark for diagnosing and tracking cultural bias in generative image models.

Method: Proposed Salient Epistemic Disentanglement (SED) that injects expert-curated bipartite graph to decouple view-disease associations and uses reinforcement learning for clinically faithful preference selection. Trained on FetalSigma-1M dataset with 20K reports from 12 medical centers.

Result: Outperforms open- and closed-source baselines across all gestational stages with +14% average gains and +61.2% higher accuracy on critical conditions, while being efficient, stable, and scalable.

Conclusion: FetalMind successfully bridges the gap in fetal ultrasound AI by addressing domain-specific challenges through clinical workflow-guided design and large-scale domain data curation, demonstrating superior performance and clinical alignment.

Abstract: Recent medical vision-language models have shown promise on tasks such as VQA, report generation, and anomaly detection. However, most are adapted to structured adult imaging and underperform in fetal ultrasound, which poses challenges of multi-view image reasoning, numerous diseases, and image diversity. To bridge this gap, we introduce FetalMind, a medical AI system tailored to fetal ultrasound for both report generation and diagnosis. Guided by clinical workflow, we propose Salient Epistemic Disentanglement (SED), which injects an expert-curated bipartite graph into the model to decouple view-disease associations and to steer preference selection along clinically faithful steps via reinforcement learning. This design mitigates variability across diseases and heterogeneity across views, reducing learning bottlenecks while aligning the model’s inference with obstetric practice. To train FetalMind at scale, we curate FetalSigma-1M dataset, the first large-scale fetal ultrasound report corpus, comprising 20K reports from twelve medical centers, addressing the scarcity of domain data. Extensive experiments show that FetalMind outperforms open- and closed-source baselines across all gestational stages, achieving +14% average gains and +61.2% higher accuracy on critical conditions while remaining efficient, stable, and scalable. Project Page: https://hexiao0275.github.io/FetalMind.

Method: Uses temporal IIR filter to integrate intensity changes from events, detects stale pixels with a novel algorithm, and applies Gaussian blur to reduce noise.

Result: Runs at 42-140 million events/s on laptop CPU, produces noisier reconstructions than neural networks with ghost images, but sufficient for tasks like fiducial marker detection.

Conclusion: FIBAR provides a simple, efficient CPU-based alternative to neural network methods for event camera image reconstruction, suitable for certain applications despite limitations.

Abstract: Reconstructing an intensity image from the events of a moving event camera is a challenging task that is typically approached with neural networks deployed on graphics processing units. This paper presents a much simpler, FIlter Based Asynchronous Reconstruction method (FIBAR). First, intensity changes signaled by events are integrated with a temporal digital IIR filter. To reduce reconstruction noise, stale pixels are detected by a novel algorithm that regulates a window of recently updated pixels. Arguing that for a moving camera, the absence of events at a pixel location likely implies a low image gradient, stale pixels are then blurred with a Gaussian filter. In contrast to most existing methods, FIBAR is asynchronous and permits image read-out at an arbitrary time. It runs on a modern laptop CPU at about 42(140) million events/s with (without) spatial filtering enabled. A few simple qualitative experiments are presented that show the difference in image reconstruction between FIBAR and a neural network-based approach (FireNet). FIBAR’s reconstruction is noisier than neural network-based methods and suffers from ghost images. However, it is sufficient for certain tasks such as the detection of fiducial markers. Code is available at https://github.com/ros-event-camera/event_image_reconstruction_fibar

Method: Proposes TBTLRR with data-adaptive tensor nuclear norm using learned unitary transforms, bilateral structure for local correlations, and ℓ₁/₂-norm + Frobenius norm regularization for noise handling. Solved via ADMM-based optimization with theoretical convergence.

Result: Extensive experiments demonstrate superiority over state-of-the-art methods in clustering performance.

Conclusion: TBTLRR effectively addresses limitations of existing methods by combining adaptive transforms, bilateral structure, and robust noise handling for superior image clustering.

Abstract: Tensor low-rank representation (TLRR) has demonstrated significant success in image clustering. However, most existing methods rely on fixed transformations and suffer from poor robustness to noise. In this paper, we propose a novel transformed bilateral tensor low-rank representation model called TBTLRR, which introduces a data-adaptive tensor nuclear norm by learning arbitrary unitary transforms, allowing for more effective capture of global correlations. In addition, by leveraging the bilateral structure of latent tensor data, TBTLRR is able to exploit local correlations between image samples and features. Furthermore, TBTLRR integrates the $\ell_{1/2}$-norm and Frobenius norm regularization terms for better dealing with complex noise in real-world scenarios. To solve the proposed nonconvex model, we develop an efficient optimization algorithm inspired by the alternating direction method of multipliers (ADMM) and provide theoretical convergence. Extensive experiments validate its superiority over the state-of-the-art methods in clustering. The code will be available at https://github.com/xianchaoxiu/TBTLRR.

Method: Development followed the software development life cycle with iterative, user-centered feedback. Used internal surveys of clinicians and computer scientists based on usability heuristics to identify requirements, followed by external clinician surveys combining usability heuristics with Technology Acceptance Model constructs.

Result: Initial testing with 4 clinicians and 4 computer scientists reported high usability (4.68/5 and 4.03/5). After refinement, 10 surgeons reported high perceived usefulness (5.07/7), ease of use (5.15/7), heuristic usability (4.38/5), and strong recommendation (9.20/10). Processing time varied with processing mode, video duration, and machine computational power.

Conclusion: Endoshare provides a transparent, user-friendly pipeline for standardized, privacy-preserving surgical video management. Compliance certification and broader interoperability validation are needed to establish it as a deployable alternative to proprietary systems.

Abstract: Video-based assessment and surgical data science can advance surgical training, research, and quality improvement. However, widespread use remains limited by heterogeneous recording formats and privacy concerns associated with video sharing. We present Endoshare, a source-available, cross-platform application for merging, standardizing, and de-identifying endoscopic videos in minimally invasive surgery. Development followed the software development life cycle with iterative, user-centered feedback. During the analysis phase, an internal survey of clinicians and computer scientists based on ten usability heuristics identified key requirements that guided a privacy-by-design architecture. In the testing phase, an external clinician survey combined the same heuristics with Technology Acceptance Model constructs to assess usability and adoption, complemented by benchmarking across different hardware configurations. Four clinicians and four computer scientists initially tested the prototype, reporting high usability (4.68 +/- 0.40/5 and 4.03 +/- 0.51/5), with the lowest score (4.00 +/- 0.93/5) relating to label clarity. After refinement, the testing phase surveyed ten surgeons who reported high perceived usefulness (5.07 +/- 1.75/7), ease of use (5.15 +/- 1.71/7), heuristic usability (4.38 +/- 0.48/5), and strong recommendation (9.20 +/- 0.79/10). Processing time varied with processing mode, video duration (both p <= 0.001), and machine computational power (p = 0.041). Endoshare provides a transparent, user-friendly pipeline for standardized, privacy-preserving surgical video management. Compliance certification and broader interoperability validation are needed to establish it as a deployable alternative to proprietary systems. The software is available at https://camma-public.github.io/Endoshare/

Method: Revealed two key principles from self-attention: adaptive routing (dynamic regulation of positional information flow) and lateral inhibition (token competition suppressing redundancy). Proposed ATConv that reformulates convolution to intrinsically inject these principles while maintaining linear complexity.

Result: ATConv with only 3×3 kernels consistently outperforms various self-attention mechanisms. AttNet (CNN family based on ATConv) achieves 84.4% ImageNet-1K Top-1 accuracy with only 27M parameters. In diffusion models, replacing self-attention with ATConv reduces ImageNet FID by 0.15 with faster sampling.

Conclusion: ATConv successfully bridges the performance gap between convolutions and self-attention by incorporating the core principles that make self-attention effective, while maintaining the efficiency advantages of convolutions.

Abstract: Self-attention (SA) has become the cornerstone of modern vision backbones for its powerful expressivity over traditional Convolutions (Conv). However, its quadratic complexity remains a critical bottleneck for practical applications. Given that Conv offers linear complexity and strong visual priors, continuing efforts have been made to promote the renaissance of Conv. However, a persistent performance chasm remains, highlighting that these modernizations have not yet captured the intrinsic expressivity that defines SA. In this paper, we re-examine the design of the CNNs, directed by a key question: what principles give SA its edge over Conv? As a result, we reveal two fundamental insights that challenge the long-standing design intuitions in prior research (e.g., Receptive field). The two findings are: (1) \textit{Adaptive routing}: SA dynamically regulates positional information flow according to semantic content, whereas Conv employs static kernels uniformly across all positions. (2) \textit{Lateral inhibition}: SA induces score competition among token weighting, effectively suppressing redundancy and sharpening representations, whereas Conv filters lack such inhibitory dynamics and exhibit considerable redundancy. Based on this, we propose \textit{Attentive Convolution} (ATConv), a principled reformulation of the convolutional operator that intrinsically injects these principles. Interestingly, with only $3\times3$ kernels, ATConv consistently outperforms various SA mechanisms in fundamental vision tasks. Building on ATConv, we introduce AttNet, a CNN family that can attain \textbf{84.4%} ImageNet-1K Top-1 accuracy with only 27M parameters. In diffusion-based image generation, replacing all SA with the proposed $3\times 3$ ATConv in SiT-XL/2 reduces ImageNet FID by 0.15 in 400k steps with faster sampling. Code is available at: github.com/price112/Attentive-Convolution.

Method: Fine-tunes variational autoencoder for better latent decoding of sketch characteristics, and integrates a new reinforcement learning reward function based on visual question answering to improve text-image alignment.

Result: Extensive experiments show StableSketcher generates sketches with improved stylistic fidelity and better alignment with prompts compared to Stable Diffusion baseline. Also introduces SketchDUO, the first dataset with instance-level sketches paired with captions and QA pairs.

Conclusion: The proposed framework successfully addresses limitations in sketch generation by diffusion models and provides a valuable dataset resource for future research in this domain.

Abstract: Although recent advancements in diffusion models have significantly enriched the quality of generated images, challenges remain in synthesizing pixel-based human-drawn sketches, a representative example of abstract expression. To combat these challenges, we propose StableSketcher, a novel framework that empowers diffusion models to generate hand-drawn sketches with high prompt fidelity. Within this framework, we fine-tune the variational autoencoder to optimize latent decoding, enabling it to better capture the characteristics of sketches. In parallel, we integrate a new reward function for reinforcement learning based on visual question answering, which improves text-image alignment and semantic consistency. Extensive experiments demonstrate that StableSketcher generates sketches with improved stylistic fidelity, achieving better alignment with prompts compared to the Stable Diffusion baseline. Additionally, we introduce SketchDUO, to the best of our knowledge, the first dataset comprising instance-level sketches paired with captions and question-answer pairs, thereby addressing the limitations of existing datasets that rely on image-label pairs. Our code and dataset will be made publicly available upon acceptance.

Method: Generate synthetic captions using multimodal large language models guided by Wikipedia-derived visual information and taxon-tailored format examples, then train BIOCAP model with these captions.

Result: BIOCAP captures rich semantics and achieves strong performance in species classification and text-image retrieval, demonstrating the value of descriptive captions beyond labels.

Conclusion: Descriptive captions provide valuable supervision for biological multimodal foundation models, bridging biological images with multimodal learning through accurate, instance-based descriptions.

Abstract: This work investigates descriptive captions as an additional source of supervision for biological multimodal foundation models. Images and captions can be viewed as complementary samples from the latent morphospace of a species, each capturing certain biological traits. Incorporating captions during training encourages alignment with this shared latent structure, emphasizing potentially diagnostic characters while suppressing spurious correlations. The main challenge, however, lies in obtaining faithful, instance-specific captions at scale. This requirement has limited the utilization of natural language supervision in organismal biology compared with many other scientific domains. We complement this gap by generating synthetic captions with multimodal large language models (MLLMs), guided by Wikipedia-derived visual information and taxon-tailored format examples. These domain-specific contexts help reduce hallucination and yield accurate, instance-based descriptive captions. Using these captions, we train BIOCAP (i.e., BIOCLIP with Captions), a biological foundation model that captures rich semantics and achieves strong performance in species classification and text-image retrieval. These results demonstrate the value of descriptive captions beyond labels in bridging biological images with multimodal foundation models.

Method: Combines deep learning-based detection with physics-based tracking using kinematics motion equations, plus outlier detection and correction module for handling occlusions and rapid direction changes.

Result: Evaluated on custom racquetball dataset, achieved up to 70% less Average Displacement Error compared to Kalman filter based trackers.

Conclusion: Demonstrates effectiveness of combining physics-based models with deep learning for real-time 3D detection and tracking of challenging small objects, with applications for robot perception on autonomous platforms.

Abstract: While computer vision has advanced considerably for general object detection and tracking, the specific problem of fast-moving tiny objects remains underexplored. This paper addresses the significant challenge of detecting and tracking rapidly moving small objects using an RGB-D camera. Our novel system combines deep learning-based detection with physics-based tracking to overcome the limitations of existing approaches. Our contributions include: (1) a comprehensive system design for object detection and tracking of fast-moving small objects in 3D space, (2) an innovative physics-based tracking algorithm that integrates kinematics motion equations to handle outliers and missed detections, and (3) an outlier detection and correction module that significantly improves tracking performance in challenging scenarios such as occlusions and rapid direction changes. We evaluated our proposed system on a custom racquetball dataset. Our evaluation shows our system surpassing kalman filter based trackers with up to 70% less Average Displacement Error. Our system has significant applications for improving robot perception on autonomous platforms and demonstrates the effectiveness of combining physics-based models with deep learning approaches for real-time 3D detection and tracking of challenging small objects.

Method: A neural rendering pipeline that stores light flow of source rays, computes relationships via cross-attention, predicts target ray colors, and iteratively generates novel views while updating occluded content.

Result: The method works well on various challenging datasets without retraining after synthetic dataset training, outperforming state-of-the-art novel view synthesis methods.

Conclusion: Inverse image-based rendering successfully generates light fields from single images, demonstrating effectiveness across diverse datasets and superior performance compared to existing approaches.

Abstract: A concept of light-fields computed from multiple view images on regular grids has proven its benefit for scene representations, and supported realistic renderings of novel views and photographic effects such as refocusing and shallow depth of field. In spite of its effectiveness of light flow computations, obtaining light fields requires either computational costs or specialized devices like a bulky camera setup and a specialized microlens array. In an effort to broaden its benefit and applicability, in this paper, we propose a novel view synthesis method for light field generation from only single images, named inverse image-based rendering. Unlike previous attempts to implicitly rebuild 3D geometry or to explicitly represent objective scenes, our method reconstructs light flows in a space from image pixels, which behaves in the opposite way to image-based rendering. To accomplish this, we design a neural rendering pipeline to render a target ray in an arbitrary viewpoint. Our neural renderer first stores the light flow of source rays from the input image, then computes the relationships among them through cross-attention, and finally predicts the color of the target ray based on these relationships. After the rendering pipeline generates the first novel view from a single input image, the generated out-of-view contents are updated to the set of source rays. This procedure is iteratively performed while ensuring the consistent generation of occluded contents. We demonstrate that our inverse image-based rendering works well with various challenging datasets without any retraining or finetuning after once trained on synthetic dataset, and outperforms relevant state-of-the-art novel view synthesis methods.

Method: Proposes LogitGap which explicitly uses the relationship between maximum logit and remaining logits, with a training-free strategy to automatically identify the most informative logits for scoring.

Result: Extensive experiments on vision-language and vision-only models show LogitGap consistently achieves state-of-the-art performance across diverse OOD detection scenarios and benchmarks.

Conclusion: LogitGap effectively enhances OOD detection by better utilizing logit space information, providing both theoretical analysis and empirical validation of its effectiveness.

Abstract: Out-of-distribution (OOD) detection is critical for ensuring the reliability of deep learning models in open-world applications. While post-hoc methods are favored for their efficiency and ease of deployment, existing approaches often underexploit the rich information embedded in the model’s logits space. In this paper, we propose LogitGap, a novel post-hoc OOD detection method that explicitly exploits the relationship between the maximum logit and the remaining logits to enhance the separability between in-distribution (ID) and OOD samples. To further improve its effectiveness, we refine LogitGap by focusing on a more compact and informative subset of the logit space. Specifically, we introduce a training-free strategy that automatically identifies the most informative logits for scoring. We provide both theoretical analysis and empirical evidence to validate the effectiveness of our approach. Extensive experiments on both vision-language and vision-only models demonstrate that LogitGap consistently achieves state-of-the-art performance across diverse OOD detection scenarios and benchmarks. Code is available at https://github.com/GIT-LJc/LogitGap.

Method: Created PartNeXt dataset with over 23,000 high-quality textured 3D models annotated with fine-grained hierarchical part labels across 50 categories, using scalable annotation methods and texture-aware labels.

Result: Benchmarking shows state-of-the-art methods struggle with fine-grained part segmentation, and 3D-LLMs have significant gaps in part-centric question answering. Training Point-SAM on PartNeXt yields substantial gains over PartNet.

Conclusion: PartNeXt opens new research avenues for structured 3D understanding by combining scalable annotation, texture-aware labels, and multi-task evaluation, demonstrating superior quality and diversity over existing datasets.

Abstract: Understanding objects at the level of their constituent parts is fundamental to advancing computer vision, graphics, and robotics. While datasets like PartNet have driven progress in 3D part understanding, their reliance on untextured geometries and expert-dependent annotation limits scalability and usability. We introduce PartNeXt, a next-generation dataset addressing these gaps with over 23,000 high-quality, textured 3D models annotated with fine-grained, hierarchical part labels across 50 categories. We benchmark PartNeXt on two tasks: (1) class-agnostic part segmentation, where state-of-the-art methods (e.g., PartField, SAMPart3D) struggle with fine-grained and leaf-level parts, and (2) 3D part-centric question answering, a new benchmark for 3D-LLMs that reveals significant gaps in open-vocabulary part grounding. Additionally, training Point-SAM on PartNeXt yields substantial gains over PartNet, underscoring the dataset’s superior quality and diversity. By combining scalable annotation, texture-aware labels, and multi-task evaluation, PartNeXt opens new avenues for research in structured 3D understanding.

Method: The method performs category-level 8D pose estimation from single RGB images, jointly estimating 8D pose and 3D keypoints of articulated bicycles. It trains with a mix of synthetic and real image data to generalize on real images, adding steering handle and pedal rotation parameters to the standard 6D pose.

Result: The method achieves competitive scores compared to state-of-the-art category-level 6D pose estimators that use rigid canonical object templates, showing promising results in 8D pose parameter accuracy.

Conclusion: The proposed 8D pose estimation method provides more fine-grained bicycle pose state and travel direction estimation than traditional 6D approaches, which is critical for autonomous driving applications involving articulated bicycles.

Abstract: In Autonomous Driving, cyclists belong to the safety-critical class of Vulnerable Road Users (VRU), and accurate estimation of their pose is critical for cyclist crossing intention classification, behavior prediction, and collision avoidance. Unlike rigid objects, articulated bicycles are composed of movable rigid parts linked by joints and constrained by a kinematic structure. 6D pose methods can estimate the 3D rotation and translation of rigid bicycles, but 6D becomes insufficient when the steering/pedals angles of the bicycle vary. That is because: 1) varying the articulated pose of the bicycle causes its 3D bounding box to vary as well, and 2) the 3D box orientation is not necessarily aligned to the orientation of the steering which determines the actual intended travel direction. In this work, we introduce a method for category-level 8D pose estimation for articulated bicycles and cyclists from a single RGB image. Besides being able to estimate the 3D translation and rotation of a bicycle from a single image, our method also estimates the rotations of its steering handles and pedals with respect to the bicycle body frame. These two new parameters enable the estimation of a more fine-grained bicycle pose state and travel direction. Our proposed model jointly estimates the 8D pose and the 3D Keypoints of articulated bicycles, and trains with a mix of synthetic and real image data to generalize on real images. We include an evaluation section where we evaluate the accuracy of our estimated 8D pose parameters, and our method shows promising results by achieving competitive scores when compared against state-of-the-art category-level 6D pose estimators that use rigid canonical object templates for matching.

Method: Proposes accumulating comprehensive knowledge in textual and visual modalities from unsupervised data to update multimodal prototypes, with adaptive update weights, dynamic priority queue for high-confidence images, and multimodal collaborative representation learning for prototype alignment.

Result: Achieves state-of-the-art performance on four benchmark datasets under both closed-world and open-world settings.

Conclusion: The approach effectively overcomes the distribution shift challenge in CZSL by leveraging multimodal knowledge accumulation and adaptive prototype updating.

Abstract: Compositional Zero-Shot Learning (CZSL) aims to recognize novel attribute-object compositions based on the knowledge learned from seen ones. Existing methods suffer from performance degradation caused by the distribution shift of label space at test time, which stems from the inclusion of unseen compositions recombined from attributes and objects. To overcome the challenge, we propose a novel approach that accumulates comprehensive knowledge in both textual and visual modalities from unsupervised data to update multimodal prototypes at test time. Building on this, we further design an adaptive update weight to control the degree of prototype adjustment, enabling the model to flexibly adapt to distribution shift during testing. Moreover, a dynamic priority queue is introduced that stores high-confidence images to acquire visual knowledge from historical images for inference. Considering the semantic consistency of multimodal knowledge, we align textual and visual prototypes by multimodal collaborative representation learning. Extensive experiments indicate that our approach achieves state-of-the-art performance on four benchmark datasets under both closed-world and open-world settings. Code will be available at https://github.com/xud-yan/TOMCAT .

Method: Uses an intermediate stochastic layer in the generator to constrain mutual information between input and output, creating a learnable latent distribution trained jointly with the generator in end-to-end fashion.

Result: Achieves competitive disentanglement scores on dSprites and Color-dSprites datasets compared to state-of-the-art β-VAEs, outperforms InfoGAN, and generates samples with better visual quality and diversity on CelebA and 3D Chairs datasets as measured by FID score.

Conclusion: IB-GAN successfully combines Information Bottleneck with GANs to create disentangled and interpretable latent representations while maintaining high sample quality, demonstrating superiority over both InfoGAN and β-VAEs in various aspects.

Abstract: We propose a new GAN-based unsupervised model for disentangled representation learning. The new model is discovered in an attempt to utilize the Information Bottleneck (IB) framework to the optimization of GAN, thereby named IB-GAN. The architecture of IB-GAN is partially similar to that of InfoGAN but has a critical difference; an intermediate layer of the generator is leveraged to constrain the mutual information between the input and the generated output. The intermediate stochastic layer can serve as a learnable latent distribution that is trained with the generator jointly in an end-to-end fashion. As a result, the generator of IB-GAN can harness the latent space in a disentangled and interpretable manner. With the experiments on dSprites and Color-dSprites dataset, we demonstrate that IB-GAN achieves competitive disentanglement scores to those of state-of-the-art \b{eta}-VAEs and outperforms InfoGAN. Moreover, the visual quality and the diversity of samples generated by IB-GAN are often better than those by \b{eta}-VAEs and Info-GAN in terms of FID score on CelebA and 3D Chairs dataset.

Method: Proposes a two-stage Pick-and-Play Memory (PPM) module inspired by human decision-making: ‘pick’ process selects relevant frames and ‘play’ process adaptively weights them for spatio-temporal aggregation.

Result: Achieves state-of-the-art performance with 0.62/1.11 TEPE on Sintel clean/final (17.3% & 9.02% improvements over BiDAStereo) with fewer computational costs.

Conclusion: PPMStereo effectively addresses the temporal consistency vs computational efficiency trade-off in stereo video depth estimation through its novel memory buffer approach.

Abstract: Temporally consistent depth estimation from stereo video is critical for real-world applications such as augmented reality, where inconsistent depth estimation disrupts the immersion of users. Despite its importance, this task remains challenging due to the difficulty in modeling long-term temporal consistency in a computationally efficient manner. Previous methods attempt to address this by aggregating spatio-temporal information but face a fundamental trade-off: limited temporal modeling provides only modest gains, whereas capturing long-range dependencies significantly increases computational cost. To address this limitation, we introduce a memory buffer for modeling long-range spatio-temporal consistency while achieving efficient dynamic stereo matching. Inspired by the two-stage decision-making process in humans, we propose a \textbf{P}ick-and-\textbf{P}lay \textbf{M}emory (PPM) construction module for dynamic \textbf{Stereo} matching, dubbed as \textbf{PPMStereo}. PPM consists of a pick' process that identifies the most relevant frames and a play’ process that weights the selected frames adaptively for spatio-temporal aggregation. This two-stage collaborative process maintains a compact yet highly informative memory buffer while achieving temporally consistent information aggregation. Extensive experiments validate the effectiveness of PPMStereo, demonstrating state-of-the-art performance in both accuracy and temporal consistency. % Notably, PPMStereo achieves 0.62/1.11 TEPE on the Sintel clean/final (17.3% & 9.02% improvements over BiDAStereo) with fewer computational costs. Codes are available at \textcolor{blue}{https://github.com/cocowy1/PPMStereo}.

Method: Developed evaluation protocol for T2V and I2V models using start frames from established datasets (I2V) and a prompt suite for diverse pedestrian interactions (T2V). Key innovation is reconstructing 2D bird’s-eye view trajectories from pixel-space without known camera parameters.

Result: Leading models have learned surprisingly effective priors for plausible multi-agent behavior, but show failure modes like merging and disappearing people.

Conclusion: The evaluation reveals both strengths and limitations of current video generation models in simulating pedestrian dynamics, highlighting areas for future improvement in multi-agent interaction modeling.

Abstract: Large-scale video generation models have demonstrated high visual realism in diverse contexts, spurring interest in their potential as general-purpose world simulators. Existing benchmarks focus on individual subjects rather than scenes with multiple interacting people. However, the plausibility of multi-agent dynamics in generated videos remains unverified. We propose a rigorous evaluation protocol to benchmark text-to-video (T2V) and image-to-video (I2V) models as implicit simulators of pedestrian dynamics. For I2V, we leverage start frames from established datasets to enable comparison with a ground truth video dataset. For T2V, we develop a prompt suite to explore diverse pedestrian densities and interactions. A key component is a method to reconstruct 2D bird’s-eye view trajectories from pixel-space without known camera parameters. Our analysis reveals that leading models have learned surprisingly effective priors for plausible multi-agent behavior. However, failure modes like merging and disappearing people highlight areas for future improvement.

Method: Proposes Suspicion Progression Analysis Network (SPAN) with suspicion score formula based on Temporal Point Process theory, Suspicion Coefficient Modulation using multimodal information, and Concept-Anchored Mapping to link actions to intention concepts.

Result: SPAN significantly outperforms existing methods on HAI dataset, reducing MSE by 19.8% and improving average mAP by 1.78%, with 2.74% mAP gain in low-frequency cases.

Conclusion: Continuous suspicion modeling enables earlier detection and proactive intervention, greatly enhancing system explainability and practical utility in security applications.

Abstract: Temporal Intention Localization (TIL) is crucial for video surveillance, focusing on identifying varying levels of suspicious intentions to improve security monitoring. However, existing discrete classification methods fail to capture the continuous nature of suspicious intentions, limiting early intervention and explainability. In this paper, we propose the Suspicion Progression Analysis Network (SPAN), which shifts from discrete classification to continuous regression, enabling the capture of fluctuating and evolving suspicious intentions. We reveal that suspicion exhibits long-term dependencies and cumulative effects, similar to Temporal Point Process (TPP) theory. Based on these insights, we define a suspicion score formula that models continuous changes while accounting for temporal characteristics. We also introduce Suspicion Coefficient Modulation, which adjusts suspicion coefficients using multimodal information to reflect the varying impacts of suspicious actions. Additionally, the Concept-Anchored Mapping method is proposed to link suspicious actions to predefined intention concepts, offering insights into both the actions and their potential underlying intentions. Extensive experiments on the HAI dataset show that SPAN significantly outperforms existing methods, reducing MSE by 19.8% and improving average mAP by 1.78%. Notably, SPAN achieves a 2.74% mAP gain in low-frequency cases, demonstrating its superior ability to capture subtle behavioral changes. Compared to discrete classification systems, our continuous suspicion modeling approach enables earlier detection and proactive intervention, greatly enhancing system explainability and practical utility in security applications.

Method: Multi-level analysis including dataset-level characterization of modality composition and disease coverage, image-level quantification of voxel spacing and intensity distributions, and evaluation of preprocessing variability across intensity normalization, bias field correction, skull stripping, spatial registration, and interpolation.

Result: Revealed strong imbalances between large healthy cohorts and smaller clinical populations, substantial heterogeneity in voxel characteristics across datasets, and measurable residual covariate shift after standardized preprocessing that cannot be eliminated by harmonization alone.

Conclusion: Public brain MRI resources exhibit significant variability requiring preprocessing-aware and domain-adaptive strategies for developing generalizable foundation models, as harmonization alone cannot eliminate inter-dataset bias.

Abstract: The development of foundation models for brain MRI depends critically on the scale, diversity, and consistency of available data, yet systematic assessments of these factors remain scarce. In this study, we analyze 54 publicly accessible brain MRI datasets encompassing over 538,031 to provide a structured, multi-level overview tailored to foundation model development. At the dataset level, we characterize modality composition, disease coverage, and dataset scale, revealing strong imbalances between large healthy cohorts and smaller clinical populations. At the image level, we quantify voxel spacing, orientation, and intensity distributions across 15 representative datasets, demonstrating substantial heterogeneity that can influence representation learning. We then perform a quantitative evaluation of preprocessing variability, examining how intensity normalization, bias field correction, skull stripping, spatial registration, and interpolation alter voxel statistics and geometry. While these steps improve within-dataset consistency, residual differences persist between datasets. Finally, feature-space case study using a 3D DenseNet121 shows measurable residual covariate shift after standardized preprocessing, confirming that harmonization alone cannot eliminate inter-dataset bias. Together, these analyses provide a unified characterization of variability in public brain MRI resources and emphasize the need for preprocessing-aware and domain-adaptive strategies in the design of generalizable brain MRI foundation models.

Method: The framework has three stages: 1) RAPO enriches prompts with retrieved modifiers and refactors them to match training distributions; 2) SSPO iteratively refines prompts using multi-source feedback (semantic alignment, spatial fidelity, temporal coherence, optical flow); 3) Fine-tunes LLM using optimized prompt pairs to internalize optimization patterns.

Result: Extensive experiments across five state-of-the-art T2V models and five benchmarks show significant gains in semantic alignment, compositional reasoning, temporal stability, and physical plausibility, outperforming existing methods by large margins.

Conclusion: RAPO++ is a model-agnostic, cost-efficient, and scalable solution that sets a new standard for prompt optimization in T2V generation, achieving substantial improvements without modifying the underlying generative backbone.

Abstract: Prompt design plays a crucial role in text-to-video (T2V) generation, yet user-provided prompts are often short, unstructured, and misaligned with training data, limiting the generative potential of diffusion-based T2V models. We present \textbf{RAPO++}, a cross-stage prompt optimization framework that unifies training-data–aligned refinement, test-time iterative scaling, and large language model (LLM) fine-tuning to substantially improve T2V generation without modifying the underlying generative backbone. In \textbf{Stage 1}, Retrieval-Augmented Prompt Optimization (RAPO) enriches user prompts with semantically relevant modifiers retrieved from a relation graph and refactors them to match training distributions, enhancing compositionality and multi-object fidelity. \textbf{Stage 2} introduces Sample-Specific Prompt Optimization (SSPO), a closed-loop mechanism that iteratively refines prompts using multi-source feedback – including semantic alignment, spatial fidelity, temporal coherence, and task-specific signals such as optical flow – yielding progressively improved video generation quality. \textbf{Stage 3} leverages optimized prompt pairs from SSPO to fine-tune the rewriter LLM, internalizing task-specific optimization patterns and enabling efficient, high-quality prompt generation even before inference. Extensive experiments across five state-of-the-art T2V models and five benchmarks demonstrate that RAPO++ achieves significant gains in semantic alignment, compositional reasoning, temporal stability, and physical plausibility, outperforming existing methods by large margins. Our results highlight RAPO++ as a model-agnostic, cost-efficient, and scalable solution that sets a new standard for prompt optimization in T2V generation. The code is available at https://github.com/Vchitect/RAPO.

Method: FlowCycle proposes an inversion-free, flow-based framework that parameterizes corruption with learnable noises and optimizes them through a cycle-consistent process with dual consistency constraints, iteratively editing source to target and recovering back to source.

Result: Extensive ablations demonstrate that FlowCycle achieves superior editing quality and consistency over state-of-the-art methods.

Conclusion: The target-aware intermediate state approach in FlowCycle enables faithful modifications while preserving source consistency, addressing limitations of current corruption-then-restoration paradigms in text-to-image editing.

Abstract: Recent advances in pre-trained text-to-image flow models have enabled remarkable progress in text-based image editing. Mainstream approaches always adopt a corruption-then-restoration paradigm, where the source image is first corrupted into an ``intermediate state’’ and then restored to the target image under the prompt guidance. However, current methods construct this intermediate state in a target-agnostic manner, i.e., they primarily focus on realizing source image reconstruction while neglecting the semantic gaps towards the specific editing target. This design inherently results in limited editability or inconsistency when the desired modifications substantially deviate from the source. In this paper, we argue that the intermediate state should be target-aware, i.e., selectively corrupting editing-relevant contents while preserving editing-irrelevant ones. To this end, we propose FlowCycle, a novel inversion-free and flow-based editing framework that parameterizes corruption with learnable noises and optimizes them through a cycle-consistent process. By iteratively editing the source to the target and recovering back to the source with dual consistency constraints, FlowCycle learns to produce a target-aware intermediate state, enabling faithful modifications while preserving source consistency. Extensive ablations have demonstrated that FlowCycle achieves superior editing quality and consistency over state-of-the-art methods.

Method: Proposes Contrastive Ultrasound Video Representation Learning (CURL) with dual-contrastive loss (spatial and temporal contrastive learning), task-specific sampling strategy, and probabilistic fine-tuning for flexible inference on long recordings.

Result: Achieved 78.01% sensitivity and 81.60% AUROC on an in-house dataset of 92 subjects with 30-minute ultrasound sessions, demonstrating reliable fetal movement detection.

Conclusion: CURL shows potential for reliable and objective fetal movement analysis, paving the way for improved prenatal monitoring and clinical decision-making through self-supervised contrastive learning.

Abstract: Accurate fetal movement (FM) detection is essential for assessing prenatal health, as abnormal movement patterns can indicate underlying complications such as placental dysfunction or fetal distress. Traditional methods, including maternal perception and cardiotocography (CTG), suffer from subjectivity and limited accuracy. To address these challenges, we propose Contrastive Ultrasound Video Representation Learning (CURL), a novel self-supervised learning framework for FM detection from extended fetal ultrasound video recordings. Our approach leverages a dual-contrastive loss, incorporating both spatial and temporal contrastive learning, to learn robust motion representations. Additionally, we introduce a task-specific sampling strategy, ensuring the effective separation of movement and non-movement segments during self-supervised training, while enabling flexible inference on arbitrarily long ultrasound recordings through a probabilistic fine-tuning approach. Evaluated on an in-house dataset of 92 subjects, each with 30-minute ultrasound sessions, CURL achieves a sensitivity of 78.01% and an AUROC of 81.60%, demonstrating its potential for reliable and objective FM analysis. These results highlight the potential of self-supervised contrastive learning for fetal movement analysis, paving the way for improved prenatal monitoring and clinical decision-making.

Method: Proposes EditInfinity using Infinity (binary-quantized generative model) with efficient image inversion mechanism integrating text prompting rectification and style preservation, plus holistic smoothing strategy.

Result: Superior performance on PIE-Bench benchmark across add, change, and delete operations compared to state-of-the-art diffusion-based baselines.

Conclusion: VQ-based models with exact intermediate representations enable more precise image inversion and editing than diffusion models, achieving better fidelity and semantic alignment.

Abstract: Adapting pretrained diffusion-based generative models for text-driven image editing with negligible tuning overhead has demonstrated remarkable potential. A classical adaptation paradigm, as followed by these methods, first infers the generative trajectory inversely for a given source image by image inversion, then performs image editing along the inferred trajectory guided by the target text prompts. However, the performance of image editing is heavily limited by the approximation errors introduced during image inversion by diffusion models, which arise from the absence of exact supervision in the intermediate generative steps. To circumvent this issue, we investigate the parameter-efficient adaptation of VQ-based generative models for image editing, and leverage their inherent characteristic that the exact intermediate quantized representations of a source image are attainable, enabling more effective supervision for precise image inversion. Specifically, we propose \emph{EditInfinity}, which adapts \emph{Infinity}, a binary-quantized generative model, for image editing. We propose an efficient yet effective image inversion mechanism that integrates text prompting rectification and image style preservation, enabling precise image inversion. Furthermore, we devise a holistic smoothing strategy which allows our \emph{EditInfinity} to perform image editing with high fidelity to source images and precise semantic alignment to the text prompts. Extensive experiments on the PIE-Bench benchmark across “add”, “change”, and “delete” editing operations, demonstrate the superior performance of our model compared to state-of-the-art diffusion-based baselines. Code available at: https://github.com/yx-chen-ust/EditInfinity.

Method: Proposes an “induce-detect-suppress” framework: actively induces hallucinations through designed contexts, uses induced instances for early detection of high-risk cases, and suppresses object-level hallucinations during decoding.

Result: Achieves consistent, significant improvements across all benchmarks, with strong detection capabilities and improved hallucination mitigation.

Conclusion: The study validates that hallucinations in longer responses are caused by context reliance rather than length itself, providing new insights for deeper exploration of LVLM hallucinations.

Abstract: Large Vision-Language Models (LVLMs) have made significant progress in recent years but are also prone to hallucination issues. They exhibit more hallucinations in longer, free-form responses, often attributed to accumulated uncertainties. In this paper, we ask: Does increased hallucination result solely from length-induced errors, or is there a deeper underlying mechanism? After a series of preliminary experiments and findings, we suggest that the risk of hallucinations is not caused by length itself but by the increased reliance on context for coherence and completeness in longer responses. Building on these insights, we propose a novel “induce-detect-suppress” framework that actively induces hallucinations through deliberately designed contexts, leverages induced instances for early detection of high-risk cases, and ultimately suppresses potential object-level hallucinations during actual decoding. Our approach achieves consistent, significant improvements across all benchmarks, demonstrating its efficacy. The strong detection and improved hallucination mitigation not only validate our framework but, more importantly, re-validate our hypothesis on context. Rather than solely pursuing performance gains, this study aims to provide new insights and serves as a first step toward a deeper exploration of hallucinations in LVLMs’ longer responses.

Method: Introduces collaborative field with instance and language fields, uses instance-to-language feature mapping and two-stage training strategy during training, and adaptive language-to-instance prompt refinement during inference.

Result: COS3D achieves leading performance on two widely-used benchmarks and shows high potential for applications like novel image-based 3D segmentation, hierarchical segmentation, and robotics.

Conclusion: The collaborative framework effectively integrates complementary language and segmentation cues, overcoming limitations of existing methods and demonstrating strong performance across various applications.

Abstract: Open-vocabulary 3D segmentation is a fundamental yet challenging task, requiring a mutual understanding of both segmentation and language. However, existing Gaussian-splatting-based methods rely either on a single 3D language field, leading to inferior segmentation, or on pre-computed class-agnostic segmentations, suffering from error accumulation. To address these limitations, we present COS3D, a new collaborative prompt-segmentation framework that contributes to effectively integrating complementary language and segmentation cues throughout its entire pipeline. We first introduce the new concept of collaborative field, comprising an instance field and a language field, as the cornerstone for collaboration. During training, to effectively construct the collaborative field, our key idea is to capture the intrinsic relationship between the instance field and language field, through a novel instance-to-language feature mapping and designing an efficient two-stage training strategy. During inference, to bridge distinct characteristics of the two fields, we further design an adaptive language-to-instance prompt refinement, promoting high-quality prompt-segmentation inference. Extensive experiments not only demonstrate COS3D’s leading performance over existing methods on two widely-used benchmarks but also show its high potential to various applications,~\ie, novel image-based 3D segmentation, hierarchical segmentation, and robotics. The code is publicly available at \href{https://github.com/Runsong123/COS3D}{https://github.com/Runsong123/COS3D}.

Method: Proposes DualGround with token-role-aware cross-modal interaction strategies that align video features with sentence-level and phrase-level semantics in a structurally disentangled manner, and a joint modeling framework that improves both global alignment and fine-grained temporal grounding.

Result: DualGround achieves state-of-the-art performance on both Moment Retrieval and Highlight Detection tasks across QVHighlights and Charades-STA benchmarks.

Conclusion: The effectiveness of disentangled semantic modeling in video-language alignment is demonstrated, showing that capturing both coarse and localized semantics enables more expressive and context-aware video grounding.

Abstract: Video Temporal Grounding (VTG) aims to localize temporal segments in long, untrimmed videos that align with a given natural language query. This task typically comprises two subtasks: Moment Retrieval (MR) and Highlight Detection (HD). While recent advances have been progressed by powerful pretrained vision-language models such as CLIP and InternVideo2, existing approaches commonly treat all text tokens uniformly during crossmodal attention, disregarding their distinct semantic roles. To validate the limitations of this approach, we conduct controlled experiments demonstrating that VTG models overly rely on [EOS]-driven global semantics while failing to effectively utilize word-level signals, which limits their ability to achieve fine-grained temporal alignment. Motivated by this limitation, we propose DualGround, a dual-branch architecture that explicitly separates global and local semantics by routing the [EOS] token through a sentence-level path and clustering word tokens into phrase-level units for localized grounding. Our method introduces (1) tokenrole- aware cross modal interaction strategies that align video features with sentence-level and phrase-level semantics in a structurally disentangled manner, and (2) a joint modeling framework that not only improves global sentence-level alignment but also enhances finegrained temporal grounding by leveraging structured phrase-aware context. This design allows the model to capture both coarse and localized semantics, enabling more expressive and context-aware video grounding. DualGround achieves state-of-the-art performance on both Moment Retrieval and Highlight Detection tasks across QVHighlights and Charades- STA benchmarks, demonstrating the effectiveness of disentangled semantic modeling in video-language alignment.

Method: Proposes mask-based positional encoding using segmentation masks to capture spatial coordinates and object silhouettes, plus a context enhancement module with horizontal/vertical strip convolutional kernels for long-range contextual features.

Result: Achieves state-of-the-art performance on CVOGL and VIGOR-Building datasets with 3.39% improvement in localization accuracy under ground-to-satellite scenarios.

Conclusion: The work provides a robust positional encoding paradigm and contextual modeling framework that advances cross-view geo-localization research by making models object-aware rather than just location-aware.

Abstract: Cross-view object geo-localization enables high-precision object localization through cross-view matching, with critical applications in autonomous driving, urban management, and disaster response. However, existing methods rely on keypoint-based positional encoding, which captures only 2D coordinates while neglecting object shape information, resulting in sensitivity to annotation shifts and limited cross-view matching capability. To address these limitations, we propose a mask-based positional encoding scheme that leverages segmentation masks to capture both spatial coordinates and object silhouettes, thereby upgrading the model from “location-aware” to “object-aware.” Furthermore, to tackle the challenge of large-span objects (e.g., elongated buildings) in satellite imagery, we design a context enhancement module. This module employs horizontal and vertical strip convolutional kernels to extract long-range contextual features, enhancing feature discrimination among strip-like objects. Integrating MPE and CEM, we present EDGeo, an end-to-end framework for robust cross-view object geo-localization. Extensive experiments on two public datasets (CVOGL and VIGOR-Building) demonstrate that our method achieves state-of-the-art performance, with a 3.39% improvement in localization accuracy under challenging ground-to-satellite scenarios. This work provides a robust positional encoding paradigm and a contextual modeling framework for advancing cross-view geo-localization research.

Method: Uses YOLOv8 nano model with custom detection head featuring deep convolutional layers and Squeeze-and-Excitation blocks for enhanced feature extraction, trained on a dataset of 30 currency classes from USD, EUR, and BDT.

Result: Achieved 97.73% accuracy, 95.23% recall, 95.85% f1-score, and 97.21% mAP50(B), with voice feedback for currency identification.

Conclusion: The system provides a practical and efficient solution to empower visually impaired individuals in handling money independently through accurate currency detection and voice feedback.

Abstract: Technologies like smartphones have become an essential in our daily lives. It has made accessible to everyone including visually impaired individuals. With the use of smartphone cameras, image capturing and processing have become more convenient. With the use of smartphones and machine learning, the life of visually impaired can be made a little easier. Daily tasks such as handling money without relying on someone can be troublesome for them. For that purpose this paper presents a real-time currency detection system designed to assist visually impaired individuals. The proposed model is trained on a dataset containing 30 classes of notes and coins, representing 3 types of currency: US dollar (USD), Euro (EUR), and Bangladeshi taka (BDT). Our approach uses a YOLOv8 nano model with a custom detection head featuring deep convolutional layers and Squeeze-and-Excitation blocks to enhance feature extraction and detection accuracy. Our model has achieved a higher accuracy of 97.73%, recall of 95.23%, f1-score of 95.85% and a mean Average Precision at IoU=0.5 (mAP50(B)) of 97.21%. Using the voice feedback after the detection would help the visually impaired to identify the currency. This paper aims to create a practical and efficient currency detection system to empower visually impaired individuals independent in handling money.

Method: Proposes Knowledge-Informed Neural Network (KINN) with ‘compression-aggregation-compression’ architecture: physics-guided compression using dictionary processor, aggregation module, and semantic compression with self-distillation. Available in CNN (0.7M) and Vision Transformer (0.95M) variants.

Result: Establishes state-of-the-art in parameter-efficient recognition on five SAR benchmarks, offering exceptional generalization in data-scarce and out-of-distribution scenarios with tangible interpretability.

Conclusion: KINN provides an effective solution to the representation trilemma and offers a new path for trustworthy AI in SAR image analysis by combining physical priors with neural networks.

Abstract: Deep learning models for complex-valued Synthetic Aperture Radar (CV-SAR) image recognition are fundamentally constrained by a representation trilemma under data-limited and domain-shift scenarios: the concurrent, yet conflicting, optimization of generalization, interpretability, and efficiency. Our work is motivated by the premise that the rich electromagnetic scattering features inherent in CV-SAR data hold the key to resolving this trilemma, yet they are insufficiently harnessed by conventional data-driven models. To this end, we introduce the Knowledge-Informed Neural Network (KINN), a lightweight framework built upon a novel “compression-aggregation-compression” architecture. The first stage performs a physics-guided compression, wherein a novel dictionary processor adaptively embeds physical priors, enabling a compact unfolding network to efficiently extract sparse, physically-grounded signatures. A subsequent aggregation module enriches these representations, followed by a final semantic compression stage that utilizes a compact classification head with self-distillation to learn maximally task-relevant and discriminative embeddings. We instantiate KINN in both CNN (0.7M) and Vision Transformer (0.95M) variants. Extensive evaluations on five SAR benchmarks confirm that KINN establishes a state-of-the-art in parameter-efficient recognition, offering exceptional generalization in data-scarce and out-of-distribution scenarios and tangible interpretability, thereby providing an effective solution to the representation trilemma and offering a new path for trustworthy AI in SAR image analysis.

Method: Two-stage training framework: supervised fine-tuning on synthesized diverse instructions for multi-perspective reasoning, followed by reinforcement learning to optimize pathway selection and composition.

Result: UI-Ins-7B and UI-Ins-32B achieve SOTA on five grounding benchmarks: 87.3% on UI-I2E-Bench, 57.0% on ScreenSpot-Pro, 84.9% on MMBench-GUI L2. UI-Ins-7B achieves 74.1% success rate on AndroidWorld.

Conclusion: The Instruction-as-Reasoning paradigm enables models to dynamically select and compose instruction pathways, demonstrating emergent reasoning and strong agentic potential while mitigating policy collapse in SFT+RL framework.

Abstract: GUI grounding, which maps natural-language instructions to actionable UI elements, is a core capability of GUI agents. Prior works largely treats instructions as a static proxy for user intent, overlooking the impact of instruction diversity and quality on grounding performance. Through a careful investigation of existing grounding datasets, we find a 23.3% flaw rate in their instructions and show that inference-time exploitation of instruction diversity yields up to a substantial 76% relative performance improvement. In this paper, we introduce the Instruction-as-Reasoning paradigm, treating instructions as dynamic analytical pathways that offer distinct perspectives and enabling the model to select the most effective pathway during reasoning. To achieve this, we propose a two-stage training framework: supervised fine-tuning (SFT) on synthesized, diverse instructions to instill multi-perspective reasoning, followed by reinforcement learning (RL) to optimize pathway selection and composition. Our resulting models, UI-Ins-7B and UI-Ins-32B, achieve state-of-the-art results on five challenging grounding benchmarks and exhibit emergent reasoning, selectively composing and synthesizing novel instruction pathways at inference. In particular, UI-Ins-32B attains the best grounding accuracy, scoring 87.3% on UI-I2E-Bench, 57.0% on ScreenSpot-Pro, and 84.9% on MMBench-GUI L2. Furthermore, our model demonstrates strong agentic potential, achieving a 74.1% success rate on AndroidWorld using UI-Ins-7B as the executor. Our in-depth analysis reveals additional insights such as how reasoning can be formulated to enhance rather than hinder grounding performance, and how our method mitigates policy collapse in the SFT+RL framework. All code and model checkpoints will be publicly released in https://github.com/alibaba/UI-Ins.

Method: Evaluated modern video foundation models (both encoder and decoder types) for breakdance video classification, including thorough analysis of finetuned decoder models.

Result: Video encoder models continue to outperform state-of-the-art video language models for prediction tasks in breakdance classification.

Conclusion: Provides guidance on selecting appropriate encoder models and detailed analysis of decoder model performance for niche sports applications.

Abstract: Large Vision Language models have seen huge application in several sports use-cases recently. Most of these works have been targeted towards a limited subset of popular sports like soccer, cricket, basketball etc; focusing on generative tasks like visual question answering, highlight generation. This work analyzes the applicability of the modern video foundation models (both encoder and decoder) for a very niche but hugely popular dance sports - breakdance. Our results show that Video Encoder models continue to outperform state-of-the-art Video Language Models for prediction tasks. We provide insights on how to choose the encoder model and provide a thorough analysis into the workings of a finetuned decoder model for breakdance video classification.

Method: Domain-aligned preprocessing pipeline with platform-wise partitioning, satellite augmentation, and orientation word removal; LLM-based caption refinement; Mixture-of-Experts framework using BGE-M3 (text) and EVA-CLIP (image) with progressive two-stage hard-negative mining training.

Result: System achieved top performance on the official leaderboard, demonstrating robust cross-modal geo-localization under heterogeneous viewpoints.

Conclusion: The proposed approach effectively handles platform heterogeneity and domain gaps in cross-modal drone navigation, providing a robust solution for geo-localization across different platforms.

Abstract: We present a winning solution to RoboSense 2025 Track 4: Cross-Modal Drone Navigation. The task retrieves the most relevant geo-referenced image from a large multi-platform corpus (satellite/drone/ground) given a natural-language query. Two obstacles are severe inter-platform heterogeneity and a domain gap between generic training descriptions and platform-specific test queries. We mitigate these with a domain-aligned preprocessing pipeline and a Mixture-of-Experts (MoE) framework: (i) platform-wise partitioning, satellite augmentation, and removal of orientation words; (ii) an LLM-based caption refinement pipeline to align textual semantics with the distinct visual characteristics of each platform. Using BGE-M3 (text) and EVA-CLIP (image), we train three platform experts using a progressive two-stage, hard-negative mining strategy to enhance discriminative power, and fuse their scores at inference. The system tops the official leaderboard, demonstrating robust cross-modal geo-localization under heterogeneous viewpoints.

Method: HyperET optimizes visual representations to align with textual counterparts using dynamic hyperbolic radius adjustment in hyperbolic space. It employs learnable matrices with Möbius multiplication operations via three configurations: diagonal scaling matrices, block-diagonal matrices, and banded matrices.

Result: Comprehensive experiments across multiple MLLM benchmarks show HyperET consistently improves both pre-training and fine-tuning MLLMs with less than 1% additional parameters.

Conclusion: HyperET provides an efficient training paradigm that effectively addresses the multi-granularity alignment problem in MLLMs using hyperbolic geometry, achieving substantial performance gains with minimal computational overhead.

Abstract: Multi-modal large language models (MLLMs) have emerged as a transformative approach for aligning visual and textual understanding. They typically require extremely high computational resources (e.g., thousands of GPUs) for training to achieve cross-modal alignment at multi-granularity levels. We argue that a key source of this inefficiency lies in the vision encoders they widely equip with, e.g., CLIP and SAM, which lack the alignment with language at multi-granularity levels. To address this issue, in this paper, we leverage hyperbolic space, which inherently models hierarchical levels and thus provides a principled framework for bridging the granularity gap between visual and textual modalities at an arbitrary granularity level. Concretely, we propose an efficient training paradigm for MLLMs, dubbed as HyperET, which can optimize visual representations to align with their textual counterparts at an arbitrary granularity level through dynamic hyperbolic radius adjustment in hyperbolic space. HyperET employs learnable matrices with M"{o}bius multiplication operations, implemented via three effective configurations: diagonal scaling matrices, block-diagonal matrices, and banded matrices, providing a flexible yet efficient parametrization strategy. Comprehensive experiments across multiple MLLM benchmarks demonstrate that HyperET consistently improves both existing pre-training and fine-tuning MLLMs clearly with less than 1% additional parameters.

Method: Uses Universal Context Model with spatial and channel-wise grouping priors, plus Instance-Adaptive Fine-Tuning that fine-tunes a small subset of network weights per instance and includes them in the bitstream.

Result: Sets new state-of-the-art in point cloud compression across 15 diverse datasets, with weight overhead being negligible compared to geometry compression savings.

Conclusion: AnyPcc effectively addresses generalization and OOD data challenges through its universal framework, demonstrating superior compression performance across diverse datasets.

Abstract: Generalization remains a critical challenge for deep learning-based point cloud geometry compression. We argue this stems from two key limitations: the lack of robust context models and the inefficient handling of out-of-distribution (OOD) data. To address both, we introduce AnyPcc, a universal point cloud compression framework. AnyPcc first employs a Universal Context Model that leverages priors from both spatial and channel-wise grouping to capture robust contextual dependencies. Second, our novel Instance-Adaptive Fine-Tuning (IAFT) strategy tackles OOD data by synergizing explicit and implicit compression paradigms. It fine-tunes a small subset of network weights for each instance and incorporates them into the bitstream, where the marginal bit cost of the weights is dwarfed by the resulting savings in geometry compression. Extensive experiments on a benchmark of 15 diverse datasets confirm that AnyPcc sets a new state-of-the-art in point cloud compression. Our code and datasets will be released to encourage reproducible research.

Method: Uses Self-Supervised Learning with facial expression datasets for initial training, followed by refinement with patch-based tri-branch network and inverse explained variance-weighted training loss function.

Result: Achieves 10.9% improvement on Gaze360, 3.8% improvement on MPIIFaceGaze, 11.6% lead on ETH-XGaze subset, with 86.4% accuracy on RAF-DB and 60.9% on Affectnet.

Conclusion: SLYKLatent’s novel components effectively enhance gaze estimation performance and demonstrate strong adaptability across different datasets, surpassing existing methods by significant margins.

Abstract: In this research, we present SLYKLatent, a novel approach for enhancing gaze estimation by addressing appearance instability challenges in datasets due to aleatoric uncertainties, covariant shifts, and test domain generalization. SLYKLatent utilizes Self-Supervised Learning for initial training with facial expression datasets, followed by refinement with a patch-based tri-branch network and an inverse explained variance-weighted training loss function. Our evaluation on benchmark datasets achieves a 10.9% improvement on Gaze360, supersedes top MPIIFaceGaze results with 3.8%, and leads on a subset of ETH-XGaze by 11.6%, surpassing existing methods by significant margins. Adaptability tests on RAF-DB and Affectnet show 86.4% and 60.9% accuracies, respectively. Ablation studies confirm the effectiveness of SLYKLatent’s novel components.

Method: AccuQuant minimizes discrepancies between full-precision and quantized diffusion models within multiple denoising steps, explicitly simulating the sampling process for quantization. Uses efficient implementation reducing memory from O(n) to O(1).

Result: Demonstrated efficacy and efficiency across various tasks and diffusion models on standard benchmarks.

Conclusion: AccuQuant effectively addresses error accumulation in diffusion model quantization through multi-step discrepancy minimization with efficient memory usage.

Abstract: We present in this paper a novel post-training quantization (PTQ) method, dubbed AccuQuant, for diffusion models. We show analytically and empirically that quantization errors for diffusion models are accumulated over denoising steps in a sampling process. To alleviate the error accumulation problem, AccuQuant minimizes the discrepancies between outputs of a full-precision diffusion model and its quantized version within a couple of denoising steps. That is, it simulates multiple denoising steps of a diffusion sampling process explicitly for quantization, accounting the accumulated errors over multiple denoising steps, which is in contrast to previous approaches to imitating a training process of diffusion models, namely, minimizing the discrepancies independently for each step. We also present an efficient implementation technique for AccuQuant, together with a novel objective, which reduces a memory complexity significantly from $\mathcal{O}(n)$ to $\mathcal{O}(1)$, where $n$ is the number of denoising steps. We demonstrate the efficacy and efficiency of AccuQuant across various tasks and diffusion models on standard benchmarks.

Method: Proposed PE-Field that extends positional encodings from 2D plane to structured 3D field, incorporating depth-aware encodings for volumetric reasoning and hierarchical encodings for fine-grained sub-patch control.

Result: PE-Field-augmented DiT achieves state-of-the-art performance on single-image novel view synthesis and generalizes to controllable spatial image editing.

Conclusion: Extending positional encodings to 3D fields enables Diffusion Transformers to better model geometry directly in 3D space, improving their capabilities for 3D-aware generation tasks.

Abstract: Diffusion Transformers (DiTs) have emerged as the dominant architecture for visual generation, powering state-of-the-art image and video models. By representing images as patch tokens with positional encodings (PEs), DiTs combine Transformer scalability with spatial and temporal inductive biases. In this work, we revisit how DiTs organize visual content and discover that patch tokens exhibit a surprising degree of independence: even when PEs are perturbed, DiTs still produce globally coherent outputs, indicating that spatial coherence is primarily governed by PEs. Motivated by this finding, we introduce the Positional Encoding Field (PE-Field), which extends positional encodings from the 2D plane to a structured 3D field. PE-Field incorporates depth-aware encodings for volumetric reasoning and hierarchical encodings for fine-grained sub-patch control, enabling DiTs to model geometry directly in 3D space. Our PE-Field-augmented DiT achieves state-of-the-art performance on single-image novel view synthesis and generalizes to controllable spatial image editing.

Method: Uses ViT-B32 as teacher and MobileNet as student with dynamic fuzzy logic KD. Implements image fusion with Gamma correction, Histogram Equalization, and wavelet-based fusion. Uses Genetic Algorithm for model selection and dynamic wait adjustment for training optimization.

Result: Achieved 99.16% accuracy on LC25000 histopathological images and 99.54% accuracy on IQOTH/NCCD CT-scan images, demonstrating robustness across different imaging modalities.

Conclusion: The proposed FuzzyDistillViT-MobileNet with dynamic fuzzy logic KD effectively handles uncertainty in medical images and achieves state-of-the-art performance in lung cancer classification across multiple imaging domains.

Abstract: This paper presents the FuzzyDistillViT-MobileNet model, a novel approach for lung cancer (LC) classification, leveraging dynamic fuzzy logic-driven knowledge distillation (KD) to address uncertainty and complexity in disease diagnosis. Unlike traditional models that rely on static KD with fixed weights, our method dynamically adjusts the distillation weight using fuzzy logic, enabling the student model to focus on high-confidence regions while reducing attention to ambiguous areas. This dynamic adjustment improves the model ability to handle varying uncertainty levels across different regions of LC images. We employ the Vision Transformer (ViT-B32) as the instructor model, which effectively transfers knowledge to the student model, MobileNet, enhancing the student generalization capabilities. The training process is further optimized using a dynamic wait adjustment mechanism that adapts the training procedure for improved convergence and performance. To enhance image quality, we introduce pixel-level image fusion improvement techniques such as Gamma correction and Histogram Equalization. The processed images (Pix1 and Pix2) are fused using a wavelet-based fusion method to improve image resolution and feature preservation. This fusion method uses the wavedec2 function to standardize images to a 224x224 resolution, decompose them into multi-scale frequency components, and recursively average coefficients at each level for better feature representation. To address computational efficiency, Genetic Algorithm (GA) is used to select the most suitable pre-trained student model from a pool of 12 candidates, balancing model performance with computational cost. The model is evaluated on two datasets, including LC25000 histopathological images (99.16% accuracy) and IQOTH/NCCD CT-scan images (99.54% accuracy), demonstrating robustness across different imaging domains.

Method: Developed Conan framework with (1) Conan-91K dataset of automatically generated reasoning traces including frame identification, evidence reasoning, and action decision, and (2) multi-stage progressive cold-start strategy with Identification-Reasoning-Action (AIR) RLVR training framework.

Result: Surpasses baseline Qwen2.5-VL-7B-Instruct by over 10% accuracy on six multi-step reasoning benchmarks, achieving state-of-the-art performance. Also generalizes effectively to long-video understanding tasks.

Conclusion: Conan provides strong evidence-grounded multi-step video reasoning capabilities with excellent scalability and robustness, addressing key limitations in current video reasoning approaches.

Abstract: Video reasoning, which requires multi-step deduction across frames, remains a major challenge for multimodal large language models (MLLMs). While reinforcement learning (RL)-based methods enhance reasoning capabilities, they often rely on text-only chains that yield ungrounded or hallucinated conclusions. Conversely, frame-retrieval approaches introduce visual grounding but still struggle with inaccurate evidence localization. To address these challenges, we present Conan, a framework for evidence-grounded multi-step video reasoning. Conan identifies contextual and evidence frames, reasons over cross-frame clues, and adaptively decides when to conclude or explore further. To achieve this, we (1) construct Conan-91K, a large-scale dataset of automatically generated reasoning traces that includes frame identification, evidence reasoning, and action decision, and (2) design a multi-stage progressive cold-start strategy combined with an Identification-Reasoning-Action (AIR) RLVR training framework to jointly enhance multi-step visual reasoning. Extensive experiments on six multi-step reasoning benchmarks demonstrate that Conan surpasses the baseline Qwen2.5-VL-7B-Instruct by an average of over 10% in accuracy, achieving state-of-the-art performance. Furthermore, Conan generalizes effectively to long-video understanding tasks, validating its strong scalability and robustness.

Method: A fully reproducible demographic attribute inference pipeline using modular transfer learning with pretrained face recognition encoders and non-linear classification heads. Audited across accuracy, fairness, and a new robustness metric based on intra-identity consistency.

Result: The proposed method outperforms strong baselines, particularly on the more challenging ethnicity attribute. Results show improved performance across multiple datasets and training setups.

Conclusion: The work contributes a reliable foundation for demographic inference in fairness auditing and promotes transparency by releasing dataset metadata, codebase, pretrained models, and evaluation toolkit.

Abstract: Fairness evaluation in face analysis systems (FAS) typically depends on automatic demographic attribute inference (DAI), which itself relies on predefined demographic segmentation. However, the validity of fairness auditing hinges on the reliability of the DAI process. We begin by providing a theoretical motivation for this dependency, showing that improved DAI reliability leads to less biased and lower-variance estimates of FAS fairness. To address this, we propose a fully reproducible DAI pipeline that replaces conventional end-to-end training with a modular transfer learning approach. Our design integrates pretrained face recognition encoders with non-linear classification heads. We audit this pipeline across three dimensions: accuracy, fairness, and a newly introduced notion of robustness, defined via intra-identity consistency. The proposed robustness metric is applicable to any demographic segmentation scheme. We benchmark the pipeline on gender and ethnicity inference across multiple datasets and training setups. Our results show that the proposed method outperforms strong baselines, particularly on ethnicity, which is the more challenging attribute. To promote transparency and reproducibility, we will publicly release the training dataset metadata, full codebase, pretrained models, and evaluation toolkit. This work contributes a reliable foundation for demographic inference in fairness auditing.

Method: Bidirectional concept distillation framework with end-to-end training, shared text encoder, adversarial losses, alignment losses, and bidirectional echoing refinement strategy.

Result: Significantly outperforms existing personalization methods in one-step diffusion personalization setup.

Conclusion: Establishes a novel paradigm for rapid and effective personalization in text-to-image diffusion models through collaborative bidirectional distillation.

Abstract: Recent advances in accelerating text-to-image (T2I) diffusion models have enabled the synthesis of high-fidelity images even in a single step. However, personalizing these models to incorporate novel concepts remains a challenge due to the limited capacity of one-step models to capture new concept distributions effectively. We propose a bidirectional concept distillation framework, EchoDistill, to enable one-step diffusion personalization (1-SDP). Our approach involves an end-to-end training process where a multi-step diffusion model (teacher) and a one-step diffusion model (student) are trained simultaneously. The concept is first distilled from the teacher model to the student, and then echoed back from the student to the teacher. During the EchoDistill, we share the text encoder between the two models to ensure consistent semantic understanding. Following this, the student model is optimized with adversarial losses to align with the real image distribution and with alignment losses to maintain consistency with the teacher’s output. Furthermore, we introduce the bidirectional echoing refinement strategy, wherein the student model leverages its faster generation capability to feedback to the teacher model. This bidirectional concept distillation mechanism not only enhances the student ability to personalize novel concepts but also improves the generative quality of the teacher model. Our experiments demonstrate that this collaborative framework significantly outperforms existing personalization methods over the 1-SDP setup, establishing a novel paradigm for rapid and effective personalization in T2I diffusion models.

Method: The framework structures a dense model into two expert branches: a thinking branch for multi-step reasoning and a non-thinking branch for rapid inference on general tasks like VQA and OCR, with a lightweight router dynamically allocating queries.

Result: Metis-HOME substantially enhances complex reasoning abilities while also improving general capabilities, reversing the degradation trend seen in reasoning-specialized models.

Conclusion: This work establishes a new paradigm for building powerful and versatile MLLMs that effectively resolves the reasoning-vs-generalization dilemma.

Abstract: Inspired by recent advancements in LLM reasoning, the field of multimodal reasoning has seen remarkable progress, achieving significant performance gains on intricate tasks such as mathematical problem-solving. Despite this progress, current multimodal large reasoning models exhibit two key limitations. They tend to employ computationally expensive reasoning even for simple queries, leading to inefficiency. Furthermore, this focus on specialized reasoning often impairs their broader, more general understanding capabilities. In this paper, we propose Metis-HOME: a Hybrid Optimized Mixture-of-Experts framework designed to address this trade-off. Metis-HOME enables a ‘‘Hybrid Thinking’’ paradigm by structuring the original dense model into two distinct expert branches: a thinking branch tailored for complex, multi-step reasoning, and a non-thinking branch optimized for rapid, direct inference on tasks like general VQA and OCR. A lightweight, trainable router dynamically allocates queries to the most suitable expert. We instantiate Metis-HOME by adapting the Qwen2.5-VL-7B into an MoE architecture. Comprehensive evaluations reveal that our approach not only substantially enhances complex reasoning abilities but also improves the model’s general capabilities, reversing the degradation trend observed in other reasoning-specialized models. Our work establishes a new paradigm for building powerful and versatile MLLMs, effectively resolving the prevalent reasoning-vs-generalization dilemma.

Method: Proposes FiFa framework with two main components: 1) FiFa-Annotator using Facial Image Concept Tree for fine-grained regional concept division and reliable data annotation, 2) FiFa-MLLM - a unified multi-task learning architecture supporting multimodal inputs/outputs for Artifact-Grounding Explanation task that generates textual explanations interleaved with segmentation masks.

Result: FiFa-MLLM outperforms strong baselines on the AGE task and achieves state-of-the-art performance on existing XDFA datasets through multiple auxiliary supervision tasks.

Conclusion: The FiFa framework successfully addresses the fine-grained awareness limitation in Explainable DeepFake Analysis by providing reliable data annotation and a unified multimodal model that grounds textual explanations in visual evidence of artifacts.

Abstract: The advancement of Multimodal Large Language Models (MLLMs) has bridged the gap between vision and language tasks, enabling the implementation of Explainable DeepFake Analysis (XDFA). However, current methods suffer from a lack of fine-grained awareness: the description of artifacts in data annotation is unreliable and coarse-grained, and the models fail to support the output of connections between textual forgery explanations and the visual evidence of artifacts, as well as the input of queries for arbitrary facial regions. As a result, their responses are not sufficiently grounded in Face Visual Context (Facext). To address this limitation, we propose the Fake-in-Facext (FiFa) framework, with contributions focusing on data annotation and model construction. We first define a Facial Image Concept Tree (FICT) to divide facial images into fine-grained regional concepts, thereby obtaining a more reliable data annotation pipeline, FiFa-Annotator, for forgery explanation. Based on this dedicated data annotation, we introduce a novel Artifact-Grounding Explanation (AGE) task, which generates textual forgery explanations interleaved with segmentation masks of manipulated artifacts. We propose a unified multi-task learning architecture, FiFa-MLLM, to simultaneously support abundant multimodal inputs and outputs for fine-grained Explainable DeepFake Analysis. With multiple auxiliary supervision tasks, FiFa-MLLM can outperform strong baselines on the AGE task and achieve SOTA performance on existing XDFA datasets. The code and data will be made open-source at https://github.com/lxq1000/Fake-in-Facext.

Method: Uses a Projective Motion Blur Model with differentiable blur creation module, predicts 3D rotation trajectory, and employs model-based restoration network trained end-to-end with post-inference trajectory optimization via reblur loss.

Result: Achieves state-of-the-art performance on both synthetic and real datasets, particularly effective for severe or spatially variant blur where end-to-end deblurring networks struggle.

Conclusion: The proposed modular framework provides interpretable camera motion estimation while enabling reconstruction of sharp image sequences, demonstrating superior performance in challenging blur scenarios.

Abstract: Motion blur caused by camera shake, particularly under large or rotational movements, remains a major challenge in image restoration. We propose a deep learning framework that jointly estimates the latent sharp image and the underlying camera motion trajectory from a single blurry image. Our method leverages the Projective Motion Blur Model (PMBM), implemented efficiently using a differentiable blur creation module compatible with modern networks. A neural network predicts a full 3D rotation trajectory, which guides a model-based restoration network trained end-to-end. This modular architecture provides interpretability by revealing the camera motion that produced the blur. Moreover, this trajectory enables the reconstruction of the sequence of sharp images that generated the observed blurry image. To further refine results, we optimize the trajectory post-inference via a reblur loss, improving consistency between the blurry input and the restored output. Extensive experiments show that our method achieves state-of-the-art performance on both synthetic and real datasets, particularly in cases with severe or spatially variant blur, where end-to-end deblurring networks struggle. Code and trained models are available at https://github.com/GuillermoCarbajal/Blur2Seq/

Method: The framework integrates SuperPoint for robust feature extraction and LightGlue for feature matching, creating a deep learning-enhanced visual SLAM system designed to handle challenging visual conditions.

Result: SELM-SLAM3 outperforms conventional ORB-SLAM3 by an average of 87.84% and exceeds state-of-the-art RGB-D SLAM systems by 36.77%. It demonstrates enhanced performance under challenging conditions like low-texture scenes and fast motion.

Conclusion: The framework provides a reliable platform for developing navigation aids for the visually impaired by significantly improving SLAM performance in challenging visual conditions.

Abstract: Despite advancements in SLAM technologies, robust operation under challenging conditions such as low-texture, motion-blur, or challenging lighting remains an open challenge. Such conditions are common in applications such as assistive navigation for the visually impaired. These challenges undermine localization accuracy and tracking stability, reducing navigation reliability and safety. To overcome these limitations, we present SELM-SLAM3, a deep learning-enhanced visual SLAM framework that integrates SuperPoint and LightGlue for robust feature extraction and matching. We evaluated our framework using TUM RGB-D, ICL-NUIM, and TartanAir datasets, which feature diverse and challenging scenarios. SELM-SLAM3 outperforms conventional ORB-SLAM3 by an average of 87.84% and exceeds state-of-the-art RGB-D SLAM systems by 36.77%. Our framework demonstrates enhanced performance under challenging conditions, such as low-texture scenes and fast motion, providing a reliable platform for developing navigation aids for the visually impaired.

Method: Two-module framework: ACamera-Exposure estimates ISO to fix underexposure, and ACamera-Color predicts color temperature and gain factors for color consistency. Optimized for real-time edge device inference.

Result: Extensive experiments show ACamera-Net consistently enhances image quality and stabilizes perception outputs, outperforming conventional auto modes and lightweight baselines.

Conclusion: The proposed network effectively addresses illumination challenges without additional enhancement modules, providing stable image quality for vision tasks across diverse lighting conditions.

Abstract: Consumer-grade camera systems often struggle to maintain stable image quality under complex illumination conditions such as low light, high dynamic range, and backlighting, as well as spatial color temperature variation. These issues lead to underexposure, color casts, and tonal inconsistency, which degrade the performance of downstream vision tasks. To address this, we propose ACamera-Net, a lightweight and scene-adaptive camera parameter adjustment network that directly predicts optimal exposure and white balance from RAW inputs. The framework consists of two modules: ACamera-Exposure, which estimates ISO to alleviate underexposure and contrast loss, and ACamera-Color, which predicts correlated color temperature and gain factors for improved color consistency. Optimized for real-time inference on edge devices, ACamera-Net can be seamlessly integrated into imaging pipelines. Trained on diverse real-world data with annotated references, the model generalizes well across lighting conditions. Extensive experiments demonstrate that ACamera-Net consistently enhances image quality and stabilizes perception outputs, outperforming conventional auto modes and lightweight baselines without relying on additional image enhancement modules.

Method: Qualitative and quantitative evaluation of four representations (geometric meshes, image-based impostors, NeRFs, 3D Gaussians) at various LoDs and viewing distances.

Result: Each representation shows distinct trade-offs between visual fidelity and computational performance.

Conclusion: The findings provide insights for designing perceptually optimized level-of-detail strategies in crowd rendering.

Abstract: In this paper, we investigate how users perceive the visual quality of crowd character representations at different levels of detail (LoD) and viewing distances. Each representation: geometric meshes, image-based impostors, Neural Radiance Fields (NeRFs), and 3D Gaussians, exhibits distinct trade-offs between visual fidelity and computational performance. Our qualitative and quantitative results provide insights to guide the design of perceptually optimized LoD strategies for crowd rendering.

Method: Proposes EmbodiedBrain framework with agent-aligned data structure, large-scale SFT with Step-Augmented Group Relative Policy Optimization (Step-GRPO) that uses preceding steps as guided precursors, and comprehensive reward system including Generative Reward Model (GRM).

Result: EmbodiedBrain achieves superior performance across all metrics, establishing new state-of-the-art for embodied foundation models, with both 7B and 32B parameter versions.

Conclusion: The work paves the way for next-generation generalist embodied agents by open-sourcing all data, model weights, and evaluation methods, including a novel challenging simulation environment.

Abstract: The realization of Artificial General Intelligence (AGI) necessitates Embodied AI agents capable of robust spatial perception, effective task planning, and adaptive execution in physical environments. However, current large language models (LLMs) and multimodal LLMs (MLLMs) for embodied tasks suffer from key limitations, including a significant gap between model design and agent requirements, an unavoidable trade-off between real-time latency and performance, and the use of unauthentic, offline evaluation metrics. To address these challenges, we propose EmbodiedBrain, a novel vision-language foundation model available in both 7B and 32B parameter sizes. Our framework features an agent-aligned data structure and employs a powerful training methodology that integrates large-scale Supervised Fine-Tuning (SFT) with Step-Augumented Group Relative Policy Optimization (Step-GRPO), which boosts long-horizon task success by integrating preceding steps as Guided Precursors. Furthermore, we incorporate a comprehensive reward system, including a Generative Reward Model (GRM) accelerated at the infrastructure level, to improve training efficiency. For enable thorough validation, we establish a three-part evaluation system encompassing General, Planning, and End-to-End Simulation Benchmarks, highlighted by the proposal and open-sourcing of a novel, challenging simulation environment. Experimental results demonstrate that EmbodiedBrain achieves superior performance across all metrics, establishing a new state-of-the-art for embodied foundation models. Towards paving the way for the next generation of generalist embodied agents, we open-source all of our data, model weight, and evaluating methods, which are available at https://zterobot.github.io/EmbodiedBrain.github.io.

Method: Proposed GenColorBench benchmark grounded in color systems (ISCC-NBS, CSS3/X11) with 44K color-focused prompts covering 400+ colors, including numerical colors absent in other benchmarks. Uses both perceptual and automated assessments.

Result: Evaluation of popular text-to-image models reveals performance variations, showing which color conventions models understand best and identifying specific failure modes in color generation.

Conclusion: GenColorBench provides the first systematic framework for assessing color precision in text-to-image generation and will guide improvements in precise color generation capabilities.

Abstract: Recent years have seen impressive advances in text-to-image generation, with image generative or unified models producing high-quality images from text. Yet these models still struggle with fine-grained color controllability, often failing to accurately match colors specified in text prompts. While existing benchmarks evaluate compositional reasoning and prompt adherence, none systematically assess color precision. Color is fundamental to human visual perception and communication, critical for applications from art to design workflows requiring brand consistency. However, current benchmarks either neglect color or rely on coarse assessments, missing key capabilities such as interpreting RGB values or aligning with human expectations. To this end, we propose GenColorBench, the first comprehensive benchmark for text-to-image color generation, grounded in color systems like ISCC-NBS and CSS3/X11, including numerical colors which are absent elsewhere. With 44K color-focused prompts covering 400+ colors, it reveals models’ true capabilities via perceptual and automated assessments. Evaluations of popular text-to-image models using GenColorBench show performance variations, highlighting which color conventions models understand best and identifying failure modes. Our GenColorBench assessments will guide improvements in precise color generation. The benchmark will be made public upon acceptance.

Method: Learn class-wise prototypes in embedding space with similarity constraints, use dictionaries to store prototypes from different images to prevent class-missing and enable contrastive learning.

Result: Extensive experiments show the method achieves better results than other state-of-the-art methods.

Conclusion: The proposed framework effectively addresses cross-modality segmentation through similarity-based prototypes and contrastive learning.

Abstract: Deep learning models have achieved great success on various vision challenges, but a well-trained model would face drastic performance degradation when applied to unseen data. Since the model is sensitive to domain shift, unsupervised domain adaptation attempts to reduce the domain gap and avoid costly annotation of unseen domains. This paper proposes a novel framework for cross-modality segmentation via similarity-based prototypes. In specific, we learn class-wise prototypes within an embedding space, then introduce a similarity constraint to make these prototypes representative for each semantic class while separable from different classes. Moreover, we use dictionaries to store prototypes extracted from different images, which prevents the class-missing problem and enables the contrastive learning of prototypes, and further improves performance. Extensive experiments show that our method achieves better results than other state-of-the-art methods.

Method: Uses a dual-key memory module combining latent appearance-geometry keys with explicit directional keys, spatial-guided memory readout, and efficient sparsification to fuse current frame features with temporally aggregated object states.

Result: Significantly outperforms state-of-the-art pose-free reconstruction baselines on real-world datasets, with performance improving with more observations while maintaining constant memory and runtime.

Conclusion: OnlineSplatter enables high-quality, object-centric 3D reconstruction from monocular video without pose/depth requirements, with constant computational cost regardless of video length.

Abstract: Free-moving object reconstruction from monocular video remains challenging, particularly without reliable pose or depth cues and under arbitrary object motion. We introduce OnlineSplatter, a novel online feed-forward framework generating high-quality, object-centric 3D Gaussians directly from RGB frames without requiring camera pose, depth priors, or bundle optimization. Our approach anchors reconstruction using the first frame and progressively refines the object representation through a dense Gaussian primitive field, maintaining constant computational cost regardless of video sequence length. Our core contribution is a dual-key memory module combining latent appearance-geometry keys with explicit directional keys, robustly fusing current frame features with temporally aggregated object states. This design enables effective handling of free-moving objects via spatial-guided memory readout and an efficient sparsification mechanism, ensuring comprehensive yet compact object coverage. Evaluations on real-world datasets demonstrate that OnlineSplatter significantly outperforms state-of-the-art pose-free reconstruction baselines, consistently improving with more observations while maintaining constant memory and runtime.

Method: Semantic-Visual Consensus Evidence Selection (SeViCES) with two modules: (1) SVCFS selects frames using temporal-aware semantic reasoning over captions and cluster-guided visual alignment via mutual information, (2) ACR fuses evidence and constrains answer space to resolve inconsistencies.

Result: Extensive experiments on long video benchmarks show SeViCES consistently outperforms state-of-the-art methods in both accuracy and robustness.

Conclusion: Consensus-driven evidence selection is crucial for effective Video-LLMs in long video understanding tasks.

Abstract: Long video understanding remains challenging due to its complex, diverse, and temporally scattered content. Although video large language models (Video-LLMs) can process videos lasting tens of minutes, applying them to truly long sequences is computationally prohibitive and often leads to unfocused or inconsistent reasoning. A promising solution is to select only the most informative frames, yet existing approaches typically ignore temporal dependencies or rely on unimodal evidence, limiting their ability to provide complete and query-relevant context. We propose a Semantic-Visual Consensus Evidence Selection (SeViCES) framework for effective and reliable long video understanding. SeViCES is training-free and model-agnostic, and introduces two key components. The Semantic-Visual Consensus Frame Selection (SVCFS) module selects frames through (1) a temporal-aware semantic branch that leverages LLM reasoning over captions, and (2) a cluster-guided visual branch that aligns embeddings with semantic scores via mutual information. The Answer Consensus Refinement (ACR) module further resolves inconsistencies between semantic- and visual-based predictions by fusing evidence and constraining the answer space. Extensive experiments on long video understanding benchmarks show that SeViCES consistently outperforms state-of-the-art methods in both accuracy and robustness, demonstrating the importance of consensus-driven evidence selection for Video-LLMs.

Method: Systematic review of 260 studies (49 on dental datasets, 211 on DL algorithms), analyzing characteristics of dental imaging, dataset acquisition methods, deep learning techniques, model architectures, optimization strategies, and training methods for different DIA tasks.

Result: The review comprehensively summarizes recent progress in DL-based dental image analysis, including foundational techniques, model categorization by tasks, performance analysis, and commonly used training/evaluation metrics in the DIA domain.

Conclusion: The paper discusses current challenges in existing research and outlines potential future directions, providing a valuable systematic reference for researchers in dental image analysis with publicly available supplementary materials.

Abstract: Efficient analysis and processing of dental images are crucial for dentists to achieve accurate diagnosis and optimal treatment planning. However, dental imaging inherently poses several challenges, such as low contrast, metallic artifacts, and variations in projection angles. Combined with the subjectivity arising from differences in clinicians’ expertise, manual interpretation often proves time-consuming and prone to inconsistency. Artificial intelligence (AI)-based automated dental image analysis (DIA) offers a promising solution to these issues and has become an integral part of computer-aided dental diagnosis and treatment. Among various AI technologies, deep learning (DL) stands out as the most widely applied and influential approach due to its superior feature extraction and representation capabilities. To comprehensively summarize recent progress in this field, we focus on the two fundamental aspects of DL research-datasets and models. In this paper, we systematically review 260 studies on DL applications in DIA, including 49 papers on publicly available dental datasets and 211 papers on DL-based algorithms. We first introduce the basic concepts of dental imaging and summarize the characteristics and acquisition methods of existing datasets. Then, we present the foundational techniques of DL and categorize relevant models and algorithms according to different DIA tasks, analyzing their network architectures, optimization strategies, training methods, and performance. Furthermore, we summarize commonly used training and evaluation metrics in the DIA domain. Finally, we discuss the current challenges of existing research and outline potential future directions. We hope that this work provides a valuable and systematic reference for researchers in this field. All supplementary materials and detailed comparison tables will be made publicly available on GitHub.

Method: A three-stage training curriculum: (i) local reconstruction, (ii) overlapping-window tiling, and (iii) long-context decoder refinement, enabling training on short slice excerpts that generalize to scans exceeding 300 slices without additional memory overhead.

Result: BTB3D improves BLEU scores and increases clinical F1 by 40% over existing methods for report generation, and reduces FID by 75% and halves FVD for text-to-CT synthesis, producing anatomically consistent 512512241 volumes.

Conclusion: Precise three-dimensional tokenization, rather than larger language backbones alone, is essential for scalable vision-language modeling in 3D medical imaging.

Abstract: Recent progress in vision-language modeling for 3D medical imaging has been fueled by large-scale computed tomography (CT) corpora with paired free-text reports, stronger architectures, and powerful pretrained models. This has enabled applications such as automated report generation and text-conditioned 3D image synthesis. Yet, current approaches struggle with high-resolution, long-sequence volumes: contrastive pretraining often yields vision encoders that are misaligned with clinical language, and slice-wise tokenization blurs fine anatomy, reducing diagnostic performance on downstream tasks. We introduce BTB3D (Better Tokens for Better 3D), a causal convolutional encoder-decoder that unifies 2D and 3D training and inference while producing compact, frequency-aware volumetric tokens. A three-stage training curriculum enables (i) local reconstruction, (ii) overlapping-window tiling, and (iii) long-context decoder refinement, during which the model learns from short slice excerpts yet generalizes to scans exceeding 300 slices without additional memory overhead. BTB3D sets a new state-of-the-art on two key tasks: it improves BLEU scores and increases clinical F1 by 40% over CT2Rep, CT-CHAT, and Merlin for report generation; and it reduces FID by 75% and halves FVD compared to GenerateCT and MedSyn for text-to-CT synthesis, producing anatomically consistent 512512241 volumes. These results confirm that precise three-dimensional tokenization, rather than larger language backbones alone, is essential for scalable vision-language modeling in 3D medical imaging. The codebase is available at: https://github.com/ibrahimethemhamamci/BTB3D

Method: Proposes UltraHR-100K dataset with 100K high-quality images >3K resolution, and a frequency-aware post-training method with Detail-Oriented Timestep Sampling (DOTS) and Soft-Weighting Frequency Regularization (SWFR) using Discrete Fourier Transform.

Result: Extensive experiments on UltraHR-eval4K benchmarks show significant improvements in fine-grained detail quality and overall fidelity of UHR image generation.

Conclusion: The proposed dataset and training method effectively enhance ultra-high-resolution text-to-image generation by improving fine-grained detail synthesis and visual fidelity.

Abstract: Ultra-high-resolution (UHR) text-to-image (T2I) generation has seen notable progress. However, two key challenges remain : 1) the absence of a large-scale high-quality UHR T2I dataset, and (2) the neglect of tailored training strategies for fine-grained detail synthesis in UHR scenarios. To tackle the first challenge, we introduce \textbf{UltraHR-100K}, a high-quality dataset of 100K UHR images with rich captions, offering diverse content and strong visual fidelity. Each image exceeds 3K resolution and is rigorously curated based on detail richness, content complexity, and aesthetic quality. To tackle the second challenge, we propose a frequency-aware post-training method that enhances fine-detail generation in T2I diffusion models. Specifically, we design (i) \textit{Detail-Oriented Timestep Sampling (DOTS)} to focus learning on detail-critical denoising steps, and (ii) \textit{Soft-Weighting Frequency Regularization (SWFR)}, which leverages Discrete Fourier Transform (DFT) to softly constrain frequency components, encouraging high-frequency detail preservation. Extensive experiments on our proposed UltraHR-eval4K benchmarks demonstrate that our approach significantly improves the fine-grained detail quality and overall fidelity of UHR image generation. The code is available at \href{https://github.com/NJU-PCALab/UltraHR-100k}{here}.

Method: The model uses a pre-trained ResNet-152 backbone for spatial feature extraction, followed by a Differentiable Soft Self-Organizing Map (Soft-SOM) for topological clustering and interpretability, and a spiking neural head for temporal processing over discrete time steps.

Result: Achieved 97.39% test accuracy on a ten-class waste dataset, outperforming state-of-the-art architectures while maintaining lightweight computational requirements suitable for real-world deployment.

Conclusion: The framework enables precise automated waste segregation, supports recycling efficiency, reduces contamination, and aligns with UN Sustainable Development Goals (SDG 11 and 12) for cleaner cities and circular economy initiatives.

Abstract: Accurate waste classification is vital for achieving sustainable waste management and reducing the environmental footprint of urbanization. Misclassification of recyclable materials contributes to landfill accumulation, inefficient recycling, and increased greenhouse gas emissions. To address these issues, this study introduces HybridSOMSpikeNet, a hybrid deep learning framework that integrates convolutional feature extraction, differentiable self-organization, and spiking-inspired temporal processing to enable intelligent and energy-efficient waste classification. The proposed model employs a pre-trained ResNet-152 backbone to extract deep spatial representations, followed by a Differentiable Soft Self-Organizing Map (Soft-SOM) that enhances topological clustering and interpretability. A spiking neural head accumulates temporal activations over discrete time steps, improving robustness and generalization. Trained on a ten-class waste dataset, HybridSOMSpikeNet achieved a test accuracy of 97.39%, outperforming several state-of-the-art architectures while maintaining a lightweight computational profile suitable for real-world deployment. Beyond its technical innovations, the framework provides tangible environmental benefits. By enabling precise and automated waste segregation, it supports higher recycling efficiency, reduces contamination in recyclable streams, and minimizes the ecological and operational costs of waste processing. The approach aligns with global sustainability priorities, particularly the United Nations Sustainable Development Goals (SDG 11 and SDG 12), by contributing to cleaner cities, circular economy initiatives, and intelligent environmental management systems.

Method: Two main techniques: (1) Weight bias correction to enable shared batch normalization and eliminate fine-tuning by neutralizing quantization-induced bias across bit-widths; (2) Bit-wise coreset sampling strategy that allows each child model to train on compact, informative subsets selected via gradient-based importance scores.

Result: Experiments on CIFAR-10/100, TinyImageNet, and ImageNet-1K with ResNet and ViT architectures show competitive or superior accuracy while reducing training time up to 7.88x compared to existing methods.

Conclusion: The proposed method effectively addresses the training overhead problem in multi-bit quantization networks without compromising model utility, achieving significant speedup while maintaining competitive performance across various datasets and architectures.

Abstract: Multi-bit quantization networks enable flexible deployment of deep neural networks by supporting multiple precision levels within a single model. However, existing approaches suffer from significant training overhead as full-dataset updates are repeated for each supported bit-width, resulting in a cost that scales linearly with the number of precisions. Additionally, extra fine-tuning stages are often required to support additional or intermediate precision options, further compounding the overall training burden. To address this issue, we propose two techniques that greatly reduce the training overhead without compromising model utility: (i) Weight bias correction enables shared batch normalization and eliminates the need for fine-tuning by neutralizing quantization-induced bias across bit-widths and aligning activation distributions; and (ii) Bit-wise coreset sampling strategy allows each child model to train on a compact, informative subset selected via gradient-based importance scores by exploiting the implicit knowledge transfer phenomenon. Experiments on CIFAR-10/100, TinyImageNet, and ImageNet-1K with both ResNet and ViT architectures demonstrate that our method achieves competitive or superior accuracy while reducing training time up to 7.88x. Our code is released at https://github.com/a2jinhee/EMQNet_jk.

Method: Proposed an agent-based architecture combining LLM reasoning with lightweight visual modules, enabling fine-grained analysis and iterative refinement of reasoning chains through a three-stage evaluation framework.

Result: Achieved significant performance gains (+10.3 on MMMU, +6.0 on MathVista over a 7B baseline), matching or surpassing much larger models.

Conclusion: Future visual reasoning models should focus on integrating a broader set of specialized tools for analyzing visual content. The framework and evaluation suite will be released to facilitate future research.

Abstract: Multimodal large language models (MLLMs) that integrate visual and textual reasoning leverage chain-of-thought (CoT) prompting to tackle complex visual tasks, yet continue to exhibit visual hallucinations and an over-reliance on textual priors. We present a systematic diagnosis of state-of-the-art vision-language models using a three-stage evaluation framework, uncovering key failure modes. To address these, we propose an agent-based architecture that combines LLM reasoning with lightweight visual modules, enabling fine-grained analysis and iterative refinement of reasoning chains. Our results highlight future visual reasoning models should focus on integrating a broader set of specialized tools for analyzing visual content. Our system achieves significant gains (+10.3 on MMMU, +6.0 on MathVista over a 7B baseline), matching or surpassing much larger models. We will release our framework and evaluation suite to facilitate future research.

Method: MixKV adapts to head-wise semantic redundancy by selectively balancing diversity and importance when compressing KV pairs, addressing the limitations of importance-only approaches.

Result: Under extreme compression (budget=64), MixKV improves baseline methods by 5.1% across five multi-modal benchmarks and achieves 8.0-9.0% gains for GUI grounding tasks while maintaining comparable inference efficiency.

Conclusion: MixKV provides an effective KV cache compression method that balances importance and diversity, demonstrating consistent performance improvements across LVLMs and extending well to LLMs with comparable gains.

Abstract: Recent large vision-language models (LVLMs) demonstrate remarkable capabilities in processing extended multi-modal sequences, yet the resulting key-value (KV) cache expansion creates a critical memory bottleneck that fundamentally limits deployment scalability. While existing KV cache compression methods focus on retaining high-importance KV pairs to minimize storage, they often overlook the modality-specific semantic redundancy patterns that emerge distinctively in multi-modal KV caches. In this work, we first analyze how, beyond simple importance, the KV cache in LVLMs exhibits varying levels of redundancy across attention heads. We show that relying solely on importance can only cover a subset of the full KV cache information distribution, leading to potential loss of semantic coverage. To address this, we propose \texttt{MixKV}, a novel method that mixes importance with diversity for optimized KV cache compression in LVLMs. \texttt{MixKV} adapts to head-wise semantic redundancy, selectively balancing diversity and importance when compressing KV pairs. Extensive experiments demonstrate that \texttt{MixKV} consistently enhances existing methods across multiple LVLMs. Under extreme compression (budget=64), \texttt{MixKV} improves baseline methods by an average of \textbf{5.1%} across five multi-modal understanding benchmarks and achieves remarkable gains of \textbf{8.0%} and \textbf{9.0%} for SnapKV and AdaKV on GUI grounding tasks, all while maintaining comparable inference efficiency. Furthermore, \texttt{MixKV} extends seamlessly to LLMs with comparable performance gains. Our code is available at \href{https://github.com/xuyang-liu16/MixKV}{\textcolor{citeblue}{https://github.com/xuyang-liu16/MixKV}}.

Method: The algorithm automatically reverse-engineers the intrinsic geometry of any spinning LiDAR sensor by inferring critical parameters including laser beam configuration, angular distributions, and per-beam calibration corrections.

Result: Comprehensive evaluation shows perfect point preservation with zero points lost across all point clouds, maintaining geometric accuracy within sensor precision limits with real-time performance.

Conclusion: This represents a paradigm shift from approximate to lossless LiDAR projections, opening new possibilities for high-precision remote sensing applications requiring complete geometric preservation.

Abstract: 3D LiDAR sensors are essential for autonomous navigation, environmental monitoring, and precision mapping in remote sensing applications. To efficiently process the massive point clouds generated by these sensors, LiDAR data is often projected into 2D range images that organize points by their angular positions and distances. While these range image representations enable efficient processing, conventional projection methods suffer from fundamental geometric inconsistencies that cause irreversible information loss, compromising high-fidelity applications. We present ALICE-LRI (Automatic LiDAR Intrinsic Calibration Estimation for Lossless Range Images), the first general, sensor-agnostic method that achieves lossless range image generation from spinning LiDAR point clouds without requiring manufacturer metadata or calibration files. Our algorithm automatically reverse-engineers the intrinsic geometry of any spinning LiDAR sensor by inferring critical parameters including laser beam configuration, angular distributions, and per-beam calibration corrections, enabling lossless projection and complete point cloud reconstruction with zero point loss. Comprehensive evaluation across the complete KITTI and DurLAR datasets demonstrates that ALICE-LRI achieves perfect point preservation, with zero points lost across all point clouds. Geometric accuracy is maintained well within sensor precision limits, establishing geometric losslessness with real-time performance. We also present a compression case study that validates substantial downstream benefits, demonstrating significant quality improvements in practical applications. This paradigm shift from approximate to lossless LiDAR projections opens new possibilities for high-precision remote sensing applications requiring complete geometric preservation.

Method: Uses an RGB-D diffusion model to iteratively generate sparse keyframes with geometric consistency, then interpolates between them using a video diffusion model. Key techniques include shared latent space for image and depth, explicit conditioning on existing scene geometry, and warp-consistent guidance.

Result: Generates realistic driving videos over 20 seconds with significant improvements: 48.6% better long-horizon FID and 43.0% better FVD scores compared to prior state-of-the-art methods.

Conclusion: AutoScape successfully generates long-horizon driving scenes with strong geometric consistency and visual quality, demonstrating substantial improvements over existing approaches.

Abstract: This paper proposes AutoScape, a long-horizon driving scene generation framework. At its core is a novel RGB-D diffusion model that iteratively generates sparse, geometrically consistent keyframes, serving as reliable anchors for the scene’s appearance and geometry. To maintain long-range geometric consistency, the model 1) jointly handles image and depth in a shared latent space, 2) explicitly conditions on the existing scene geometry (i.e., rendered point clouds) from previously generated keyframes, and 3) steers the sampling process with a warp-consistent guidance. Given high-quality RGB-D keyframes, a video diffusion model then interpolates between them to produce dense and coherent video frames. AutoScape generates realistic and geometrically consistent driving videos of over 20 seconds, improving the long-horizon FID and FVD scores over the prior state-of-the-art by 48.6% and 43.0%, respectively.

Method: Proposed a unified dual-encoder model with attention-driven feature fusion that combines convolutional neural networks (CNNs) and vision transformers (ViTs) for semantic segmentation.

Result: Achieved μIoU/μDice scores of 76.79%/86.87% on GCPS dataset and 64.93%/76.60% on PUMA dataset, outperforming state-of-the-art benchmarks.

Conclusion: The attention-driven fusion of CNN and ViT features within a dual-encoder framework significantly improves semantic segmentation performance in histopathological image analysis.

Abstract: Automated histopathological image analysis plays a vital role in computer-aided diagnosis of various diseases. Among developed algorithms, deep learning-based approaches have demonstrated excellent performance in multiple tasks, including semantic tissue segmentation in histological images. In this study, we propose a novel approach based on attention-driven feature fusion of convolutional neural networks (CNNs) and vision transformers (ViTs) within a unified dual-encoder model to improve semantic segmentation performance. Evaluation on two publicly available datasets showed that our model achieved {\mu}IoU/{\mu}Dice scores of 76.79%/86.87% on the GCPS dataset and 64.93%/76.60% on the PUMA dataset, outperforming state-of-the-art and baseline benchmarks. The implementation of our method is publicly available in a GitHub repository: https://github.com/NimaTorbati/ACS-SegNet

Method: DyPE dynamically adjusts the model’s positional encoding at each diffusion step to match the frequency spectrum with the current generative stage, leveraging the spectral progression inherent in diffusion where low-frequency structures converge early and high-frequencies take more steps.

Result: DyPE enables generation of images at resolutions far beyond training data (e.g., 16 million pixels using FLUX) with no additional sampling cost, achieving state-of-the-art fidelity in ultra-high-resolution image generation with performance gains increasing at higher resolutions.

Conclusion: DyPE provides an effective training-free solution for extending diffusion transformers to ultra-high resolutions by dynamically adapting positional encodings to the diffusion process’s spectral characteristics.

Abstract: Diffusion Transformer models can generate images with remarkable fidelity and detail, yet training them at ultra-high resolutions remains extremely costly due to the self-attention mechanism’s quadratic scaling with the number of image tokens. In this paper, we introduce Dynamic Position Extrapolation (DyPE), a novel, training-free method that enables pre-trained diffusion transformers to synthesize images at resolutions far beyond their training data, with no additional sampling cost. DyPE takes advantage of the spectral progression inherent to the diffusion process, where low-frequency structures converge early, while high-frequencies take more steps to resolve. Specifically, DyPE dynamically adjusts the model’s positional encoding at each diffusion step, matching their frequency spectrum with the current stage of the generative process. This approach allows us to generate images at resolutions that exceed the training resolution dramatically, e.g., 16 million pixels using FLUX. On multiple benchmarks, DyPE consistently improves performance and achieves state-of-the-art fidelity in ultra-high-resolution image generation, with gains becoming even more pronounced at higher resolutions. Project page is available at https://noamissachar.github.io/DyPE/.

Method: SV combines multiple lightweight draft experts (small VLMs) that generate diverse reasoning paths, and a strong verdict model that synthesizes these paths. It includes consensus expert selection to forward only high-agreement paths to the verdict stage.

Result: SV achieves consistent gains on challenging benchmarks including InfographicVQA, ChartMuseum, ChartQAPro, and HR-Bench 4K, providing both error correction and cost-efficiency compared to large proprietary models.

Conclusion: The framework successfully synthesizes correct insights from multiple partially accurate reasoning paths, achieving improved performance on information-intensive visual reasoning tasks while maintaining computational efficiency.

Abstract: Large Vision-Language Models (VLMs) have achieved remarkable progress in multimodal understanding, yet they struggle when reasoning over information-intensive images that densely interleave textual annotations with fine-grained graphical elements. The main challenges lie in precisely localizing critical cues in dense layouts and multi-hop reasoning to integrate dispersed evidence. We propose Speculative Verdict (SV), a training-free framework inspired by speculative decoding that combines multiple lightweight draft experts with a large verdict model. In the draft stage, small VLMs act as draft experts to generate reasoning paths that provide diverse localization candidates; in the verdict stage, a strong VLM synthesizes these paths to produce the final answer, minimizing computational cost while recovering correct answers. To further improve efficiency and accuracy, SV introduces a consensus expert selection mechanism that forwards only high-agreement reasoning paths to the verdict. Empirically, SV achieves consistent gains on challenging information-intensive and high-resolution visual question answering benchmarks, including InfographicVQA, ChartMuseum, ChartQAPro, and HR-Bench 4K. By synthesizing correct insights from multiple partially accurate reasoning paths, SV achieves both error correction and cost-efficiency compared to large proprietary models or training pipelines. Code is available at https://github.com/Tinaliu0123/speculative-verdict

Method: Introduces α-Flow, a family of objectives that unifies trajectory flow matching, Shortcut Model, and MeanFlow, using a curriculum strategy to smoothly transition from trajectory flow matching to MeanFlow.

Result: α-Flow consistently outperforms MeanFlow across scales and settings on ImageNet-1K 256x256, achieving SOTA FID scores of 2.58 (1-NFE) and 2.15 (2-NFE) with vanilla DiT backbones.

Conclusion: The curriculum-based α-Framework effectively resolves optimization conflicts in MeanFlow, enabling better convergence and superior performance in few-step generative modeling.

Abstract: MeanFlow has recently emerged as a powerful framework for few-step generative modeling trained from scratch, but its success is not yet fully understood. In this work, we show that the MeanFlow objective naturally decomposes into two parts: trajectory flow matching and trajectory consistency. Through gradient analysis, we find that these terms are strongly negatively correlated, causing optimization conflict and slow convergence. Motivated by these insights, we introduce $\alpha$-Flow, a broad family of objectives that unifies trajectory flow matching, Shortcut Model, and MeanFlow under one formulation. By adopting a curriculum strategy that smoothly anneals from trajectory flow matching to MeanFlow, $\alpha$-Flow disentangles the conflicting objectives, and achieves better convergence. When trained from scratch on class-conditional ImageNet-1K 256x256 with vanilla DiT backbones, $\alpha$-Flow consistently outperforms MeanFlow across scales and settings. Our largest $\alpha$-Flow-XL/2+ model achieves new state-of-the-art results using vanilla DiT backbones, with FID scores of 2.58 (1-NFE) and 2.15 (2-NFE).

Method: Two-stage flow matching pipeline: (1) coarse stage for initial 3D geometry and pose recovery, (2) refinement stage integrating pose-aligned image features to enhance structural fidelity and appearance details. Represents camera poses and 3D shape as distributions in shared 3D latent space.

Result: Outperforms leading 3D reconstruction methods with over 3 dB PSNR gain and over 10% Chamfer Distance reduction. Matches monocular estimators on pose accuracy and delivers superior visual fidelity over baseline 3D generative models.

Conclusion: Cupid provides an effective unified framework for single-image 3D reconstruction that achieves state-of-the-art performance in shape accuracy, pose estimation, and visual quality.

Abstract: This work proposes a new generation-based 3D reconstruction method, named Cupid, that accurately infers the camera pose, 3D shape, and texture of an object from a single 2D image. Cupid casts 3D reconstruction as a conditional sampling process from a learned distribution of 3D objects, and it jointly generates voxels and pixel-voxel correspondences, enabling robust pose and shape estimation under a unified generative framework. By representing both input camera poses and 3D shape as a distribution in a shared 3D latent space, Cupid adopts a two-stage flow matching pipeline: (1) a coarse stage that produces initial 3D geometry with associated 2D projections for pose recovery; and (2) a refinement stage that integrates pose-aligned image features to enhance structural fidelity and appearance details. Extensive experiments demonstrate Cupid outperforms leading 3D reconstruction methods with an over 3 dB PSNR gain and an over 10% Chamfer Distance reduction, while matching monocular estimators on pose accuracy and delivering superior visual fidelity over baseline 3D generative models. For an immersive view of the 3D results generated by Cupid, please visit cupid3d.github.io.

Method: Developed a radar-camera fusion MOT framework with online calibration using common features, feature matching, and category-consistency checking to associate detections from both sensors more accurately than position matching alone.

Result: The framework demonstrated improved tracking precision and streamlined radar-camera mapping process in real-world experiments conducted in both controlled environments and actual traffic scenarios.

Conclusion: This is the first work to investigate radar-camera common features integration and online calibration for MOT, successfully enhancing tracking efficiency while minimizing manual interventions through effective sensor fusion.

Abstract: This paper presents a Multi-Object Tracking (MOT) framework that fuses radar and camera data to enhance tracking efficiency while minimizing manual interventions. Contrary to many studies that underutilize radar and assign it a supplementary role–despite its capability to provide accurate range/depth information of targets in a world 3D coordinate system–our approach positions radar in a crucial role. Meanwhile, this paper utilizes common features to enable online calibration to autonomously associate detections from radar and camera. The main contributions of this work include: (1) the development of a radar-camera fusion MOT framework that exploits online radar-camera calibration to simplify the integration of detection results from these two sensors, (2) the utilization of common features between radar and camera data to accurately derive real-world positions of detected objects, and (3) the adoption of feature matching and category-consistency checking to surpass the limitations of mere position matching in enhancing sensor association accuracy. To the best of our knowledge, we are the first to investigate the integration of radar-camera common features and their use in online calibration for achieving MOT. The efficacy of our framework is demonstrated by its ability to streamline the radar-camera mapping process and improve tracking precision, as evidenced by real-world experiments conducted in both controlled environments and actual traffic scenarios. Code is available at https://github.com/radar-lab/Radar_Camera_MOT

Method: Proposes a segmentation framework based on image generation where MLLM outputs visual tokens that are detokenized into images using universal VQ-VAE, with next-scale-prediction strategy for parallel token generation to reduce inference latency.

Result: Outperforms prior state-of-the-art approaches on multiple segmentation datasets with significant inference speed improvements while maintaining strong multimodal understanding capabilities.

Conclusion: ARGenSeg successfully integrates image segmentation into MLLMs through generation-based paradigm, achieving superior performance and efficiency compared to existing methods.

Abstract: We propose a novel AutoRegressive Generation-based paradigm for image Segmentation (ARGenSeg), achieving multimodal understanding and pixel-level perception within a unified framework. Prior works integrating image segmentation into multimodal large language models (MLLMs) typically employ either boundary points representation or dedicated segmentation heads. These methods rely on discrete representations or semantic prompts fed into task-specific decoders, which limits the ability of the MLLM to capture fine-grained visual details. To address these challenges, we introduce a segmentation framework for MLLM based on image generation, which naturally produces dense masks for target objects. We leverage MLLM to output visual tokens and detokenize them into images using an universal VQ-VAE, making the segmentation fully dependent on the pixel-level understanding of the MLLM. To reduce inference latency, we employ a next-scale-prediction strategy to generate required visual tokens in parallel. Extensive experiments demonstrate that our method surpasses prior state-of-the-art approaches on multiple segmentation datasets with a remarkable boost in inference speed, while maintaining strong understanding capabilities.

Method: Simple end-to-end pure transformer model for autoregressive video prediction using continuous pixel-space representations, comparing various spatiotemporal self-attention layouts, trained unsupervised on physical simulation datasets.

Result: Significantly extends time horizon for physically accurate predictions by up to 50% compared to existing latent-space approaches, while maintaining comparable performance on common video quality metrics. Also enables accurate estimation of PDE simulation parameters.

Conclusion: The work provides a platform for attention-based spatiotemporal video modeling through a simple, parameter-efficient, and interpretable approach that generalizes well to out-of-distribution simulation parameters.

Abstract: Inspired by the performance and scalability of autoregressive large language models (LLMs), transformer-based models have seen recent success in the visual domain. This study investigates a transformer adaptation for video prediction with a simple end-to-end approach, comparing various spatiotemporal self-attention layouts. Focusing on causal modeling of physical simulations over time; a common shortcoming of existing video-generative approaches, we attempt to isolate spatiotemporal reasoning via physical object tracking metrics and unsupervised training on physical simulation datasets. We introduce a simple yet effective pure transformer model for autoregressive video prediction, utilizing continuous pixel-space representations for video prediction. Without the need for complex training strategies or latent feature-learning components, our approach significantly extends the time horizon for physically accurate predictions by up to 50% when compared with existing latent-space approaches, while maintaining comparable performance on common video quality metrics. In addition, we conduct interpretability experiments to identify network regions that encode information useful to perform accurate estimations of PDE simulation parameters via probing models, and find that this generalizes to the estimation of out-of-distribution simulation parameters. This work serves as a platform for further attention-based spatiotemporal modeling of videos via a simple, parameter efficient, and interpretable approach.

Method: Uses an unmixing bottleneck: extracts endmember signatures from HSI, predicts abundance maps from MSI using MLP, reconstructs spectra via linear mixing, and trains self-supervised using MSI’s spectral response function.

Result: Outperforms state-of-the-art unsupervised/self-supervised baselines and remains competitive with supervised methods, even with single-band MSI. Trains in under a minute.

Conclusion: SpectraMorph provides interpretable hyperspectral super-resolution that is robust, fast-training, and effective even with limited MSI bands, bridging the gap between performance and interpretability.

Abstract: Hyperspectral sensors capture dense spectra per pixel but suffer from low spatial resolution, causing blurred boundaries and mixed-pixel effects. Co-registered companion sensors such as multispectral, RGB, or panchromatic cameras provide high-resolution spatial detail, motivating hyperspectral super-resolution through the fusion of hyperspectral and multispectral images (HSI-MSI). Existing deep learning based methods achieve strong performance but rely on opaque regressors that lack interpretability and often fail when the MSI has very few bands. We propose SpectraMorph, a physics-guided self-supervised fusion framework with a structured latent space. Instead of direct regression, SpectraMorph enforces an unmixing bottleneck: endmember signatures are extracted from the low-resolution HSI, and a compact multilayer perceptron predicts abundance-like maps from the MSI. Spectra are reconstructed by linear mixing, with training performed in a self-supervised manner via the MSI sensor’s spectral response function. SpectraMorph produces interpretable intermediates, trains in under a minute, and remains robust even with a single-band (pan-chromatic) MSI. Experiments on synthetic and real-world datasets show SpectraMorph consistently outperforming state-of-the-art unsupervised/self-supervised baselines while remaining very competitive against supervised baselines.

Method: Uses latent-variable extension of Denoising Diffusion Bridge Models with shared latent space, contrastive alignment loss for semantic consistency, domain-agnostic encoder-decoder architecture, and predictive loss for cross-domain translation.

Result: Performs strongly on diverse MT tasks including multi-view to 3D shape generation, image super-resolution, and multi-view scene synthesis, establishing a new strong baseline.

Conclusion: LDDBM provides an effective general framework for modality translation that supports arbitrary modality pairs and overcomes limitations of existing approaches.

Abstract: Recent advances in generative modeling have positioned diffusion models as state-of-the-art tools for sampling from complex data distributions. While these models have shown remarkable success across single-modality domains such as images and audio, extending their capabilities to Modality Translation (MT), translating information across different sensory modalities, remains an open challenge. Existing approaches often rely on restrictive assumptions, including shared dimensionality, Gaussian source priors, and modality-specific architectures, which limit their generality and theoretical grounding. In this work, we propose the Latent Denoising Diffusion Bridge Model (LDDBM), a general-purpose framework for modality translation based on a latent-variable extension of Denoising Diffusion Bridge Models. By operating in a shared latent space, our method learns a bridge between arbitrary modalities without requiring aligned dimensions. We introduce a contrastive alignment loss to enforce semantic consistency between paired samples and design a domain-agnostic encoder-decoder architecture tailored for noise prediction in latent space. Additionally, we propose a predictive loss to guide training toward accurate cross-domain translation and explore several training strategies to improve stability. Our approach supports arbitrary modality pairs and performs strongly on diverse MT tasks, including multi-view to 3D shape generation, image super-resolution, and multi-view scene synthesis. Comprehensive experiments and ablations validate the effectiveness of our framework, establishing a new strong baseline in general modality translation. For more information, see our project page: https://sites.google.com/view/lddbm/home.

Method: Introduces a layered canvas where each subject is on a distinct layer for occlusion-free composition, and a locking mechanism that preserves selected layers while allowing others to adapt to context. Uses positional embeddings and complementary data sampling without architectural changes.

Result: Extensive experiments show LayerComposer achieves superior spatial control and identity preservation compared to state-of-the-art methods in multi-subject personalized image generation.

Conclusion: LayerComposer provides an interactive framework that enables intuitive layer manipulation similar to professional image-editing software, successfully addressing limitations in spatial composition control and multi-subject scaling.

Abstract: Despite their impressive visual fidelity, existing personalized generative models lack interactive control over spatial composition and scale poorly to multiple subjects. To address these limitations, we present LayerComposer, an interactive framework for personalized, multi-subject text-to-image generation. Our approach introduces two main contributions: (1) a layered canvas, a novel representation in which each subject is placed on a distinct layer, enabling occlusion-free composition; and (2) a locking mechanism that preserves selected layers with high fidelity while allowing the remaining layers to adapt flexibly to the surrounding context. Similar to professional image-editing software, the proposed layered canvas allows users to place, resize, or lock input subjects through intuitive layer manipulation. Our versatile locking mechanism requires no architectural changes, relying instead on inherent positional embeddings combined with a new complementary data sampling strategy. Extensive experiments demonstrate that LayerComposer achieves superior spatial control and identity preservation compared to the state-of-the-art methods in multi-subject personalized image generation.

Method: Uses Window Cross-Attention for precise directorial control by localizing text prompts to specific shots, and Sparse Inter-Shot Self-Attention (dense within shots, sparse between them) for efficient minute-scale generation.

Result: Sets new state-of-the-art in narrative coherence and develops emergent abilities including persistent memory for characters/scenes and intuitive grasp of cinematic techniques.

Conclusion: Marks a pivotal shift from clip synthesis towards automated filmmaking, making end-to-end cinematic creation a tangible future.

Abstract: State-of-the-art text-to-video models excel at generating isolated clips but fall short of creating the coherent, multi-shot narratives, which are the essence of storytelling. We bridge this “narrative gap” with HoloCine, a model that generates entire scenes holistically to ensure global consistency from the first shot to the last. Our architecture achieves precise directorial control through a Window Cross-Attention mechanism that localizes text prompts to specific shots, while a Sparse Inter-Shot Self-Attention pattern (dense within shots but sparse between them) ensures the efficiency required for minute-scale generation. Beyond setting a new state-of-the-art in narrative coherence, HoloCine develops remarkable emergent abilities: a persistent memory for characters and scenes, and an intuitive grasp of cinematic techniques. Our work marks a pivotal shift from clip synthesis towards automated filmmaking, making end-to-end cinematic creation a tangible future. Our code is available at: https://holo-cine.github.io/.

Method: Uses second-order digital IIR filters for per-pixel brightness reconstruction, analyzes falling edges and zero-level crossings for more accurate period estimation, and implements as an open-source ROS node.

Result: The algorithm is more robust to high-frequency noise than baseline methods, can detect frequencies up to 64kHz for single pixels, and runs efficiently on laptop CPUs at over 50M events/second.

Conclusion: Hardware implementation closer to the sensor is needed for improved full-sensor frequency imaging due to bandwidth limitations, and the open-source implementation produces results comparable to proprietary solutions.

Abstract: Due to their high temporal resolution and large dynamic range, event cameras are uniquely suited for the analysis of time-periodic signals in an image. In this work we present an efficient and fully asynchronous event camera algorithm for detecting the fundamental frequency at which image pixels flicker. The algorithm employs a second-order digital infinite impulse response (IIR) filter to perform an approximate per-pixel brightness reconstruction and is more robust to high-frequency noise than the baseline method we compare to. We further demonstrate that using the falling edge of the signal leads to more accurate period estimates than the rising edge, and that for certain signals interpolating the zero-level crossings can further increase accuracy. Our experiments find that the outstanding capabilities of the camera in detecting frequencies up to 64kHz for a single pixel do not carry over to full sensor imaging as readout bandwidth limitations become a serious obstacle. This suggests that a hardware implementation closer to the sensor will allow for greatly improved frequency imaging. We discuss the important design parameters for fullsensor frequency imaging and present Frequency Cam, an open-source implementation as a ROS node that can run on a single core of a laptop CPU at more than 50 million events per second. It produces results that are qualitatively very similar to those obtained from the closed source vibration analysis module in Prophesee’s Metavision Toolkit. The code for Frequency Cam and a demonstration video can be found at https://github.com/ros-event-camera/frequency_cam

Method: Proposes Tex-ViT (Texture-Vision Transformer) that enhances CNN features by combining ResNet with a vision transformer. Uses a texture module that operates in parallel on ResNet sections before down-sampling, then feeds into a dual branch cross-attention vision transformer to capture global texture correlations.

Result: Achieved 98% accuracy in cross-domain scenarios on FF++, DFDCPreview, and Celeb-DF datasets. Model performed well under various post-processing conditions (blurring, compression, noise) and demonstrated superior generalization compared to state-of-the-art models.

Conclusion: Tex-ViT effectively learns shared distinguishing textural characteristics in manipulated samples, making it applicable to various situations and resistant to post-processing procedures. The hybrid approach successfully addresses limitations of traditional CNNs and pure vision transformers.

Abstract: Deepfakes, which employ GAN to produce highly realistic facial modification, are widely regarded as the prevailing method. Traditional CNN have been able to identify bogus media, but they struggle to perform well on different datasets and are vulnerable to adversarial attacks due to their lack of robustness. Vision transformers have demonstrated potential in the realm of image classification problems, but they require enough training data. Motivated by these limitations, this publication introduces Tex-ViT (Texture-Vision Transformer), which enhances CNN features by combining ResNet with a vision transformer. The model combines traditional ResNet features with a texture module that operates in parallel on sections of ResNet before each down-sampling operation. The texture module then serves as an input to the dual branch of the cross-attention vision transformer. It specifically focuses on improving the global texture module, which extracts feature map correlation. Empirical analysis reveals that fake images exhibit smooth textures that do not remain consistent over long distances in manipulations. Experiments were performed on different categories of FF++, such as DF, f2f, FS, and NT, together with other types of GAN datasets in cross-domain scenarios. Furthermore, experiments also conducted on FF++, DFDCPreview, and Celeb-DF dataset underwent several post-processing situations, such as blurring, compression, and noise. The model surpassed the most advanced models in terms of generalization, achieving a 98% accuracy in cross-domain scenarios. This demonstrates its ability to learn the shared distinguishing textural characteristics in the manipulated samples. These experiments provide evidence that the proposed model is capable of being applied to various situations and is resistant to many post-processing procedures.

Method: Introduces Residual Kolmogorov-Arnold Network (RKAN) as a plug-in module that can be added to any stage of traditional deep networks to integrate supportive polynomial feature transformations.

Result: RKAN offers consistent improvements over baseline models across different vision tasks and benchmarks, achieving cutting-edge performance.

Conclusion: RKAN provides an efficient solution to enhance CNN performance while addressing computational inefficiency and optimization challenges in deep networks.

Abstract: Despite their immense success, deep convolutional neural networks (CNNs) can be difficult to optimize and costly to train due to hundreds of layers within the network depth. Conventional convolutional operations are fundamentally limited by their linear nature along with fixed activations, where many layers are needed to learn meaningful patterns in data. Because of the sheer size of these networks, this approach is simply computationally inefficient, and poses overfitting or gradient explosion risks, especially in small datasets. As a result, we introduce a “plug-in” module, called Residual Kolmogorov-Arnold Network (RKAN). Our module is highly compact, so it can be easily added into any stage (level) of traditional deep networks, where it learns to integrate supportive polynomial feature transformations to existing convolutional frameworks. RKAN offers consistent improvements over baseline models in different vision tasks and widely tested benchmarks, accomplishing cutting-edge performance on them.

Method: The framework distills graphics engine capabilities into a video-generation model (Stable Video Diffusion) through fine-tuning on a small synthetic dataset. Users can directly insert and manipulate point lights in 3D scenes within images to generate video sequences.

Result: The model generalizes well to real-world scenes despite being trained only on synthetic data, producing plausible shadows and inter-reflections. It demonstrates that video foundation models can capture rich lighting, material, and shape information.

Conclusion: Video foundation models with minimal training can perform realistic relighting without explicit asset reconstruction or ray-tracing, representing a new paradigm for scene illumination manipulation.

Abstract: Manipulating the illumination of a 3D scene within a single image represents a fundamental challenge in computer vision and graphics. This problem has traditionally been addressed using inverse rendering techniques, which involve explicit 3D asset reconstruction and costly ray-tracing simulations. Meanwhile, recent advancements in visual foundation models suggest that a new paradigm could soon be possible – one that replaces explicit physical models with networks that are trained on large amounts of image and video data. In this paper, we exploit the implicit scene understanding of a video diffusion model, particularly Stable Video Diffusion, to relight a single image. We introduce GenLit, a framework that distills the ability of a graphics engine to perform light manipulation into a video-generation model, enabling users to directly insert and manipulate a point light in the 3D world within a given image and generate results directly as a video sequence. We find that a model fine-tuned on only a small synthetic dataset generalizes to real-world scenes, enabling single-image relighting with plausible and convincing shadows and inter-reflections. Our results highlight the ability of video foundation models to capture rich information about lighting, material, and shape, and our findings indicate that such models, with minimal training, can be used to perform relighting without explicit asset reconstruction or ray-tracing. . Project page: https://genlit.is.tue.mpg.de/.

Method: FairGen uses attribute adapters that learn latent directions via noise composition in a self-discovering process, and a distribution indicator that multiplies with adapters to guide generation towards prescribed distributions.

Result: Extensive experiments show FairGen outperforms previous state-of-the-art methods by a large margin in debiasing gender, racial, and intersectional biases.

Conclusion: FairGen provides an effective, lightweight solution for debiasing multiple attributes simultaneously in Diffusion Models without requiring retraining or reference datasets.

Abstract: While Diffusion Models (DM) exhibit remarkable performance across various image generative tasks, they nonetheless reflect the inherent bias presented in the training set. As DMs are now widely used in real-world applications, these biases could perpetuate a distorted worldview and hinder opportunities for minority groups. Existing methods on debiasing DMs usually requires model retraining with a human-crafted reference dataset or additional classifiers, which suffer from two major limitations: (1) collecting reference datasets causes expensive annotation cost; (2) the debiasing performance is heavily constrained by the quality of the reference dataset or the additional classifier. To address the above limitations, we propose FairGen, a plug-and-play method that learns attribute latent directions in a self-discovering manner, thus eliminating the reliance on such reference dataset. Specifically, FairGen consists of two parts: a set of attribute adapters and a distribution indicator. Each adapter in the set aims to learn an attribute latent direction, and is optimized via noise composition through a self-discovering process. Then, the distribution indicator is multiplied by the set of adapters to guide the generation process towards the prescribed distribution. Our method enables debiasing multiple attributes in DMs simultaneously, while remaining lightweight and easily integrable with other DMs, eliminating the need for retraining. Extensive experiments on debiasing gender, racial, and their intersectional biases show that our method outperforms previous SOTA by a large margin.

Method: Proposed a hierarchical ability taxonomy, collected images and annotations from public datasets, built a semi-automatic data pipeline to create problems, and evaluated 25 mainstream MLLMs on the benchmark.

Result: The benchmark includes development and test sets with 1800 problems each, supporting both English and Chinese. Evaluations examined ability correlations, target position impact, CoT prompting effects, and identified abilities needing specialist model supplementation.

Conclusion: Face-Human-Bench provides a comprehensive evaluation framework for face and human understanding in MLLMs, with publicly available dataset and evaluation code to advance multi-modal assistant capabilities.

Abstract: Faces and humans are crucial elements in social interaction and are widely included in everyday photos and videos. Therefore, a deep understanding of faces and humans will enable multi-modal assistants to achieve improved response quality and broadened application scope. Currently, the multi-modal assistant community lacks a comprehensive and scientific evaluation of face and human understanding abilities. In this paper, we first propose a hierarchical ability taxonomy that includes three levels of abilities. Then, based on this taxonomy, we collect images and annotations from publicly available datasets in the face and human community and build a semi-automatic data pipeline to produce problems for the new benchmark. Finally, the obtained Face-Human-Bench includes a development set and a test set, each with 1800 problems, supporting both English and Chinese. We conduct evaluations over 25 mainstream multi-modal large language models (MLLMs) with our Face-Human-Bench, focusing on the correlation between abilities, the impact of the relative position of targets on performance, and the impact of Chain of Thought (CoT) prompting on performance. We also explore which abilities of MLLMs need to be supplemented by specialist models. The dataset and evaluation code have been made publicly available at https://face-human-bench.github.io.

Method: Augment 2D videos with 3D point trajectories aligned in pixel space to create PointVid dataset, fine-tune latent diffusion models to track 2D objects with 3D coordinates, and regularize object shape and motion to eliminate artifacts.

Result: The method enhances quality of generated RGB videos, alleviates common issues like object morphing, and enables handling of contact-rich scenarios where 3D perception is crucial.

Conclusion: The proposed 3D augmentation and regularization framework can be seamlessly integrated into existing video diffusion models to improve their visual plausibility and physical realism.

Abstract: We present a novel video generation framework that integrates 3-dimensional geometry and dynamic awareness. To achieve this, we augment 2D videos with 3D point trajectories and align them in pixel space. The resulting 3D-aware video dataset, PointVid, is then used to fine-tune a latent diffusion model, enabling it to track 2D objects with 3D Cartesian coordinates. Building on this, we regularize the shape and motion of objects in the video to eliminate undesired artifacts, e.g., non-physical deformation. Consequently, we enhance the quality of generated RGB videos and alleviate common issues like object morphing, which are prevalent in current video models due to a lack of shape awareness. With our 3D augmentation and regularization, our model is capable of handling contact-rich scenarios such as task-oriented videos, where 3D information is essential for perceiving shape and motion of interacting solids. Our method can be seamlessly integrated into existing video diffusion models to improve their visual plausibility.

Method: Represent each ground image pixel as a 3D Gaussian with semantic/spatial features, synthesize BEV feature maps for pose estimation, and use icosphere-based supervision for panoramic queries.

Result: Significant improvement in localization accuracy on KITTI and VIGOR datasets compared to prior approaches, handling both pinhole and panoramic query images.

Conclusion: Feature-based Gaussian primitives effectively resolve height ambiguity in cross-view localization, providing superior performance over existing methods.

Abstract: This paper addresses the problem of weakly supervised cross-view localization, where the goal is to estimate the pose of a ground camera relative to a satellite image with noisy ground truth annotations. A common approach to bridge the cross-view domain gap for pose estimation is Bird’s-Eye View (BEV) synthesis. However, existing methods struggle with height ambiguity due to the lack of depth information in ground images and satellite height maps. Previous solutions either assume a flat ground plane or rely on complex models, such as cross-view transformers. We propose BevSplat, a novel method that resolves height ambiguity by using feature-based Gaussian primitives. Each pixel in the ground image is represented by a 3D Gaussian with semantic and spatial features, which are synthesized into a BEV feature map for relative pose estimation. Additionally, to address challenges with panoramic query images, we introduce an icosphere-based supervision strategy for the Gaussian primitives. We validate our method on the widely used KITTI and VIGOR datasets, which include both pinhole and panoramic query images. Experimental results show that BevSplat significantly improves localization accuracy over prior approaches.

Method: A semi-automated pipeline using fine-tuned YOLOv8 detector and ByteTrack, followed by manual correction, was used to create the dataset. The study benchmarks 28 object detectors, 23 identification models, and tracking methods.

Result: Object detectors show near-perfect performance on lenient IoU (mAP50: 95.2-98.9%) but significant divergence on stricter metrics (mAP50:95: 56.5-66.4%). Smaller architectures like ConvNextV2 Nano achieve the best balance for identification (73.35% accuracy, 50.82% Top-1 KNN). Tracking methods achieve high detection accuracy (MOTA > 0.92) but struggle with identity preservation (IDF1 ≈ 0.27).

Conclusion: The 8-Calves dataset bridges a gap by providing temporal richness and realistic challenges, serving as a resource for advancing agricultural vision models.

Abstract: Automated livestock monitoring is crucial for precision farming, but robust computer vision models are hindered by a lack of datasets reflecting real-world group challenges. We introduce the 8-Calves dataset, a challenging benchmark for multi-animal detection, tracking, and identification. It features a one-hour video of eight Holstein Friesian calves in a barn, with frequent occlusions, motion blur, and diverse poses. A semi-automated pipeline using a fine-tuned YOLOv8 detector and ByteTrack, followed by manual correction, provides over 537,000 bounding boxes with temporal identity labels. We benchmark 28 object detectors, showing near-perfect performance on a lenient IoU threshold (mAP50: 95.2-98.9%) but significant divergence on stricter metrics (mAP50:95: 56.5-66.4%), highlighting fine-grained localization challenges. Our identification benchmark across 23 models reveals a trade-off: scaling model size improves classification accuracy but compromises retrieval. Smaller architectures like ConvNextV2 Nano achieve the best balance (73.35% accuracy, 50.82% Top-1 KNN). Pre-training focused on semantic learning (e.g., BEiT) yielded superior transferability. For tracking, leading methods achieve high detection accuracy (MOTA > 0.92) but struggle with identity preservation (IDF1 $\approx$ 0.27), underscoring a key challenge in occlusion-heavy scenarios. The 8-Calves dataset bridges a gap by providing temporal richness and realistic challenges, serving as a resource for advancing agricultural vision models. The dataset and code are available at https://huggingface.co/datasets/tonyFang04/8-calves.

Method: Developed OpenMIBOOD framework with three benchmarks from diverse medical domains, categorizing 14 datasets into covariate-shifted in-distribution, near-OOD, and far-OOD categories, and evaluated 24 post-hoc OOD detection methods.

Result: Findings from broad-scale OOD benchmarks in natural image domains do not translate to medical applications, highlighting the critical need for medical-specific benchmarks.

Conclusion: OpenMIBOOD supports the advancement of reliable and trustworthy AI systems in healthcare by mitigating risks of exposing models to inputs outside their training distribution, with the framework available as an open repository.

Abstract: The growing reliance on Artificial Intelligence (AI) in critical domains such as healthcare demands robust mechanisms to ensure the trustworthiness of these systems, especially when faced with unexpected or anomalous inputs. This paper introduces the Open Medical Imaging Benchmarks for Out-Of-Distribution Detection (OpenMIBOOD), a comprehensive framework for evaluating out-of-distribution (OOD) detection methods specifically in medical imaging contexts. OpenMIBOOD includes three benchmarks from diverse medical domains, encompassing 14 datasets divided into covariate-shifted in-distribution, near-OOD, and far-OOD categories. We evaluate 24 post-hoc methods across these benchmarks, providing a standardized reference to advance the development and fair comparison of OOD detection methods. Results reveal that findings from broad-scale OOD benchmarks in natural image domains do not translate to medical applications, underscoring the critical need for such benchmarks in the medical field. By mitigating the risk of exposing AI models to inputs outside their training distribution, OpenMIBOOD aims to support the advancement of reliable and trustworthy AI systems in healthcare. The repository is available at https://github.com/remic-othr/OpenMIBOOD.

Method: The dataset was created by recording high-resolution LiDAR, camera, and pose data in rural areas during rain conditions, providing diverse and information-rich data for challenging scenarios.

Result: The paper presents analysis of the recorded data and provides baseline results for panoptic segmentation, semantic segmentation, and 3D occupancy prediction methods on LiDAR point clouds.

Conclusion: Panoptic-CUDAL fills an important gap by providing a specialized dataset for developing and testing autonomous driving systems in challenging rural environments with adverse weather conditions.

Abstract: Existing autonomous driving datasets are predominantly oriented towards well-structured urban settings and favourable weather conditions, leaving the complexities of rural environments and adverse weather conditions largely unaddressed. Although some datasets encompass variations in weather and lighting, bad weather scenarios do not appear often. Rainfall can significantly impair sensor functionality, introducing noise and reflections in LiDAR and camera data and reducing the system’s capabilities for reliable environmental perception and safe navigation. This paper introduces the Panoptic-CUDAL dataset, a novel dataset purpose-built for panoptic segmentation in rural areas subject to rain. By recording high-resolution LiDAR, camera, and pose data, Panoptic-CUDAL offers a diverse, information-rich dataset in a challenging scenario. We present the analysis of the recorded data and provide baseline results for panoptic, semantic segmentation, and 3D occupancy prediction methods on LiDAR point clouds. The dataset can be found here: https://robotics.sydney.edu.au/our-research/intelligent-transportation-systems, https://vision.rwth-aachen.de/panoptic-cudal

Method: Developed a degraded imaging model based on Retinex theory and atmospheric scattering, and created a prompt-modulated restoration network with text encoder for user specifications and spatial-frequency visual adapter for autonomous degradation perception.

Result: Extensive experiments show ControlFusion outperforms state-of-the-art methods in fusion quality and degradation handling, especially for real-world and compound degradations at various levels.

Conclusion: ControlFusion effectively addresses composite degradations in real-world imaging through its controllable framework with language-vision prompts, demonstrating superior performance over existing methods.

Abstract: Current image fusion methods struggle to address the composite degradations encountered in real-world imaging scenarios and lack the flexibility to accommodate user-specific requirements. In response to these challenges, we propose a controllable image fusion framework with language-vision prompts, termed ControlFusion, which adaptively neutralizes composite degradations. On the one hand, we develop a degraded imaging model that integrates physical imaging mechanisms, including the Retinex theory and atmospheric scattering principle, to simulate composite degradations, thereby providing potential for addressing real-world complex degradations from the data level. On the other hand, we devise a prompt-modulated restoration and fusion network that dynamically enhances features with degradation prompts, enabling our method to accommodate composite degradation of varying levels. Specifically, considering individual variations in quality perception of users, we incorporate a text encoder to embed user-specified degradation types and severity levels as degradation prompts. We also design a spatial-frequency collaborative visual adapter that autonomously perceives degradations in source images, thus eliminating the complete dependence on user instructions. Extensive experiments demonstrate that ControlFusion outperforms SOTA fusion methods in fusion quality and degradation handling, particularly in countering real-world and compound degradations with various levels. The source code is publicly available at https://github.com/Linfeng-Tang/ControlFusion.

Method: Uses semantic matching and position-constrained attention fusion to transfer visual details from reference subjects, plus a novel noise initialization strategy to preserve geometry priors for robust feature matching.

Result: Significantly outperforms existing zero-shot and training-free approaches in both subject fidelity and text alignment, and can extend to multi-subject generation.

Conclusion: FreeGraftor enables precise subject identity transfer while maintaining text-aligned scene synthesis without model fine-tuning, making it practical for real-world deployment.

Abstract: Subject-driven image generation aims to synthesize novel scenes that faithfully preserve subject identity from reference images while adhering to textual guidance. However, existing methods struggle with a critical trade-off between fidelity and efficiency. Tuning-based approaches rely on time-consuming and resource-intensive, subject-specific optimization, while zero-shot methods often fail to maintain adequate subject consistency. In this work, we propose FreeGraftor, a training-free framework that addresses these limitations through cross-image feature grafting. Specifically, FreeGraftor leverages semantic matching and position-constrained attention fusion to transfer visual details from reference subjects to the generated images. Additionally, our framework introduces a novel noise initialization strategy to preserve the geometry priors of reference subjects, facilitating robust feature matching. Extensive qualitative and quantitative experiments demonstrate that our method enables precise subject identity transfer while maintaining text-aligned scene synthesis. Without requiring model fine-tuning or additional training, FreeGraftor significantly outperforms existing zero-shot and training-free approaches in both subject fidelity and text alignment. Furthermore, our framework can seamlessly extend to multi-subject generation, making it practical for real-world deployment. Our code is available at https://github.com/Nihukat/FreeGraftor.

Method: Proposes balanced contact sampling to fairly represent diverse contact statistics, and vertex-level class-balanced (VCB) loss that reweights loss contribution per vertex based on contact frequency to address spatial imbalance.

Result: The framework effectively learns dense hand contact estimation from large-scale data without suffering from class and spatial imbalance issues.

Conclusion: HACO successfully addresses key challenges in hand contact estimation through novel sampling and loss strategies, enabling robust learning from imbalanced datasets.

Abstract: Hands are essential to human interaction, and exploring contact between hands and the world can promote comprehensive understanding of their function. Recently, there have been growing number of hand interaction datasets that cover interaction with object, other hand, scene, and body. Despite the significance of the task and increasing high-quality data, how to effectively learn dense hand contact estimation remains largely underexplored. There are two major challenges for learning dense hand contact estimation. First, there exists class imbalance issue from hand contact datasets where majority of regions are not in contact. Second, hand contact datasets contain spatial imbalance issue with most of hand contact exhibited in finger tips, resulting in challenges for generalization towards contacts in other hand regions. To tackle these issues, we present a framework that learns dense HAnd COntact estimation (HACO) from imbalanced data. To resolve the class imbalance issue, we introduce balanced contact sampling, which builds and samples from multiple sampling groups that fairly represent diverse contact statistics for both contact and non-contact vertices. Moreover, to address the spatial imbalance issue, we propose vertex-level class-balanced (VCB) loss, which incorporates spatially varying contact distribution by separately reweighting loss contribution of each vertex based on its contact frequency across dataset. As a result, we effectively learn to predict dense hand contact estimation with large-scale hand contact data without suffering from class and spatial imbalance issue. The codes are available at https://github.com/dqj5182/HACO_RELEASE.

Method: Developed a full-pipeline toolkit specifically designed for concept erasure and conducted systematic study of NSFW concept erasure methods by examining the interplay between mechanisms and empirical observations.

Result: The study provides in-depth insights and practical guidance for effective application of concept erasure methods in real-world scenarios.

Conclusion: This work advances understanding of content safety in diffusion models and establishes a foundation for future research and development in NSFW content prevention.

Abstract: Text-to-image diffusion models have gained widespread application across various domains, demonstrating remarkable creative potential. However, the strong generalization capabilities of diffusion models can inadvertently lead to the generation of not-safe-for-work (NSFW) content, posing significant risks to their safe deployment. While several concept erasure methods have been proposed to mitigate the issue associated with NSFW content, a comprehensive evaluation of their effectiveness across various scenarios remains absent. To bridge this gap, we introduce a full-pipeline toolkit specifically designed for concept erasure and conduct the first systematic study of NSFW concept erasure methods. By examining the interplay between the underlying mechanisms and empirical observations, we provide in-depth insights and practical guidance for the effective application of concept erasure methods in various real-world scenarios, with the aim of advancing the understanding of content safety in diffusion models and establishing a solid foundation for future research and development in this critical area.

Method: Proposes Mesh-RFT framework using Masked Direct Preference Optimization (M-DPO) with quality-aware face masking for localized refinement. Introduces objective topology-aware scoring system with Boundary Edge Ratio (BER) and Topology Score (TS) metrics to evaluate geometric integrity and topological regularity at both object and face levels.

Result: M-DPO reduces Hausdorff Distance (HD) by 24.6% and improves Topology Score (TS) by 3.8% over pre-trained models. Outperforms global DPO methods with 17.4% HD reduction and 4.9% TS gain.

Conclusion: Mesh-RFT is the first method to optimize mesh quality at individual face granularity, resolving localized errors while preserving global coherence, achieving state-of-the-art performance in production-ready mesh generation.

Abstract: Existing pretrained models for 3D mesh generation often suffer from data biases and produce low-quality results, while global reinforcement learning (RL) methods rely on object-level rewards that struggle to capture local structure details. To address these challenges, we present Mesh-RFT, a novel fine-grained reinforcement fine-tuning framework that employs Masked Direct Preference Optimization (M-DPO) to enable localized refinement via quality-aware face masking. To facilitate efficient quality evaluation, we introduce an objective topology-aware scoring system to evaluate geometric integrity and topological regularity at both object and face levels through two metrics: Boundary Edge Ratio (BER) and Topology Score (TS). By integrating these metrics into a fine-grained RL strategy, Mesh-RFT becomes the first method to optimize mesh quality at the granularity of individual faces, resolving localized errors while preserving global coherence. Experiment results show that our M-DPO approach reduces Hausdorff Distance (HD) by 24.6% and improves Topology Score (TS) by 3.8% over pre-trained models, while outperforming global DPO methods with a 17.4% HD reduction and 4.9% TS gain. These results demonstrate Mesh-RFT’s ability to improve geometric integrity and topological regularity, achieving new state-of-the-art performance in production-ready mesh generation. Project Page: https://hitcslj.github.io/mesh-rft/.

Method: Developed REOBench benchmark with high-resolution optical remote sensing images, evaluating models across 6 tasks and 12 corruption types (appearance-based and geometric). Systematically assessed models trained with masked image modeling, contrastive learning, and vision-language pre-training paradigms.

Result: Existing Earth observation foundation models experience significant performance degradation (1-20% drop) when exposed to input corruptions. Degradation severity varies by task, architecture, backbone size, and corruption type. Vision-language models show enhanced robustness, especially in multimodal tasks.

Conclusion: Current Earth observation foundation models are vulnerable to real-world corruptions. REOBench provides actionable insights for developing more robust and reliable models, with vision-language approaches showing promise for enhanced robustness.

Abstract: Earth observation foundation models have shown strong generalization across multiple Earth observation tasks, but their robustness under real-world perturbations remains underexplored. To bridge this gap, we introduce REOBench, the first comprehensive benchmark for evaluating the robustness of Earth observation foundation models across six tasks and twelve types of image corruptions, including both appearance-based and geometric perturbations. To ensure realistic and fine-grained evaluation, our benchmark focuses on high-resolution optical remote sensing images, which are widely used in critical applications such as urban planning and disaster response. We conduct a systematic evaluation of a broad range of models trained using masked image modeling, contrastive learning, and vision-language pre-training paradigms. Our results reveal that (1) existing Earth observation foundation models experience significant performance degradation when exposed to input corruptions. (2) The severity of degradation varies across tasks, model architectures, backbone sizes, and types of corruption, with performance drop varying from less than 1% to over 20%. (3) Vision-language models show enhanced robustness, particularly in multimodal tasks. REOBench underscores the vulnerability of current Earth observation foundation models to real-world corruptions and provides actionable insights for developing more robust and reliable models. Code and data are publicly available at https://github.com/lx709/REOBench.

Method: Proposes MODEM with Morton-Order 2D-Selective-Scan Module (MOS2D) that integrates Morton-coded spatial ordering with selective state-space models, and Dual Degradation Estimation Module (DDEM) that disentangles global and local degradation priors to dynamically condition the restoration process.

Result: Extensive experiments show MODEM achieves state-of-the-art results across multiple benchmarks and weather types, demonstrating effectiveness in modeling complex degradation dynamics.

Conclusion: MODEM effectively addresses adverse weather image restoration by accurately estimating degradation patterns and providing adaptive, context-aware restoration guidance through its novel architecture.

Abstract: Restoring images degraded by adverse weather remains a significant challenge due to the highly non-uniform and spatially heterogeneous nature of weather-induced artifacts, e.g., fine-grained rain streaks versus widespread haze. Accurately estimating the underlying degradation can intuitively provide restoration models with more targeted and effective guidance, enabling adaptive processing strategies. To this end, we propose a Morton-Order Degradation Estimation Mechanism (MODEM) for adverse weather image restoration. Central to MODEM is the Morton-Order 2D-Selective-Scan Module (MOS2D), which integrates Morton-coded spatial ordering with selective state-space models to capture long-range dependencies while preserving local structural coherence. Complementing MOS2D, we introduce a Dual Degradation Estimation Module (DDEM) that disentangles and estimates both global and local degradation priors. These priors dynamically condition the MOS2D modules, facilitating adaptive and context-aware restoration. Extensive experiments and ablation studies demonstrate that MODEM achieves state-of-the-art results across multiple benchmarks and weather types, highlighting its effectiveness in modeling complex degradation dynamics. Our code will be released at https://github.com/hainuo-wang/MODEM.git.

Method: Introduces Max-Former with two frequency-enhancing operators: (1) extra Max-Pool in patch embedding to restore high-frequency signals, and (2) Depth-Wise Convolution replacing self-attention. Also extends the insight to convolution-based networks with Max-ResNet-18.

Result: Max-Former achieves 82.39% top-1 accuracy on ImageNet with only 63.99M parameters, surpassing Spikformer by +7.58%. Max-ResNet-18 achieves state-of-the-art performance: 97.17% on CIFAR-10 and 83.06% on CIFAR-100.

Conclusion: The frequency bias in SNNs is the root cause of degraded performance, not binary activations. The proposed frequency-enhancing approach provides a simple yet effective solution that inspires future research on SNNs’ distinctive nature.

Abstract: Spiking Neural Networks promise brain-inspired and energy-efficient computation by transmitting information through binary (0/1) spikes. Yet, their performance still lags behind that of artificial neural networks, often assumed to result from information loss caused by sparse and binary activations. In this work, we challenge this long-standing assumption and reveal a previously overlooked frequency bias: spiking neurons inherently suppress high-frequency components and preferentially propagate low-frequency information. This frequency-domain imbalance, we argue, is the root cause of degraded feature representation in SNNs. Empirically, on Spiking Transformers, adopting Avg-Pooling (low-pass) for token mixing lowers performance to 76.73% on Cifar-100, whereas replacing it with Max-Pool (high-pass) pushes the top-1 accuracy to 79.12%. Accordingly, we introduce Max-Former that restores high-frequency signals through two frequency-enhancing operators: (1) extra Max-Pool in patch embedding, and (2) Depth-Wise Convolution in place of self-attention. Notably, Max-Former attains 82.39% top-1 accuracy on ImageNet using only 63.99M parameters, surpassing Spikformer (74.81%, 66.34M) by +7.58%. Extending our insight beyond transformers, our Max-ResNet-18 achieves state-of-the-art performance on convolution-based benchmarks: 97.17% on CIFAR-10 and 83.06% on CIFAR-100. We hope this simple yet effective solution inspires future research to explore the distinctive nature of spiking neural networks. Code is available: https://github.com/bic-L/MaxFormer.

Method: Parameterizes complex 3D deformation fields as a composition of rigid transforms, leveraging the biological constraint that individual bones do not bend in typical motion, without requiring expensive deformation regularizers or patient-specific hyperparameter optimization.

Result: PolyPose successfully aligns preoperative volumes to as few as two X-rays, providing crucial 3D guidance in challenging sparse-view and limited-angle settings where current registration methods fail, as demonstrated across diverse datasets from orthopedic surgery and radiotherapy.

Conclusion: The polyrigid formulation with anatomically plausible priors enables robust deformable 2D/3D registration in under-determined settings, making volumetric guidance feasible during interventional procedures with minimal 2D imaging requirements.

Abstract: Determining the 3D pose of a patient from a limited set of 2D X-ray images is a critical task in interventional settings. While preoperative volumetric imaging (e.g., CT and MRI) provides precise 3D localization and visualization of anatomical targets, these modalities cannot be acquired during procedures, where fast 2D imaging (X-ray) is used instead. To integrate volumetric guidance into intraoperative procedures, we present PolyPose, a simple and robust method for deformable 2D/3D registration. PolyPose parameterizes complex 3D deformation fields as a composition of rigid transforms, leveraging the biological constraint that individual bones do not bend in typical motion. Unlike existing methods that either assume no inter-joint movement or fail outright in this under-determined setting, our polyrigid formulation enforces anatomically plausible priors that respect the piecewise-rigid nature of human movement. This approach eliminates the need for expensive deformation regularizers that require patient- and procedure-specific hyperparameter optimization. Across extensive experiments on diverse datasets from orthopedic surgery and radiotherapy, we show that this strong inductive bias enables PolyPose to successfully align the patient’s preoperative volume to as few as two X-rays, thereby providing crucial 3D guidance in challenging sparse-view and limited-angle settings where current registration methods fail. Additional visualizations, tutorials, and code are available at https://polypose.csail.mit.edu.

Method: Created Roboflow100-VL benchmark with 100 multi-modal object detection datasets containing diverse concepts, and evaluated state-of-the-art models in zero-shot, few-shot, semi-supervised, and fully-supervised settings to compare performance across data regimes.

Result: VLMs like GroundingDINO and Qwen2.5-VL achieve less than 2% zero-shot accuracy on challenging medical imaging datasets in Roboflow100-VL, demonstrating poor generalization to out-of-distribution concepts. The winning team in their CVPR 2025 competition outperformed their baseline by 17 mAP.

Conclusion: There is a critical need for few-shot concept alignment in VLMs to handle out-of-distribution classes and imaging modalities, as current models show significant limitations in generalization despite their strong performance on common objects.

Abstract: Vision-language models (VLMs) trained on internet-scale data achieve remarkable zero-shot detection performance on common objects like car, truck, and pedestrian. However, state-of-the-art models still struggle to generalize to out-of-distribution classes, tasks and imaging modalities not typically found in their pre-training. Rather than simply re-training VLMs on more visual data, we argue that one should align VLMs to new concepts with annotation instructions containing a few visual examples and rich textual descriptions. To this end, we introduce Roboflow100-VL, a large-scale collection of 100 multi-modal object detection datasets with diverse concepts not commonly found in VLM pre-training. We evaluate state-of-the-art models on our benchmark in zero-shot, few-shot, semi-supervised, and fully-supervised settings, allowing for comparison across data regimes. Notably, we find that VLMs like GroundingDINO and Qwen2.5-VL achieve less than 2% zero-shot accuracy on challenging medical imaging datasets within Roboflow100-VL, demonstrating the need for few-shot concept alignment. Lastly, we discuss our recent CVPR 2025 Foundational FSOD competition and share insights from the community. Notably, the winning team significantly outperforms our baseline by 17 mAP! Our code and dataset are available at https://github.com/roboflow/rf100-vl and https://universe.roboflow.com/rf100-vl/.

Method: Uses a small calibration set to divide pruning into multiple stages. Early stages focus on global impact on subsequent layers, while deeper stages prioritize local output consistency.

Result: Achieves 78% compression rate while preserving 96.7% of original model performance on average across various LVLMs and benchmarks.

Conclusion: BTP effectively balances local and global pruning impacts, demonstrating broad effectiveness for reducing computational overhead in vision-language models.

Abstract: Large Vision-Language Models (LVLMs) have shown impressive performance across multi-modal tasks by encoding images into thousands of tokens. However, the large number of image tokens results in significant computational overhead, and the use of dynamic high-resolution inputs further increases this burden. Previous approaches have attempted to reduce the number of image tokens through token pruning, typically by selecting tokens based on attention scores or image token diversity. Through empirical studies, we observe that existing methods often overlook the joint impact of pruning on both the current layer’s output (local) and the outputs of subsequent layers (global), leading to suboptimal pruning decisions. To address this challenge, we propose Balanced Token Pruning (BTP), a plug-and-play method for pruning vision tokens. Specifically, our method utilizes a small calibration set to divide the pruning process into multiple stages. In the early stages, our method emphasizes the impact of pruning on subsequent layers, whereas in the deeper stages, the focus shifts toward preserving the consistency of local outputs. Extensive experiments across various LVLMs demonstrate the broad effectiveness of our approach on multiple benchmarks. Our method achieves a 78% compression rate while preserving 96.7% of the original models’ performance on average. Our code is available at https://github.com/EmbodiedCity/NeurIPS2025-Balanced-Token-Pruning.

Method: Sherlock framework with trajectory-level self-correction objective, visual perturbation-based preference data construction, and dynamic β for preference tuning. Uses only 20k annotated data for initial training.

Result: Achieves 64.1% accuracy with direct generation and 65.4% after self-correction across 8 benchmarks, outperforming LLaVA-CoT (63.2%), Mulberry (63.9%), and LlamaV-o1 (63.4%) while using <20% of annotated data.

Conclusion: Sherlock enables reasoning VLMs to acquire self-correction capabilities with minimal supervision and continue self-improving without external guidance, demonstrating effective self-correction and data efficiency.

Abstract: Reasoning Vision-Language Models (VLMs) have shown promising performance on complex multimodal tasks. However, they still face significant challenges: they are highly sensitive to reasoning errors, require large volumes of annotated data or accurate verifiers, and struggle to generalize beyond specific domains. To address these limitations, we explore self-correction as a strategy to enhance reasoning VLMs. We first conduct an in-depth analysis of reasoning VLMs’ self-correction abilities and identify key gaps. Based on our findings, we introduce Sherlock, a self-correction and self-improvement training framework. Sherlock introduces a trajectory-level self-correction objective, a preference data construction method based on visual perturbation, and a dynamic $\beta$ for preference tuning. Once the model acquires self-correction capabilities using only 20k randomly sampled annotated data, it continues to self-improve without external supervision. Built on the Llama3.2-Vision-11B model, Sherlock achieves remarkable results across eight benchmarks, reaching an average accuracy of 64.1 with direct generation and 65.4 after self-correction. It outperforms LLaVA-CoT (63.2), Mulberry (63.9), and LlamaV-o1 (63.4) while using less than 20% of the annotated data.

Method: SeG-SR framework with three modules: Semantic Feature Extraction Module (SFEM) using pretrained VLM to extract semantic knowledge, Semantic Localization Module (SLM) deriving semantic guidance, and Learnable Modulation Module (LMM) using semantic guidance to modulate SR network features.

Result: Achieved state-of-the-art performance on three datasets, with PSNR of 29.3042 dB and SSIM of 0.7961 for x4 SR on UCMerced dataset. Consistently improved performance across various SR architectures.

Conclusion: Semantic guidance from vision-language models effectively incorporates high-level scene understanding into remote sensing image super-resolution, leading to improved performance and generalizability across different architectures.

Abstract: High-resolution (HR) remote sensing imagery plays a vital role in a wide range of applications, including urban planning and environmental monitoring. However, due to limitations in sensors and data transmission links, the images acquired in practice often suffer from resolution degradation. Remote Sensing Image Super-Resolution (RSISR) aims to reconstruct HR images from low-resolution (LR) inputs, providing a cost-effective and efficient alternative to direct HR image acquisition. Existing RSISR methods primarily focus on low-level characteristics in pixel space, while neglecting the high-level understanding of remote sensing scenes. This may lead to semantically inconsistent artifacts in the reconstructed results. Motivated by this observation, our work aims to explore the role of high-level semantic knowledge in improving RSISR performance. We propose a Semantic-Guided Super-Resolution framework, SeG-SR, which leverages Vision-Language Models (VLMs) to extract semantic knowledge from input images and uses it to guide the super resolution (SR) process. Specifically, we first design a Semantic Feature Extraction Module (SFEM) that utilizes a pretrained VLM to extract semantic knowledge from remote sensing images. Next, we propose a Semantic Localization Module (SLM), which derives a series of semantic guidance from the extracted semantic knowledge. Finally, we develop a Learnable Modulation Module (LMM) that uses semantic guidance to modulate the features extracted by the SR network, effectively incorporating high-level scene understanding into the SR pipeline. We validate the effectiveness and generalizability of SeG-SR through extensive experiments: SeG-SR achieves state-of-the-art performance on three datasets, and consistently improves performance across various SR architectures. Notably, for the x4 SR task on UCMerced dataset, it attained a PSNR of 29.3042 dB and an SSIM of 0.7961.

Method: Proposed Predictive Future Modeling (PreFM) framework featuring: (a) predictive multimodal future modeling to integrate future audio-visual cues, (b) modality-agnostic robust representation, and (c) focal temporal prioritization for improved precision and generalization.

Result: PreFM significantly outperforms state-of-the-art methods on UnAV-100 and LLP datasets by large margins with significantly fewer parameters.

Conclusion: PreFM offers an effective approach for real-time multimodal video understanding, balancing high performance with computational efficiency for streaming video applications.

Abstract: Audio-visual event parsing plays a crucial role in understanding multimodal video content, but existing methods typically rely on offline processing of entire videos with huge model sizes, limiting their real-time applicability. We introduce Online Audio-Visual Event Parsing (On-AVEP), a novel paradigm for parsing audio, visual, and audio-visual events by sequentially analyzing incoming video streams. The On-AVEP task necessitates models with two key capabilities: (1) Accurate online inference, to effectively distinguish events with unclear and limited context in online settings, and (2) Real-time efficiency, to balance high performance with computational constraints. To cultivate these, we propose the Predictive Future Modeling (PreFM) framework featured by (a) predictive multimodal future modeling to infer and integrate beneficial future audio-visual cues, thereby enhancing contextual understanding and (b) modality-agnostic robust representation along with focal temporal prioritization to improve precision and generalization. Extensive experiments on the UnAV-100 and LLP datasets show PreFM significantly outperforms state-of-the-art methods by a large margin with significantly fewer parameters, offering an insightful approach for real-time multimodal video understanding. Code is available at https://github.com/XiaoYu-1123/PreFM.

Method: Created TreeOfLife-200M (214M images), trained BioCLIP 2 using contrastive vision-language training with hierarchical supervision to distinguish species, then analyzed embedding spaces at inter-species and intra-species levels.

Result: BioCLIP 2 achieved extraordinary accuracy on biological tasks like habitat classification and trait prediction. Embeddings aligned with functional/ecological meanings (beak sizes, habitats) and preserved intra-species variations in orthogonal subspaces. Properties improved with larger datasets.

Conclusion: Large-scale contrastive training with hierarchical supervision enables emergent biological understanding in vision models, creating meaningful embedding spaces that capture both inter-species distinctions and intra-species variations, with scaling effects enhancing these properties.

Abstract: Foundation models trained at scale exhibit remarkable emergent behaviors, learning new capabilities beyond their initial training objectives. We find such emergent behaviors in biological vision models via large-scale contrastive vision-language training. To achieve this, we first curate TreeOfLife-200M, comprising 214 million images of living organisms, the largest and most diverse biological organism image dataset to date. We then train BioCLIP 2 on TreeOfLife-200M to distinguish different species. Despite the narrow training objective, BioCLIP 2 yields extraordinary accuracy when applied to various biological visual tasks such as habitat classification and trait prediction. We identify emergent properties in the learned embedding space of BioCLIP 2. At the inter-species level, the embedding distribution of different species aligns closely with functional and ecological meanings (e.g., beak sizes and habitats). At the intra-species level, instead of being diminished, the intra-species variations (e.g., life stages and sexes) are preserved and better separated in subspaces orthogonal to inter-species distinctions. We provide formal proof and analyses to explain why hierarchical supervision and contrastive objectives encourage these emergent properties. Crucially, our results reveal that these properties become increasingly significant with larger-scale training data, leading to a biologically meaningful embedding space.

Method: Proposes ABMILX with global correlation-based attention refinement and multi-head mechanisms to address optimization challenges in sparse-attention MIL. Uses efficient multi-scale random patch sampling strategy for end-to-end training.

Result: An end-to-end trained ResNet with ABMILX surpasses state-of-the-art foundation models under the two-stage paradigm across multiple challenging benchmarks, while remaining computationally efficient (<10 RTX3090 hours).

Conclusion: The work demonstrates the potential of end-to-end learning in computational pathology and calls for greater research focus in this area, showing that properly addressed optimization challenges can lead to superior performance.

Abstract: Pre-trained encoders for offline feature extraction followed by multiple instance learning (MIL) aggregators have become the dominant paradigm in computational pathology (CPath), benefiting cancer diagnosis and prognosis. However, performance limitations arise from the absence of encoder fine-tuning for downstream tasks and disjoint optimization with MIL. While slide-level supervised end-to-end (E2E) learning is an intuitive solution to this issue, it faces challenges such as high computational demands and suboptimal results. These limitations motivate us to revisit E2E learning. We argue that prior work neglects inherent E2E optimization challenges, leading to performance disparities compared to traditional two-stage methods. In this paper, we pioneer the elucidation of optimization challenge caused by sparse-attention MIL and propose a novel MIL called ABMILX. It mitigates this problem through global correlation-based attention refinement and multi-head mechanisms. With the efficient multi-scale random patch sampling strategy, an E2E trained ResNet with ABMILX surpasses SOTA foundation models under the two-stage paradigm across multiple challenging benchmarks, while remaining computationally efficient (<10 RTX3090 hours). We show the potential of E2E learning in CPath and calls for greater research focus in this area. The code is https://github.com/DearCaat/E2E-WSI-ABMILX.

Method: Three-stage generation: BEV layout production, 3D lifting, and object placement refinement. Uses Chain-of-Thought Activation based on 3D-Front dataset and CoT-Grounded Generative Layout Reward for spatial reasoning. Addresses asset-layout mismatches via Iterative Asset-Layout Alignment.

Result: Extensive experiments show DirectLayout achieves impressive semantic consistency, generalization and physical plausibility in 3D scene layout generation.

Conclusion: DirectLayout successfully enables open-vocabulary 3D indoor scene synthesis with fine-grained user control through LLM-based spatial reasoning.

Abstract: Realistic 3D indoor scene synthesis is vital for embodied AI and digital content creation. It can be naturally divided into two subtasks: object generation and layout generation. While recent generative models have significantly advanced object-level quality and controllability, layout generation remains challenging due to limited datasets. Existing methods either overfit to these datasets or rely on predefined constraints to optimize numerical layout that sacrifice flexibility. As a result, they fail to generate scenes that are both open-vocabulary and aligned with fine-grained user instructions. We introduce DirectLayout, a framework that directly generates numerical 3D layouts from text descriptions using generalizable spatial reasoning of large language models (LLMs). DirectLayout decomposes the generation into three stages: producing a Bird’s-Eye View (BEV) layout, lifting it into 3D space, and refining object placements. To enable explicit spatial reasoning and help the model grasp basic principles of object placement, we employ Chain-of-Thought (CoT) Activation based on the 3D-Front dataset. Additionally, we design CoT-Grounded Generative Layout Reward to enhance generalization and spatial planning. During inference, DirectLayout addresses asset-layout mismatches via Iterative Asset-Layout Alignment through in-context learning. Extensive experiments demonstrate that DirectLayout achieves impressive semantic consistency, generalization and physical plausibility.

Method: Reimagines UNet decoding as solving an initial value problem (IVP), treating skip connections as discrete nodes. Uses adaptive ordinary differential equation method based on linear multistep method principles for multi-scale feature fusion. Architecture-independent approach adaptable to various U-Net-like networks.

Result: Experiments on ACDC, KiTS2023, MSD brain tumor, and ISIC2017/2018 skin lesion segmentation datasets show improved feature utilization, reduced network parameters, and maintained high performance.

Conclusion: The proposed method effectively addresses UNet’s skip connection limitations through ODE-based multi-scale fusion, providing a flexible solution that enhances feature interaction while reducing computational complexity.

Abstract: Medical image segmentation is a critical task in computer vision, with UNet serving as a milestone architecture. The typical component of UNet family is the skip connection, however, their skip connections face two significant limitations: (1) they lack effective interaction between features at different scales, and (2) they rely on simple concatenation or addition operations, which constrain efficient information integration. While recent improvements to UNet have focused on enhancing encoder and decoder capabilities, these limitations remain overlooked. To overcome these challenges, we propose a novel multi-scale feature fusion method that reimagines the UNet decoding process as solving an initial value problem (IVP), treating skip connections as discrete nodes. By leveraging principles from the linear multistep method, we propose an adaptive ordinary differential equation method to enable effective multi-scale feature fusion. Our approach is independent of the encoder and decoder architectures, making it adaptable to various U-Net-like networks. Experiments on ACDC, KiTS2023, MSD brain tumor, and ISIC2017/2018 skin lesion segmentation datasets demonstrate improved feature utilization, reduced network parameters, and maintained high performance. The code is available at https://github.com/nayutayuki/FuseUNet.

Method: The approach performs 2D instance segmentation on multi-view images to generate instance masks, matches instance IDs across views using a specialized module, develops instance NeRF to render implicit scenes with color, density, semantic and instance information, and converts this into high-precision point clouds based on volume density.

Result: PlantSegNeRF achieved significant improvements: 16.1% precision, 18.3% recall, 17.8% F1-score, and 24.2% IoU in semantic segmentation, and 11.7% mPrec, 38.2% mRec, 32.2% mCov, and 25.3% mWCov in instance segmentation across all plant species.

Conclusion: The study extends organ-level plant phenotyping capabilities and provides a high-throughput method to supply high-quality 3D data for developing large-scale models in plant science.

Abstract: Organ segmentation of plant point clouds is a prerequisite for the high-resolution and accurate extraction of organ-level phenotypic traits. Although the fast development of deep learning has boosted much research on segmentation of plant point clouds, the existing techniques for organ segmentation still face limitations in resolution, segmentation accuracy, and generalizability across various plant species. In this study, we proposed a novel approach called plant segmentation neural radiance fields (PlantSegNeRF), aiming to directly generate high-precision instance point clouds from multi-view RGB image sequences for a wide range of plant species. PlantSegNeRF performed 2D instance segmentation on the multi-view images to generate instance masks for each organ with a corresponding ID. The multi-view instance IDs corresponding to the same plant organ were then matched and refined using a specially designed instance matching module. The instance NeRF was developed to render an implicit scene, containing color, density, semantic and instance information. The implicit scene was ultimately converted into high-precision plant instance point clouds based on the volume density. The results proved that in semantic segmentation of point clouds, PlantSegNeRF outperformed the commonly used methods, demonstrating an average improvement of 16.1%, 18.3%, 17.8%, and 24.2% in precision, recall, F1-score, and IoU compared to the second-best results on structurally complex species. More importantly, PlantSegNeRF exhibited significant advantages in plant point cloud instance segmentation tasks. Across all plant species, it achieved average improvements of 11.7%, 38.2%, 32.2% and 25.3% in mPrec, mRec, mCov, mWCov, respectively. This study extends the organ-level plant phenotyping and provides a high-throughput way to supply high-quality 3D data for the development of large-scale models in plant science.

Method: Integrates multi-modal embeddings from a lightweight VLM with SAM2, using frozen pre-trained components, positional tie-breaker embeddings, and cross-attention layers for instance awareness.

Result: Achieves state-of-the-art performance in open-vocabulary semantic, instance, and panoptic segmentation across multiple benchmarks with strong zero-shot generalization.

Conclusion: OpenWorldSAM provides an efficient and flexible framework for open-vocabulary segmentation that generalizes well to unseen categories without additional training.

Abstract: The ability to segment objects based on open-ended language prompts remains a critical challenge, requiring models to ground textual semantics into precise spatial masks while handling diverse and unseen categories. We present OpenWorldSAM, a framework that extends the prompt-driven Segment Anything Model v2 (SAM2) to open-vocabulary scenarios by integrating multi-modal embeddings extracted from a lightweight vision-language model (VLM). Our approach is guided by four key principles: i) Unified prompting: OpenWorldSAM supports a diverse range of prompts, including category-level and sentence-level language descriptions, providing a flexible interface for various segmentation tasks. ii) Efficiency: By freezing the pre-trained components of SAM2 and the VLM, we train only 4.5 million parameters on the COCO-stuff dataset, achieving remarkable resource efficiency. iii) Instance Awareness: We enhance the model’s spatial understanding through novel positional tie-breaker embeddings and cross-attention layers, enabling effective segmentation of multiple instances. iv) Generalization: OpenWorldSAM exhibits strong zero-shot capabilities, generalizing well on unseen categories and an open vocabulary of concepts without additional training. Extensive experiments demonstrate that OpenWorldSAM achieves state-of-the-art performance in open-vocabulary semantic, instance, and panoptic segmentation across multiple benchmarks. Code is available at https://github.com/GinnyXiao/OpenWorldSAM.

Method: Created SnapMoGen dataset with 20K motion clips and 122K detailed textual descriptions, and developed MoMask++ model that transforms motion into multi-scale token sequences and uses a single generative masked transformer to generate all tokens.

Result: MoMask++ achieves state-of-the-art performance on both HumanML3D and SnapMoGen benchmarks, and demonstrates ability to process casual user prompts using LLM reformatting.

Conclusion: The combination of high-quality dataset (SnapMoGen) and improved model (MoMask++) enables fine-grained text-to-motion generation with better controllability and generalization capabilities.

Abstract: Text-to-motion generation has experienced remarkable progress in recent years. However, current approaches remain limited to synthesizing motion from short or general text prompts, primarily due to dataset constraints. This limitation undermines fine-grained controllability and generalization to unseen prompts. In this paper, we introduce SnapMoGen, a new text-motion dataset featuring high-quality motion capture data paired with accurate, expressive textual annotations. The dataset comprises 20K motion clips totaling 44 hours, accompanied by 122K detailed textual descriptions averaging 48 words per description (vs. 12 words of HumanML3D). Importantly, these motion clips preserve original temporal continuity as they were in long sequences, facilitating research in long-term motion generation and blending. We also improve upon previous generative masked modeling approaches. Our model, MoMask++, transforms motion into multi-scale token sequences that better exploit the token capacity, and learns to generate all tokens using a single generative masked transformer. MoMask++ achieves state-of-the-art performance on both HumanML3D and SnapMoGen benchmarks. Additionally, we demonstrate the ability to process casual user prompts by employing an LLM to reformat inputs to align with the expressivity and narration style of SnapMoGen. Project webpage: https://snap-research.github.io/SnapMoGen/

Method: FDAM uses two techniques: Attention Inversion (AttInv) to generate complementary high-pass filtering by inverting attention matrices, and Frequency Dynamic Scaling (FreqScale) to weight different frequency components for fine-grained adjustments.

Result: FDAM avoids representation collapse and achieves consistent performance improvements across various models (SegFormer, DeiT, MaskDINO) in semantic segmentation, object detection, and instance segmentation, with state-of-the-art results in remote sensing detection.

Conclusion: FDAM effectively addresses frequency vanishing in Vision Transformers through circuit-theory-inspired frequency modulation, enhancing performance across multiple vision tasks.

Abstract: Vision Transformers (ViTs) have significantly advanced computer vision, demonstrating strong performance across various tasks. However, the attention mechanism in ViTs makes each layer function as a low-pass filter, and the stacked-layer architecture in existing transformers suffers from frequency vanishing. This leads to the loss of critical details and textures. We propose a novel, circuit-theory-inspired strategy called Frequency-Dynamic Attention Modulation (FDAM), which can be easily plugged into ViTs. FDAM directly modulates the overall frequency response of ViTs and consists of two techniques: Attention Inversion (AttInv) and Frequency Dynamic Scaling (FreqScale). Since circuit theory uses low-pass filters as fundamental elements, we introduce AttInv, a method that generates complementary high-pass filtering by inverting the low-pass filter in the attention matrix, and dynamically combining the two. We further design FreqScale to weight different frequency components for fine-grained adjustments to the target response function. Through feature similarity analysis and effective rank evaluation, we demonstrate that our approach avoids representation collapse, leading to consistent performance improvements across various models, including SegFormer, DeiT, and MaskDINO. These improvements are evident in tasks such as semantic segmentation, object detection, and instance segmentation. Additionally, we apply our method to remote sensing detection, achieving state-of-the-art results in single-scale settings. The code is available at https://github.com/Linwei-Chen/FDAM.

Method: Uses instruction-editing datasets as hard negative image-text pairs with symmetric contrastive loss, and incorporates long descriptive captions with rotary positional encodings.

Result: Achieves substantial gains on MMVP benchmark and fine-grained visual recognition tasks while maintaining robust zero-shot performance on broader tasks; reduces visual hallucinations in MLLMs.

Conclusion: Targeted instruction-based contrastive learning combined with comprehensive descriptive information significantly enhances VLMs’ fine-grained understanding capabilities.

Abstract: Despite the success of Vision-Language Models (VLMs) like CLIP in aligning vision and language, their proficiency in detailed, fine-grained visual comprehension remains a key challenge. We present CLIP-IN, a novel framework that bolsters CLIP’s fine-grained perception through two core innovations. Firstly, we leverage instruction-editing datasets, originally designed for image manipulation, as a unique source of hard negative image-text pairs. Coupled with a symmetric hard negative contrastive loss, this enables the model to effectively distinguish subtle visual-semantic differences. Secondly, CLIP-IN incorporates long descriptive captions, utilizing rotary positional encodings to capture rich semantic context often missed by standard CLIP. Our experiments demonstrate that CLIP-IN achieves substantial gains on the MMVP benchmark and various fine-grained visual recognition tasks, without compromising robust zero-shot performance on broader classification and retrieval tasks. Critically, integrating CLIP-IN’s visual representations into Multimodal Large Language Models significantly reduces visual hallucinations and enhances reasoning abilities. This work underscores the considerable potential of synergizing targeted, instruction-based contrastive learning with comprehensive descriptive information to elevate the fine-grained understanding of VLMs.

Method: Proposes a Text-Conditioned State Space Model (TCSSM) that dynamically predicts input-dependent parameters using both bi-temporal images and geo-disaster-related descriptions. This facilitates alignment between visual data and textual descriptions to extract domain-invariant features across domains.

Result: Extensive experiments demonstrate superior performance against state-of-the-art models. The proposed method consistently outperforms existing approaches in handling domain shifts in CDVQA tasks.

Conclusion: The TCSSM framework effectively addresses domain shift in CDVQA by unifying bi-temporal imagery and textual information, enabling better domain generalization. The BrightVQA dataset facilitates future research in this area, and the code will be publicly available.

Abstract: The Earth’s surface is constantly changing, and detecting these changes provides valuable insights that benefit various aspects of human society. While traditional change detection methods have been employed to detect changes from bi-temporal images, these approaches typically require expert knowledge for accurate interpretation. To enable broader and more flexible access to change information by non-expert users, the task of Change Detection Visual Question Answering (CDVQA) has been introduced. However, existing CDVQA methods have been developed under the assumption that training and testing datasets share similar distributions. This assumption does not hold in real-world applications, where domain shifts often occur. In this paper, the CDVQA task is revisited with a focus on addressing domain shift. To this end, a new multi-modal and multi-domain dataset, BrightVQA, is introduced to facilitate domain generalization research in CDVQA. Furthermore, a novel state space model, termed Text-Conditioned State Space Model (TCSSM), is proposed. The TCSSM framework is designed to leverage both bi-temporal imagery and geo-disaster-related textual information in an unified manner to extract domain-invariant features across domains. Input-dependent parameters existing in TCSSM are dynamically predicted by using both bi-temporal images and geo-disaster-related description, thereby facilitating the alignment between bi-temporal visual data and the associated textual descriptions. Extensive experiments are conducted to evaluate the proposed method against state-of-the-art models, and superior performance is consistently demonstrated. The code and dataset will be made publicly available upon acceptance at https://github.com/Elman295/TCSSM.

Method: ViSpec uses a lightweight vision adaptor to compress image tokens into compact representations integrated into draft model’s attention while preserving positional information. It also extracts global image features to augment text tokens for multimodal coherence. A specialized training dataset is curated by repurposing existing datasets and generating extended outputs.

Result: ViSpec achieves, to our knowledge, the first substantial speedup in VLM speculative decoding, significantly outperforming existing methods that only achieve <1.5x speedup.

Conclusion: The proposed ViSpec framework successfully addresses the challenge of accelerating VLMs through speculative decoding, demonstrating that large VLMs can effectively filter redundant image information layer by layer while maintaining textual comprehension.

Abstract: Speculative decoding is a widely adopted technique for accelerating inference in large language models (LLMs), yet its application to vision-language models (VLMs) remains underexplored, with existing methods achieving only modest speedups (<1.5x). This gap is increasingly significant as multimodal capabilities become central to large-scale models. We hypothesize that large VLMs can effectively filter redundant image information layer by layer without compromising textual comprehension, whereas smaller draft models struggle to do so. To address this, we introduce Vision-Aware Speculative Decoding (ViSpec), a novel framework tailored for VLMs. ViSpec employs a lightweight vision adaptor module to compress image tokens into a compact representation, which is seamlessly integrated into the draft model’s attention mechanism while preserving original image positional information. Additionally, we extract a global feature vector for each input image and augment all subsequent text tokens with this feature to enhance multimodal coherence. To overcome the scarcity of multimodal datasets with long assistant responses, we curate a specialized training dataset by repurposing existing datasets and generating extended outputs using the target VLM with modified prompts. Our training strategy mitigates the risk of the draft model exploiting direct access to the target model’s hidden states, which could otherwise lead to shortcut learning when training solely on target model outputs. Extensive experiments validate ViSpec, achieving, to our knowledge, the first substantial speedup in VLM speculative decoding. Code is available at https://github.com/KangJialiang/ViSpec.

Method: Uses local pixel dependencies based on Markov Random Fields to reconstruct images and expose texture/edge inconsistencies, then applies FerretNet - a 1.1M parameter lightweight neural network trained on ProGAN dataset.

Result: Achieves 97.1% average accuracy on open-world benchmark with 22 generative models, demonstrating strong generalization despite training only on ProGAN data.

Conclusion: FerretNet provides efficient and robust synthetic image detection by leveraging generation artifacts and local pixel dependencies, offering a lightweight solution with excellent cross-model generalization.

Abstract: The increasing realism of synthetic images generated by advanced models such as VAEs, GANs, and LDMs poses significant challenges for synthetic image detection. To address this issue, we explore two artifact types introduced during the generation process: (1) latent distribution deviations and (2) decoding-induced smoothing effects, which manifest as inconsistencies in local textures, edges, and color transitions. Leveraging local pixel dependencies (LPD) properties rooted in Markov Random Fields, we reconstruct synthetic images using neighboring pixel information to expose disruptions in texture continuity and edge coherence. Building upon LPD, we propose FerretNet, a lightweight neural network with only 1.1M parameters that delivers efficient and robust synthetic image detection. Extensive experiments demonstrate that FerretNet, trained exclusively on the 4-class ProGAN dataset, achieves an average accuracy of 97.1% on an open-world benchmark comprising 22 generative models. Our code and datasets are publicly available at https://github.com/xigua7105/FerretNet.

Method: Jailbreaking with Loss-guided Image Perturbation (JaiLIP) minimizes a joint objective combining MSE loss between clean/adversarial images and the model’s harmful-output loss.

Result: JaiLIP generates highly effective and imperceptible adversarial images that outperform existing methods in producing toxicity, and works effectively in practical domains like transportation.

Conclusion: Image-based jailbreak attacks present practical challenges for VLMs, highlighting the need for efficient defense mechanisms.

Abstract: Vision-Language Models (VLMs) have remarkable abilities in generating multimodal reasoning tasks. However, potential misuse or safety alignment concerns of VLMs have increased significantly due to different categories of attack vectors. Among various attack vectors, recent studies have demonstrated that image-based perturbations are particularly effective in generating harmful outputs. In the literature, many existing techniques have been proposed to jailbreak VLMs, leading to unstable performance and visible perturbations. In this study, we propose Jailbreaking with Loss-guided Image Perturbation (JaiLIP), a jailbreaking attack in the image space that minimizes a joint objective combining the mean squared error (MSE) loss between clean and adversarial image with the models harmful-output loss. We evaluate our proposed method on VLMs using standard toxicity metrics from Perspective API and Detoxify. Experimental results demonstrate that our method generates highly effective and imperceptible adversarial images, outperforming existing methods in producing toxicity. Moreover, we have evaluated our method in the transportation domain to demonstrate the attacks practicality beyond toxic text generation in specific domain. Our findings emphasize the practical challenges of image-based jailbreak attacks and the need for efficient defense mechanisms for VLMs.

Method: Uses lightweight dual-branch conditioner to inject signals from degraded input and restored proxy, timestep- and layer-adaptive modulation schedule, caption-free semantic alignment via SigLIP features, and scalable curation pipeline for structure-rich supervision.

Result: Consistently outperforms strong open-source and commercial baselines across synthetic and in-the-wild benchmarks, with ablation studies verifying component necessity.

Conclusion: For large DiTs, when, where, and what to condition on — rather than adding parameters or relying on text prompts — is the governing lever for robust and caption-free universal image restoration in the wild.

Abstract: Universal image restoration (UIR) aims to recover images degraded by unknown mixtures while preserving semantics – conditions under which discriminative restorers and UNet-based diffusion priors often oversmooth, hallucinate, or drift. We present LucidFlux, a caption-free UIR framework that adapts a large diffusion transformer (Flux.1) without image captions. LucidFlux introduces a lightweight dual-branch conditioner that injects signals from the degraded input and a lightly restored proxy to respectively anchor geometry and suppress artifacts. Then, a timestep- and layer-adaptive modulation schedule is designed to route these cues across the backbone’s hierarchy, in order to yield coarse-to-fine and context-aware updates that protect the global structure while recovering texture. After that, to avoid the latency and instability of text prompts or MLLM captions, we enforce caption-free semantic alignment via SigLIP features extracted from the proxy. A scalable curation pipeline further filters large-scale data for structure-rich supervision. Across synthetic and in-the-wild benchmarks, LucidFlux consistently outperforms strong open-source and commercial baselines, and ablation studies verify the necessity of each component. LucidFlux shows that, for large DiTs, when, where, and what to condition on – rather than adding parameters or relying on text prompts – is the governing lever for robust and caption-free universal image restoration in the wild.

Method: Created a novel Vietnamese multimodal medical dataset with PET/CT volumes and clinical reports, plus a training framework with data augmentation and expert-validated test sets.

Result: Incorporating the dataset significantly improves existing VLMs’ performance on downstream tasks, demonstrating enhanced capabilities for medical imaging in low-resource languages.

Conclusion: This dataset and benchmark advance robust VLMs for medical imaging, particularly benefiting low-resource languages and Vietnamese clinical applications.

Abstract: Vision-Language Foundation Models (VLMs), trained on large-scale multimodal datasets, have driven significant advances in Artificial Intelligence (AI) by enabling rich cross-modal reasoning. Despite their success in general domains, applying these models to medical imaging remains challenging due to the limited availability of diverse imaging modalities and multilingual clinical data. Most existing medical VLMs are trained on a subset of imaging modalities and focus primarily on high-resource languages, thus limiting their generalizability and clinical utility. To address these limitations, we introduce a novel Vietnamese-language multimodal medical dataset consisting of 2,757 whole-body PET/CT volumes from independent patients and their corresponding full-length clinical reports. This dataset is designed to fill two pressing gaps in medical AI development: (1) the lack of PET/CT imaging data in existing VLMs training corpora, which hinders the development of models capable of handling functional imaging tasks; and (2) the underrepresentation of low-resource languages, particularly the Vietnamese language, in medical vision-language research. To the best of our knowledge, this is the first dataset to provide comprehensive PET/CT-report pairs in Vietnamese. We further introduce a training framework to enhance VLMs’ learning, including data augmentation and expert-validated test sets. We conduct comprehensive experiments benchmarking state-of-the-art VLMs on downstream tasks. The experimental results show that incorporating our dataset significantly improves the performance of existing VLMs. We believe this dataset and benchmark will serve as a pivotal step in advancing the development of more robust VLMs for medical imaging, especially for low-resource languages and clinical use in Vietnamese healthcare. The source code is available at https://github.com/AIoT-Lab-BKAI/ViPET-ReportGen.

Method: Proposes Cross-modal Iterative Prompting (CIP) that conditions LLMs on class names and support images to generate precise descriptions, and Cross-modal Geometric Alignment (CGA) that aligns textual, support, and synthetic visual representations by minimizing kernelized volume of 3D parallelotope.

Result: Establishes new state-of-the-art performance across ten diverse benchmarks including standard, cross-domain, and fine-grained few-shot learning scenarios.

Conclusion: VT-FSL effectively bridges vision and text with LLMs, providing precise cross-modal prompts and geometry-aware alignment for improved few-shot learning performance.

Abstract: Few-shot learning (FSL) aims to recognize novel concepts from only a few labeled support samples. Recent studies enhance support features by incorporating additional semantic information or designing complex semantic fusion modules. However, they still suffer from hallucinating semantics that contradict the visual evidence due to the lack of grounding in actual instances, resulting in noisy guidance and costly corrections. To address these issues, we propose a novel framework, bridging Vision and Text with LLMs for Few-Shot Learning (VT-FSL), which constructs precise cross-modal prompts conditioned on Large Language Models (LLMs) and support images, seamlessly integrating them through a geometry-aware alignment. It mainly consists of Cross-modal Iterative Prompting (CIP) and Cross-modal Geometric Alignment (CGA). Specifically, the CIP conditions an LLM on both class names and support images to generate precise class descriptions iteratively in a single structured reasoning pass. These descriptions not only enrich the semantic understanding of novel classes but also enable the zero-shot synthesis of semantically consistent images. The descriptions and synthetic images act respectively as complementary textual and visual prompts, providing high-level class semantics and low-level intra-class diversity to compensate for limited support data. Furthermore, the CGA jointly aligns the fused textual, support, and synthetic visual representations by minimizing the kernelized volume of the 3-dimensional parallelotope they span. It captures global and nonlinear relationships among all representations, enabling structured and consistent multimodal integration. The proposed VT-FSL method establishes new state-of-the-art performance across ten diverse benchmarks, including standard, cross-domain, and fine-grained few-shot learning scenarios. Code is available at https://github.com/peacelwh/VT-FSL.

Method: Generate 3D pseudo-ground-truth labels using foundation models (Grounded-SAM and Metric3Dv2) and leverage temporal information for label densification. Also propose EasyOcc model that learns solely from these labels without complex rendering.

Result: 45% mIoU improvement (9.73 to 14.09) when integrated into OccNeRF. EasyOcc achieves 13.86 mIoU and 7.71 mIoU on full scene without camera mask, outperforming previous best by 31%.

Conclusion: Foundation models, temporal context, and loss computation space choice are critical for self-supervised learning in comprehensive scene understanding.

Abstract: Self-supervised models have recently achieved notable advancements, particularly in the domain of semantic occupancy prediction. These models utilize sophisticated loss computation strategies to compensate for the absence of ground-truth labels. For instance, techniques such as novel view synthesis, cross-view rendering, and depth estimation have been explored to address the issue of semantic and depth ambiguity. However, such techniques typically incur high computational costs and memory usage during the training stage, especially in the case of novel view synthesis. To mitigate these issues, we propose 3D pseudo-ground-truth labels generated by the foundation models Grounded-SAM and Metric3Dv2, and harness temporal information for label densification. Our 3D pseudo-labels can be easily integrated into existing models, which yields substantial performance improvements, with mIoU increasing by 45%, from 9.73 to 14.09, when implemented into the OccNeRF model. This stands in contrast to earlier advancements in the field, which are often not readily transferable to other architectures. Additionally, we propose a streamlined model, EasyOcc, achieving 13.86 mIoU. This model conducts learning solely from our labels, avoiding complex rendering strategies mentioned previously. Furthermore, our method enables models to attain state-of-the-art performance when evaluated on the full scene without applying the camera mask, with EasyOcc achieving 7.71 mIoU, outperforming the previous best model by 31%. These findings highlight the critical importance of foundation models, temporal context, and the choice of loss computation space in self-supervised learning for comprehensive scene understanding.

Method: DragFlow introduces region-based editing with affine transformations for richer feature supervision, integrates pretrained personalization adapters (IP-Adapter) for subject consistency, uses gradient mask-based hard constraints for background preservation, and employs MLLMs to resolve task ambiguities.

Result: Extensive experiments on DragBench-DR and the novel ReD Bench benchmark show that DragFlow surpasses both point-based and region-based baselines, achieving state-of-the-art performance in drag-based image editing.

Conclusion: DragFlow successfully harnesses FLUX’s strong generative priors for drag-based editing through region-based supervision and enhanced consistency mechanisms, setting a new standard for distortion-free drag-based image editing.

Abstract: Drag-based image editing has long suffered from distortions in the target region, largely because the priors of earlier base models, Stable Diffusion, are insufficient to project optimized latents back onto the natural image manifold. With the shift from UNet-based DDPMs to more scalable DiT with flow matching (e.g., SD3.5, FLUX), generative priors have become significantly stronger, enabling advances across diverse editing tasks. However, drag-based editing has yet to benefit from these stronger priors. This work proposes the first framework to effectively harness FLUX’s rich prior for drag-based editing, dubbed DragFlow, achieving substantial gains over baselines. We first show that directly applying point-based drag editing to DiTs performs poorly: unlike the highly compressed features of UNets, DiT features are insufficiently structured to provide reliable guidance for point-wise motion supervision. To overcome this limitation, DragFlow introduces a region-based editing paradigm, where affine transformations enable richer and more consistent feature supervision. Additionally, we integrate pretrained open-domain personalization adapters (e.g., IP-Adapter) to enhance subject consistency, while preserving background fidelity through gradient mask-based hard constraints. Multimodal large language models (MLLMs) are further employed to resolve task ambiguities. For evaluation, we curate a novel Region-based Dragging benchmark (ReD Bench) featuring region-level dragging instructions. Extensive experiments on DragBench-DR and ReD Bench show that DragFlow surpasses both point-based and region-based baselines, setting a new state-of-the-art in drag-based image editing. Code and datasets will be publicly available upon publication.

Method: Combines Gram matrix-based style extraction with metric learning for intra-class compactness and inter-class separation to extract style embeddings. Introduces SEDD metric to quantify domain gap.

Result: The method successfully quantifies synthetic-to-real gap across various datasets and sim-to-real methods, providing a standardized profiling-based quality control paradigm.

Conclusion: This work enables systematic diagnosis and targeted enhancement of synthetic datasets, advancing data-driven autonomous driving system development.

Abstract: Ensuring the reliability of autonomous driving perception systems requires extensive environment-based testing, yet real-world execution is often impractical. Synthetic datasets have therefore emerged as a promising alternative, offering advantages such as cost-effectiveness, bias free labeling, and controllable scenarios. However, the domain gap between synthetic and real-world datasets remains a major obstacle to model generalization. To address this challenge from a data-centric perspective, this paper introduces a profile extraction and discovery framework for characterizing the style profiles underlying both synthetic and real image datasets. We propose Style Embedding Distribution Discrepancy (SEDD) as a novel evaluation metric. Our framework combines Gram matrix-based style extraction with metric learning optimized for intra-class compactness and inter-class separation to extract style embeddings. Furthermore, we establish a benchmark using publicly available datasets. Experiments are conducted on a variety of datasets and sim-to-real methods, and the results show that our method is capable of quantifying the synthetic-to-real gap. This work provides a standardized profiling-based quality control paradigm that enables systematic diagnosis and targeted enhancement of synthetic datasets, advancing future development of data-driven autonomous driving systems.

Method: Introduced AURA benchmark dataset with multiple annotators, evaluated four VAD models across two marine scenes, and implemented robust frame selection strategies for meaningful video segment extraction.

Result: VAD performance varies dramatically across models, highly sensitive to training data amount and visual content variability defining “normal” scenes. Soft and consensus labels proved valuable.

Conclusion: The study offers practical approach for supporting scientific exploration and scalable biodiversity monitoring through automated underwater anomaly detection systems.

Abstract: Underwater video monitoring is a promising strategy for assessing marine biodiversity, but the vast volume of uneventful footage makes manual inspection highly impractical. In this work, we explore the use of visual anomaly detection (VAD) based on deep neural networks to automatically identify interesting or anomalous events. We introduce AURA, the first multi-annotator benchmark dataset for underwater VAD, and evaluate four VAD models across two marine scenes. We demonstrate the importance of robust frame selection strategies to extract meaningful video segments. Our comparison against multiple annotators reveals that VAD performance of current models varies dramatically and is highly sensitive to both the amount of training data and the variability in visual content that defines “normal” scenes. Our results highlight the value of soft and consensus labels and offer a practical approach for supporting scientific exploration and scalable biodiversity monitoring.

Method: A graph-based framework that represents 3D CT volumes as structured graphs, where axial slice triplets serve as nodes processed through spectral graph convolution to capture inter-slice dependencies.

Result: The method achieves strong cross-dataset generalization across 3 independent institution datasets and shows competitive performance compared to state-of-the-art visual encoders.

Conclusion: The proposed 2.5D graph-based approach effectively handles multi-label classification of 3D CT scans while maintaining clinical deployment compatibility, with broader applicability demonstrated in radiology report generation and abdominal CT data.

Abstract: With the growing volume of CT examinations, there is an increasing demand for automated tools such as organ segmentation, abnormality detection, and report generation to support radiologists in managing their clinical workload. Multi-label classification of 3D Chest CT scans remains a critical yet challenging problem due to the complex spatial relationships inherent in volumetric data and the wide variability of abnormalities. Existing methods based on 3D convolutional neural networks struggle to capture long-range dependencies, while Vision Transformers often require extensive pre-training on large-scale, domain-specific datasets to perform competitively. In this work of academic research, we propose a 2.5D alternative by introducing a new graph-based framework that represents 3D CT volumes as structured graphs, where axial slice triplets serve as nodes processed through spectral graph convolution, enabling the model to reason over inter-slice dependencies while maintaining complexity compatible with clinical deployment. Our method, trained and evaluated on 3 datasets from independent institutions, achieves strong cross-dataset generalization, and shows competitive performance compared to state-of-the-art visual encoders. We further conduct comprehensive ablation studies to evaluate the impact of various aggregation strategies, edge-weighting schemes, and graph connectivity patterns. Additionally, we demonstrate the broader applicability of our approach through transfer experiments on automated radiology report generation and abdominal CT data.

Method: Created mmWalk dataset with manually controlled walking trajectories, multi-view sensor data, and accessibility-oriented features. Generated mmWalkVQA benchmark with visual question-answer triplets across 9 categories. Evaluated VLMs and fine-tuned models.

Result: State-of-the-art VLMs struggle with risk assessment and navigational tasks. mmWalk-finetuned model shows effectiveness on real-world datasets for advancing multi-modal walking assistance.

Conclusion: mmWalk dataset successfully addresses BLV navigation challenges and enables development of more effective walking assistance systems through comprehensive multi-modal data and specialized benchmarks.

Abstract: Walking assistance in extreme or complex environments remains a significant challenge for people with blindness or low vision (BLV), largely due to the lack of a holistic scene understanding. Motivated by the real-world needs of the BLV community, we build mmWalk, a simulated multi-modal dataset that integrates multi-view sensor and accessibility-oriented features for outdoor safe navigation. Our dataset comprises 120 manually controlled, scenario-categorized walking trajectories with 62k synchronized frames. It contains over 559k panoramic images across RGB, depth, and semantic modalities. Furthermore, to emphasize real-world relevance, each trajectory involves outdoor corner cases and accessibility-specific landmarks for BLV users. Additionally, we generate mmWalkVQA, a VQA benchmark with over 69k visual question-answer triplets across 9 categories tailored for safe and informed walking assistance. We evaluate state-of-the-art Vision-Language Models (VLMs) using zero- and few-shot settings and found they struggle with our risk assessment and navigational tasks. We validate our mmWalk-finetuned model on real-world datasets and show the effectiveness of our dataset for advancing multi-modal walking assistance.

Method: Proposes Joint Learning of Causal Representation and Reasoning using Structural Causal Model (SCM). First analyzes hidden confounders in base class representations and causal relationships between base and novel classes. Uses causal representation prototypes to eliminate confounders and capture causal representations, then models causal relationships using graph structures for reasoning from base to novel classes.

Result: Extensive experiments and visualization results on 3D and 2D NCD semantic segmentation demonstrate the method’s superior performance.

Conclusion: The proposed causal representation and reasoning approach effectively addresses the 3D-NCD problem by uncovering essential point cloud representations through causal modeling, enabling accurate segmentation of novel classes using only base class supervision.

Abstract: In this paper, we focus on Novel Class Discovery for Point Cloud Segmentation (3D-NCD), aiming to learn a model that can segment unlabeled (novel) 3D classes using only the supervision from labeled (base) 3D classes. The key to this task is to setup the exact correlations between the point representations and their base class labels, as well as the representation correlations between the points from base and novel classes. A coarse or statistical correlation learning may lead to the confusion in novel class inference. lf we impose a causal relationship as a strong correlated constraint upon the learning process, the essential point cloud representations that accurately correspond to the classes should be uncovered. To this end, we introduce a structural causal model (SCM) to re-formalize the 3D-NCD problem and propose a new method, i.e., Joint Learning of Causal Representation and Reasoning. Specifically, we first analyze hidden confounders in the base class representations and the causal relationships between the base and novel classes through SCM. We devise a causal representation prototype that eliminates confounders to capture the causal representations of base classes. A graph structure is then used to model the causal relationships between the base classes’ causal representation prototypes and the novel class prototypes, enabling causal reasoning from base to novel classes. Extensive experiments and visualization results on 3D and 2D NCD semantic segmentation demonstrate the superiorities of our method.

Method: Developed a unified benchmark with cognitively grounded taxonomy categorizing tasks into four quadrants, and created an automated pipeline to generate diverse spatial reasoning questions, resulting in Spatial-DISE dataset with evaluation and training VQA pairs.

Result: Evaluation of 28 state-of-the-art VLMs shows large and consistent gaps to human competence, particularly on multi-step multi-view spatial reasoning tasks.

Conclusion: Spatial-DISE provides a robust framework, valuable dataset, and clear direction for advancing VLMs toward human-like spatial intelligence, with benchmark and code to be publicly released.

Abstract: Spatial reasoning ability is crucial for Vision Language Models (VLMs) to support real-world applications in diverse domains including robotics, augmented reality, and autonomous navigation. Unfortunately, existing benchmarks are inadequate in assessing spatial reasoning ability, especially the \emph{intrinsic-dynamic} spatial reasoning which is a fundamental aspect of human spatial cognition. In this paper, we propose a unified benchmark, \textbf{Spatial-DISE}, based on a cognitively grounded taxonomy that categorizes tasks into four fundamental quadrants: \textbf{I}ntrinsic-\textbf{S}tatic, Intrinsic-\textbf{D}ynamic, \textbf{E}xtrinsic-Static, and Extrinsic-Dynamic spatial reasoning. Moreover, to address the issue of data scarcity, we develop a scalable and automated pipeline to generate diverse and verifiable spatial reasoning questions, resulting in a new \textbf{Spatial-DISE} dataset that includes Spatial-DISE Bench (559 evaluation VQA pairs) and Spatial-DISE-12K (12K+ training VQA pairs). Our comprehensive evaluation across 28 state-of-the-art VLMs reveals that, current VLMs have a large and consistent gap to human competence, especially on multi-step multi-view spatial reasoning. Spatial-DISE offers a robust framework, valuable dataset, and clear direction for future research toward human-like spatial intelligence. Benchmark, dataset, and code will be publicly released.

Method: Proposes Identity-Preserving Reward-guided Optimization (IPRO) - a tuning algorithm that uses face identity scorer, backpropagates reward signals through sampling chain, employs facial scoring with ground-truth videos as feature pools, and incorporates KL-divergence regularization.

Result: Extensive experiments on Wan 2.2 I2V model and in-house I2V model demonstrate the method’s effectiveness in enhancing identity preservation.

Conclusion: IPRO provides a direct and effective approach to improve identity consistency in human-centric video generation without requiring architectural changes or auxiliary modules.

Abstract: Recent advances in image-to-video (I2V) generation have achieved remarkable progress in synthesizing high-quality, temporally coherent videos from static images. Among all the applications of I2V, human-centric video generation includes a large portion. However, existing I2V models encounter difficulties in maintaining identity consistency between the input human image and the generated video, especially when the person in the video exhibits significant expression changes and movements. This issue becomes critical when the human face occupies merely a small fraction of the image. Since humans are highly sensitive to identity variations, this poses a critical yet under-explored challenge in I2V generation. In this paper, we propose Identity-Preserving Reward-guided Optimization (IPRO), a novel video diffusion framework based on reinforcement learning to enhance identity preservation. Instead of introducing auxiliary modules or altering model architectures, our approach introduces a direct and effective tuning algorithm that optimizes diffusion models using a face identity scorer. To improve performance and accelerate convergence, our method backpropagates the reward signal through the last steps of the sampling chain, enabling richer gradient feedback. We also propose a novel facial scoring mechanism that treats faces in ground-truth videos as facial feature pools, providing multi-angle facial information to enhance generalization. A KL-divergence regularization is further incorporated to stabilize training and prevent overfitting to the reward signal. Extensive experiments on Wan 2.2 I2V model and our in-house I2V model demonstrate the effectiveness of our method. Our project and code are available at https://ipro-alimama.github.io/.

Method: Introduces VaCo with Modular Task Queries (MTQs) and Visual Alignment Layers (VALs) to activate specific visual signals under VFM supervision, using Token Gateway Mask (TGM) to coordinate representation conflicts across multiple VFMs.

Result: Extensive experiments show VaCo significantly improves performance of different MLLMs on various benchmarks, demonstrating superior visual comprehension capabilities.

Conclusion: VaCo effectively bridges the gap between text-only supervision and vision-centric requirements in MLLMs, enabling better integration of visual information for enhanced multimodal understanding.

Abstract: Multimodal large language models (MLLMs) integrate image features from visual encoders with LLMs, demonstrating advanced comprehension capabilities. However, mainstream MLLMs are solely supervised by the next-token prediction of textual tokens, neglecting critical vision-centric information essential for analytical abilities. To track this dilemma, we introduce VaCo, which optimizes MLLM representations through Vision-Centric activation and Coordination from multiple vision foundation models (VFMs). VaCo introduces visual discriminative alignment to integrate task-aware perceptual features extracted from VFMs, thereby unifying the optimization of both textual and visual outputs in MLLMs. Specifically, we incorporate the learnable Modular Task Queries (MTQs) and Visual Alignment Layers (VALs) into MLLMs, activating specific visual signals under the supervision of diverse VFMs. To coordinate representation conflicts across VFMs, the crafted Token Gateway Mask (TGM) restricts the information flow among multiple groups of MTQs. Extensive experiments demonstrate that VaCo significantly improves the performance of different MLLMs on various benchmarks, showcasing its superior capabilities in visual comprehension.

Method: Proposes a unified framework with two components: Geometric Prior Enhancement Module (GPEM) that leverages part-level and structural cues for object consistency under degraded conditions, and Semantic Alignment Injection Mechanism (SAIM) that enriches language embeddings with domain-specific priors.

Result: The framework consistently outperforms existing open-vocabulary baselines in both In-Domain and Cross-Domain settings on the MARIS benchmark.

Conclusion: Establishes a strong foundation for future underwater perception research by addressing key challenges in underwater open-vocabulary instance segmentation.

Abstract: Most existing underwater instance segmentation approaches are constrained by close-vocabulary prediction, limiting their ability to recognize novel marine categories. To support evaluation, we introduce \textbf{MARIS} (\underline{Mar}ine Open-Vocabulary \underline{I}nstance \underline{S}egmentation), the first large-scale fine-grained benchmark for underwater Open-Vocabulary (OV) segmentation, featuring a limited set of seen categories and diverse unseen categories. Although OV segmentation has shown promise on natural images, our analysis reveals that transfer to underwater scenes suffers from severe visual degradation (e.g., color attenuation) and semantic misalignment caused by lack underwater class definitions. To address these issues, we propose a unified framework with two complementary components. The Geometric Prior Enhancement Module (\textbf{GPEM}) leverages stable part-level and structural cues to maintain object consistency under degraded visual conditions. The Semantic Alignment Injection Mechanism (\textbf{SAIM}) enriches language embeddings with domain-specific priors, mitigating semantic ambiguity and improving recognition of unseen categories. Experiments show that our framework consistently outperforms existing OV baselines both In-Domain and Cross-Domain setting on MARIS, establishing a strong foundation for future underwater perception research.

Method: Encoder-decoder architecture with sparse convolution on ResNet-18 backbone, SPLite decoder for faster decoding, and quantization-aware training for memory optimization.

Result: 42% end-to-end efficiency improvement, 3.1x frame rate boost on Raspberry Pi 5, 2.98x overall speed-up, with minimal accuracy loss (PA-MPJPE from 9.0mm to 9.1mm on FreiHAND).

Conclusion: The proposed framework achieves comparable accuracy to state-of-the-art methods while significantly enhancing computational efficiency for edge deployment.

Abstract: With the increasing ubiquity of AR/VR devices, the deployment of deep learning models on edge devices has become a critical challenge. These devices require real-time inference, low power consumption, and minimal latency. Many framework designers face the conundrum of balancing efficiency and performance. We design a light framework that adopts an encoder-decoder architecture and introduces several key contributions aimed at improving both efficiency and accuracy. We apply sparse convolution on a ResNet-18 backbone to exploit the inherent sparsity in hand pose images, achieving a 42% end-to-end efficiency improvement. Moreover, we propose our SPLite decoder. This new architecture significantly boosts the decoding process’s frame rate by 3.1x on the Raspberry Pi 5, while maintaining accuracy on par. To further optimize performance, we apply quantization-aware training, reducing memory usage while preserving accuracy (PA-MPJPE increases only marginally from 9.0 mm to 9.1 mm on FreiHAND). Overall, our system achieves a 2.98x speed-up on a Raspberry Pi 5 CPU (BCM2712 quad-core Arm A76 processor). Our method is also evaluated on compound benchmark datasets, demonstrating comparable accuracy to state-of-the-art approaches while significantly enhancing computational efficiency.

Method: Uses consistency models to learn self-consistent mapping between noisy and clean motion states, with Transformer-based spatiotemporal architecture and temporal embeddings for long-range dependencies.

Result: Achieves comparable or superior accuracy to state-of-the-art diffusion models on Human3.6M and HumanEva-I datasets while reducing inference steps by up to two orders of magnitude.

Conclusion: HumanCM provides an efficient alternative to diffusion models for human motion prediction with similar performance but much faster inference.

Abstract: We present HumanCM, a one-step human motion prediction framework built upon consistency models. Instead of relying on multi-step denoising as in diffusion-based methods, HumanCM performs efficient single-step generation by learning a self-consistent mapping between noisy and clean motion states. The framework adopts a Transformer-based spatiotemporal architecture with temporal embeddings to model long-range dependencies and preserve motion coherence. Experiments on Human3.6M and HumanEva-I demonstrate that HumanCM achieves comparable or superior accuracy to state-of-the-art diffusion models while reducing inference steps by up to two orders of magnitude.

Method: Created an extended dataset with: camera modality containing full and mini versions with four occlusion types; radar and LiDAR with parameterised occlusion scripts implementing three degradation types each for flexible and repeatable data generation.

Result: Provides the first multi-sensor occlusion dataset with controlled and reproducible degradations, enabling consistent evaluation of perception models under partial sensor failures and environmental interference.

Conclusion: This resource aims to advance research on robust sensor fusion, resilience analysis, and safety-critical perception in automated driving by enabling systematic testing under adverse conditions.

Abstract: Robust perception in automated driving requires reliable performance under adverse conditions, where sensors may be affected by partial failures or environmental occlusions. Although existing autonomous driving datasets inherently contain sensor noise and environmental variability, very few enable controlled, parameterised, and reproducible degradations across multiple sensing modalities. This gap limits the ability to systematically evaluate how perception and fusion architectures perform under well-defined adverse conditions. To address this limitation, we introduce the Occluded nuScenes Dataset, a novel extension of the widely used nuScenes benchmark. For the camera modality, we release both the full and mini versions with four types of occlusions, two adapted from public implementations and two newly designed. For radar and LiDAR, we provide parameterised occlusion scripts that implement three types of degradations each, enabling flexible and repeatable generation of occluded data. This resource supports consistent, reproducible evaluation of perception models under partial sensor failures and environmental interference. By releasing the first multi-sensor occlusion dataset with controlled and reproducible degradations, we aim to advance research on robust sensor fusion, resilience analysis, and safety-critical perception in automated driving.

Method: Propose EMIM module that constructs affinity matrix in cost volume style by sampling key candidate tokens from query-based neighboring area in next frame using sliding-window. The affinity matrix is used for both contextual aggregation and motion feature generation.

Result: Validated on four datasets, performs better than state-of-the-art approaches, especially on motion-sensitive datasets Something-Something V1 & V2.

Conclusion: EMIM effectively integrates motion modeling properties into transformers, achieving superior performance on motion-sensitive action recognition tasks.

Abstract: Recently, action recognition has been dominated by transformer-based methods, thanks to their spatiotemporal contextual aggregation capacities. However, despite the significant progress achieved on scene-related datasets, they do not perform well on motion-sensitive datasets due to the lack of elaborate motion modeling designs. Meanwhile, we observe that the widely-used cost volume in traditional action recognition is highly similar to the affinity matrix defined in self-attention, but equipped with powerful motion modeling capacities. In light of this, we propose to integrate those effective motion modeling properties into the existing transformer in a unified and neat way, with the proposal of the Explicit Motion Information Mining module (EMIM). In EMIM, we propose to construct the desirable affinity matrix in a cost volume style, where the set of key candidate tokens is sampled from the query-based neighboring area in the next frame in a sliding-window manner. Then, the constructed affinity matrix is used to aggregate contextual information for appearance modeling and is converted into motion features for motion modeling as well. We validate the motion modeling capacities of our method on four widely-used datasets, and our method performs better than existing state-of-the-art approaches, especially on motion-sensitive datasets, i.e., Something-Something V1 & V2. Our project is available at https://github.com/PeiqinZhuang/EMIM .

Method: Two-stage learning: (1) train spectral enhancement network via reverse distillation for robust background spectral representations; (2) train spatial network (rebellious student) using decorrelation losses to enforce feature orthogonality while maintaining reconstruction fidelity.

Result: Experiments on HAD100 benchmark show substantial improvements over established baselines with modest computational overhead, confirming the effectiveness of the complementary learning paradigm.

Conclusion: The proposed Rebellious Student framework successfully learns complementary spatial patterns that spectral features fail to capture, enabling parameter-free and training-free anomaly detection when paired with conventional detectors.

Abstract: A recent class of hyperspectral anomaly detection methods that can be trained once on background datasets and then universally deployed – without per-scene retraining or parameter tuning – has demonstrated remarkable efficiency and robustness. Building upon this paradigm, we focus on the integration of spectral and spatial cues and introduce a novel “Rebellious Student” framework for complementary feature learning. Unlike conventional teacher-student paradigms driven by imitation, our method intentionally trains the spatial branch to diverge from the spectral teacher, thereby learning complementary spatial patterns that the teacher fails to capture. A two-stage learning strategy is adopted: (1) a spectral enhancement network is first trained via reverse distillation to obtain robust background spectral representations; and (2) a spatial network – the rebellious student – is subsequently optimized using decorrelation losses that enforce feature orthogonality while maintaining reconstruction fidelity to avoid irrelevant noise. Once trained, the framework enhances both spectral and spatial background features, enabling parameter-free and training-free anomaly detection when paired with conventional detectors. Experiments on the HAD100 benchmark show substantial improvements over several established baselines with modest computational overhead, confirming the effectiveness of the proposed complementary learning paradigm. Our code is publicly available at https://github.com/xjpp2016/FERS.

Method: Defines VCD in the frequency space of video frame features to capture temporal information effectively, then fine-tunes video generation models using this metric within a reward-based framework.

Result: Experimental results show that fine-tuning with VCD significantly enhances temporal consistency across multiple I2V datasets without degrading other performance metrics compared to previous methods.

Conclusion: VCD effectively addresses temporal consistency limitations in image-to-video generation through frequency-domain analysis, providing a specialized reward function that improves video coherence while maintaining overall quality.

Abstract: Reward-based fine-tuning of video diffusion models is an effective approach to improve the quality of generated videos, as it can fine-tune models without requiring real-world video datasets. However, it can sometimes be limited to specific performances because conventional reward functions are mainly aimed at enhancing the quality across the whole generated video sequence, such as aesthetic appeal and overall consistency. Notably, the temporal consistency of the generated video often suffers when applying previous approaches to image-to-video (I2V) generation tasks. To address this limitation, we propose Video Consistency Distance (VCD), a novel metric designed to enhance temporal consistency, and fine-tune a model with the reward-based fine-tuning framework. To achieve coherent temporal consistency relative to a conditioning image, VCD is defined in the frequency space of video frame features to capture frame information effectively through frequency-domain analysis. Experimental results across multiple I2V datasets demonstrate that fine-tuning a video generation model with VCD significantly enhances temporal consistency without degrading other performance compared to the previous method.

Method: CBDiff uses a conditional Bernoulli diffusion model with Bernoulli noise to reflect binary/sparse properties of forgery masks, and incorporates Time-Step Cross-Attention (TSCAttention) for semantic feature guidance.

Result: Extensive experiments on eight benchmark datasets show CBDiff significantly outperforms state-of-the-art methods.

Conclusion: CBDiff demonstrates strong potential for real-world deployment by enhancing prediction credibility and mitigating error risks through multiple plausible localization maps.

Abstract: Image Forgery Localization (IFL) is a crucial task in image forensics, aimed at accurately identifying manipulated or tampered regions within an image at the pixel level. Existing methods typically generate a single deterministic localization map, which often lacks the precision and reliability required for high-stakes applications such as forensic analysis and security surveillance. To enhance the credibility of predictions and mitigate the risk of errors, we introduce an advanced Conditional Bernoulli Diffusion Model (CBDiff). Given a forged image, CBDiff generates multiple diverse and plausible localization maps, thereby offering a richer and more comprehensive representation of the forgery distribution. This approach addresses the uncertainty and variability inherent in tampered regions. Furthermore, CBDiff innovatively incorporates Bernoulli noise into the diffusion process to more faithfully reflect the inherent binary and sparse properties of forgery masks. Additionally, CBDiff introduces a Time-Step Cross-Attention (TSCAttention), which is specifically designed to leverage semantic feature guidance with temporal steps to improve manipulation detection. Extensive experiments on eight publicly benchmark datasets demonstrate that CBDiff significantly outperforms existing state-of-the-art methods, highlighting its strong potential for real-world deployment.

Method: The algorithm uses Matrix Product States (MPS) to represent spin configurations and employs a discrete driving schedule with transverse magnetic field driver Hamiltonian. It updates MPS using Density Matrix Renormalization Group (DMRG) method to iteratively minimize energy via sweeps across the spin chain.

Result: The algorithm reliably identifies global minima (not just near-optimal solutions) across diverse QUBO instances. Successfully solved intermediate-level Sudoku puzzles with over 200 Ising spins and MaxCut problems from Biq Mac library with up to 251 nodes and 3,265 edges.

Conclusion: The quantum-inspired approach demonstrates scalability, generalizability, and suitability for industrial-scale QUBO applications, providing an effective method for finding global optima in complex optimization problems.

Abstract: We propose and evaluate a quantum-inspired algorithm for solving Quadratic Unconstrained Binary Optimization (QUBO) problems, which are mathematically equivalent to finding ground states of Ising spin-glass Hamiltonians. The algorithm employs Matrix Product States (MPS) to compactly represent large superpositions of spin configurations and utilizes a discrete driving schedule to guide the MPS toward the ground state. At each step, a driver Hamiltonian – incorporating a transverse magnetic field – is combined with the problem Hamiltonian to enable spin flips and facilitate quantum tunneling. The MPS is updated using the standard Density Matrix Renormalization Group (DMRG) method, which iteratively minimizes the system’s energy via multiple sweeps across the spin chain. Despite its heuristic nature, the algorithm reliably identifies global minima, not merely near-optimal solutions, across diverse QUBO instances. We first demonstrate its effectiveness on intermediate-level Sudoku puzzles from publicly available sources, involving over $200$ Ising spins with long-range couplings dictated by constraint satisfaction. We then apply the algorithm to MaxCut problems from the Biq Mac library, successfully solving instances with up to $251$ nodes and $3,265$ edges. We discuss the advantages of this quantum-inspired approach, including its scalability, generalizability, and suitability for industrial-scale QUBO applications.

Method: Developed ARB framework with five submetrics (accuracy, reasoning reliability, uncertainty discipline, policy consistency, transparency) and evaluated four frontier models (GPT-4/5, Claude 4.5 Sonnet, Gemini 2.5 Pro, Llama 3 70B) using open technoeconomic datasets under identical conditions.

Result: GPT-4/5 and Claude 4.5 Sonnet achieved consistent and policy-compliant reasoning (Analytical Reliability Index >90), Gemini 2.5 Pro showed moderate stability, and Llama 3 70B remained below professional thresholds. Statistical validation confirmed significant and reproducible differences.

Conclusion: ARB establishes the first quantitative method for verifying causal, probabilistic, and policy-driven reasoning in AI systems for energy applications, providing a reference framework for trustworthy analytical applications in the global energy transition.

Abstract: Artificial intelligence and machine learning are increasingly used for forecasting, optimization, and policy design in the energy sector, yet no standardized framework exists to evaluate whether these systems reason correctly. Current validation practices focus on predictive accuracy or computational efficiency, leaving the logical integrity of analytical conclusions untested. This study introduces the Analytical Reliability Benchmark (ARB), a reproducible framework that quantifies reasoning reliability in large language models applied to energy system analysis. The benchmark integrates five submetrics: accuracy, reasoning reliability, uncertainty discipline, policy consistency, and transparency, and evaluates model performance across deterministic, probabilistic, and epistemic scenarios using open technoeconomic datasets (NREL ATB 2024, DOE H2A/H2New, IEA WEO 2024). Four frontier models (GPT-4/5, Claude 4.5 Sonnet, Gemini 2.5 Pro, Llama 3 70B) were tested under identical factual and regulatory conditions. Results show that reasoning reliability can be objectively measured. GPT-4/5 and Claude 4.5 Sonnet achieved consistent and policy-compliant reasoning (Analytical Reliability Index greater than 90), Gemini 2.5 Pro demonstrated moderate stability, and Llama 3 70B remained below professional thresholds. Statistical validation confirmed that these differences are significant and reproducible. The ARB establishes the first quantitative method in the energy literature for verifying causal, probabilistic, and policy-driven reasoning in artificial intelligence systems, providing a reference framework for trustworthy and transparent analytical applications in the global energy transition.

Method: Unifies structured reasoning-acting, contextual memory, and efficient execution through: (i) explicit subtask management with tree-structured exploration, (ii) web state replay with background reasoning, and (iii) page action memory for sharing explored actions.

Result: Achieved 35.8% task success rate on WebArena benchmark and reduced execution time by up to 40.4% compared to state-of-the-art methods.

Conclusion: Branch-and-Browse is a reliable and efficient framework for LLM-based web agents that demonstrates significant improvements in both success rate and execution efficiency.

Abstract: Autonomous web agents powered by large language models (LLMs) show strong potential for performing goal-oriented tasks such as information retrieval, report generation, and online transactions. These agents mark a key step toward practical embodied reasoning in open web environments. However, existing approaches remain limited in reasoning depth and efficiency: vanilla linear methods fail at multi-step reasoning and lack effective backtracking, while other search strategies are coarse-grained and computationally costly. We introduce Branch-and-Browse, a fine-grained web agent framework that unifies structured reasoning-acting, contextual memory, and efficient execution. It (i) employs explicit subtask management with tree-structured exploration for controllable multi-branch reasoning, (ii) bootstraps exploration through efficient web state replay with background reasoning, and (iii) leverages a page action memory to share explored actions within and across sessions. On the WebArena benchmark, Branch-and-Browse achieves a task success rate of 35.8% and reduces execution time by up to 40.4% relative to state-of-the-art methods. These results demonstrate that Branch-and-Browse is a reliable and efficient framework for LLM-based web agents.

Method: Model CoT as rule-based stochastic processes over directed acyclic graphs (DAGs), introduce logical closeness metric, create DAG-MATH CoT format benchmark to guide LLMs’ reasoning trajectories for evaluation.

Result: Analysis reveals statistically significant differences in reasoning fidelity among LLM families even when PASS@k metrics are comparable, highlighting gaps between final-answer accuracy and rule-consistent derivation.

Conclusion: The framework provides a balance between free-form CoT and formal proof systems, offering actionable diagnostics for evaluating LLMs’ mathematical reasoning capabilities.

Abstract: Large Language Models (LLMs) demonstrate strong performance on mathematical problems when prompted with Chain-of-Thought (CoT), yet it remains unclear whether this success stems from search, rote procedures, or rule-consistent reasoning. To address this, we propose modeling CoT as a certain rule-based stochastic process over directed acyclic graphs (DAGs), where nodes represent intermediate derivation states and edges encode rule applications. Within this framework, we introduce logical closeness, a metric that quantifies how well a model’s CoT trajectory (i.e., the LLM’s final output) adheres to the DAG structure, providing evaluation beyond classical PASS@k metrics. Building on this, we introduce the DAG-MATH CoT format and construct a benchmark that guides LLMs to generate CoT trajectories in this format, thereby enabling the evaluation of their reasoning ability under our framework. Across standard mathematical reasoning datasets, our analysis uncovers statistically significant differences in reasoning fidelity among representative LLM families-even when PASS@k is comparable-highlighting gaps between final-answer accuracy and rule-consistent derivation. Our framework provides a balance between free-form CoT and formal proofs systems, offering actionable diagnostics for LLMs reasoning evaluation. Our benchmark and code are available at: https://github.com/YuanheZ/DAG-MATH-Formatted-CoT.

Method: Surfer 2 integrates hierarchical context management, decoupled planning and execution, and self-verification with adaptive recovery to enable reliable operation over long task horizons.

Result: Achieves 97.1% accuracy on WebVoyager, 69.6% on WebArena, 60.1% on OSWorld, and 87.1% on AndroidWorld, exceeding human performance on all benchmarks with multiple attempts.

Conclusion: Systematic orchestration amplifies foundation model capabilities and enables general-purpose computer control through visual interaction alone, though next-generation vision language models are needed for Pareto-optimal cost-efficiency.

Abstract: Building agents that generalize across web, desktop, and mobile environments remains an open challenge, as prior systems rely on environment-specific interfaces that limit cross-platform deployment. We introduce Surfer 2, a unified architecture operating purely from visual observations that achieves state-of-the-art performance across all three environments. Surfer 2 integrates hierarchical context management, decoupled planning and execution, and self-verification with adaptive recovery, enabling reliable operation over long task horizons. Our system achieves 97.1% accuracy on WebVoyager, 69.6% on WebArena, 60.1% on OSWorld, and 87.1% on AndroidWorld, outperforming all prior systems without task-specific fine-tuning. With multiple attempts, Surfer 2 exceeds human performance on all benchmarks. These results demonstrate that systematic orchestration amplifies foundation model capabilities and enables general-purpose computer control through visual interaction alone, while calling for a next-generation vision language model to achieve Pareto-optimal cost-efficiency.

Method: RELATE uses shared modality-specific encoders for categorical, numerical, textual, and temporal attributes, followed by a Perceiver-style cross-attention module that aggregates features into fixed-size, permutation-invariant node representations.

Result: On the RelBench benchmark with ReLGNN and HGT, RELATE achieves performance within 3% of schema-specific encoders while reducing parameter counts by up to 5x.

Conclusion: RELATE enables varying schemas and multi-dataset pretraining for general-purpose GNNs, paving the way toward foundation models for relational graph data.

Abstract: Relational multi-table data is common in domains such as e-commerce, healthcare, and scientific research, and can be naturally represented as heterogeneous temporal graphs with multi-modal node attributes. Existing graph neural networks (GNNs) rely on schema-specific feature encoders, requiring separate modules for each node type and feature column, which hinders scalability and parameter sharing. We introduce RELATE (Relational Encoder for Latent Aggregation of Typed Entities), a schema-agnostic, plug-and-play feature encoder that can be used with any general purpose GNN. RELATE employs shared modality-specific encoders for categorical, numerical, textual, and temporal attributes, followed by a Perceiver-style cross-attention module that aggregates features into a fixed-size, permutation-invariant node representation. We evaluate RELATE on ReLGNN and HGT in the RelBench benchmark, where it achieves performance within 3% of schema-specific encoders while reducing parameter counts by up to 5x. This design supports varying schemas and enables multi-dataset pretraining for general-purpose GNNs, paving the way toward foundation models for relational graph data.

Method: UVeye layered solution for vehicle fraud detection, combining multiple verification approaches to combat AI-generated fake evidence.

Result: The solution represents a major advancement in detecting, mitigating and deterring AI-enabled insurance fraud in the vehicle sector.

Conclusion: Combating AI-driven insurance fraud remains challenging due to the evolving nature of fraud tactics and limitations of current detection systems, requiring continuous innovation in fraud prevention technology.

Abstract: Generative AI is supercharging insurance fraud by making it easier to falsify accident evidence at scale and in rapid time. Insurance fraud is a pervasive and costly problem, amounting to tens of billions of dollars in losses each year. In the vehicle insurance sector, fraud schemes have traditionally involved staged accidents, exaggerated damage, or forged documents. The rise of generative AI, including deepfake image and video generation, has introduced new methods for committing fraud at scale. Fraudsters can now fabricate highly realistic crash photos, damage evidence, and even fake identities or documents with minimal effort, exploiting AI tools to bolster false insurance claims. Insurers have begun deploying countermeasures such as AI-based deepfake detection software and enhanced verification processes to detect and mitigate these AI-driven scams. However, current mitigation strategies face significant limitations. Detection tools can suffer from false positives and negatives, and sophisticated fraudsters continuously adapt their tactics to evade automated checks. This cat-and-mouse arms race between generative AI and detection technology, combined with resource and cost barriers for insurers, means that combating AI-enabled insurance fraud remains an ongoing challenge. In this white paper, we present UVeye layered solution for vehicle fraud, representing a major leap forward in the ability to detect, mitigate and deter this new wave of fraud.

Method: Used data from 129 master’s students with 23 leadership personality characteristics from five tests. Applied exploratory data analysis, correlation analysis, and tuned seven ML algorithms (SVM, LR, KNN, DT, GB, RF, XGBoost, LightGBM) with feature selection using Pearson correlation.

Result: Random Forest classifier achieved highest performance: 87.50% accuracy with 17 personality traits plus leadership mark, and 85.71% accuracy without leadership mark.

Conclusion: The study provides an effective method to identify students’ strengths/weaknesses early and select personalized learning strategies using AI and personality traits.

Abstract: The study explores the potential of AI technologies in personalized learning, suggesting the prediction of academic success through leadership personality traits and machine learning modelling. The primary data were obtained from 129 master’s students in the Environmental Engineering Department, who underwent five leadership personality tests with 23 characteristics. Students used self-assessment tools that included Personality Insight, Workplace Culture, Motivation at Work, Management Skills, and Emotion Control tests. The test results were combined with the average grade obtained from academic reports. The study employed exploratory data analysis and correlation analysis. Feature selection utilized Pearson correlation coefficients of personality traits. The average grades were separated into three categories: fail, pass, and excellent. The modelling process was performed by tuning seven ML algorithms, such as SVM, LR, KNN, DT, GB, RF, XGBoost and LightGBM. The highest predictive performance was achieved with the RF classifier, which yielded an accuracy of 87.50% for the model incorporating 17 personality trait features and the leadership mark feature, and an accuracy of 85.71% for the model excluding this feature. In this way, the study offers an additional opportunity to identify students’ strengths and weaknesses at an early stage of their education process and select the most suitable strategies for personalized learning.

Method: A simple and efficient protocol using modest 8-billion-parameter open-source LLMs to encode and decode messages within coherent text of the same length.

Result: High-quality results achieved with local processing on a laptop in seconds, enabling scenarios like covert deployment of unfiltered LLMs within safe model responses.

Conclusion: This protocol demonstrates radical decoupling of text from intent, challenges understanding of LLM knowledge, and raises urgent AI safety questions.

Abstract: A meaningful text can be hidden inside another, completely different yet still coherent and plausible, text of the same length. For example, a tweet containing a harsh political critique could be embedded in a tweet that celebrates the same political leader, or an ordinary product review could conceal a secret manuscript. This uncanny state of affairs is now possible thanks to Large Language Models, and in this paper we present a simple and efficient protocol to achieve it. We show that even modest 8-billion-parameter open-source LLMs are sufficient to obtain high-quality results, and a message as long as this abstract can be encoded and decoded locally on a laptop in seconds. The existence of such a protocol demonstrates a radical decoupling of text from authorial intent, further eroding trust in written communication, already shaken by the rise of LLM chatbots. We illustrate this with a concrete scenario: a company could covertly deploy an unfiltered LLM by encoding its answers within the compliant responses of a safe model. This possibility raises urgent questions for AI safety and challenges our understanding of what it means for a Large Language Model to know something.

Method: Uses component-based orchestration for deterministic routing between LLMs, hybrid retrieval with OpenSearch and finance-domain embeddings, and multi-stage recommendation combining rule heuristics, behavioral modeling, and contextual bandits. Deployed on-premises with Docker Swarm and vLLM across 24 NVIDIA H100 GPUs.

Result: Demonstrated through human QA and system metrics that grounded generation with explicit routing and layered safety can deliver trustworthy AI insights in financial applications.

Conclusion: The system successfully shows that proactive, grounded generation with proper safety measures can enable trustworthy AI deployment in regulated financial environments.

Abstract: We present AI PB, a production-scale generative agent deployed in real retail finance. Unlike reactive chatbots that answer queries passively, AI PB proactively generates grounded, compliant, and user-specific investment insights. It integrates (i) a component-based orchestration layer that deterministically routes between internal and external LLMs based on data sensitivity, (ii) a hybrid retrieval pipeline using OpenSearch and the finance-domain embedding model, and (iii) a multi-stage recommendation mechanism combining rule heuristics, sequential behavioral modeling, and contextual bandits. Operating fully on-premises under Korean financial regulations, the system employs Docker Swarm and vLLM across 24 X NVIDIA H100 GPUs. Through human QA and system metrics, we demonstrate that grounded generation with explicit routing and layered safety can deliver trustworthy AI insights in high-stakes finance.

Method: A multi-agent system with three roles (Parsing, Detection, Explanation) in a conversational workflow, using XGBoost as the baseline detector and providing narrative explanations grounded in underlying features through an open-source web UI.

Result: The baseline detector achieved strong accuracy on a Bitcoin mixing dataset (Wasabi Wallet, 2020-2024), while HCLA added interpretability and interactive refinement capabilities.

Conclusion: Human-in-the-loop design improves transparency and trust in financial forensics by enabling natural language interaction and context-aware explanations for anomaly detection.

Abstract: We present HCLA, a human-centered multi-agent system for anomaly detection in digital asset transactions. The system links three roles: Parsing, Detection, and Explanation, into a conversational workflow that lets non-experts ask questions in natural language, inspect structured analytics, and obtain context-aware rationales. Implemented with an open-source web UI, HCLA translates user intents into a schema for a classical detector (XGBoost in our prototype) and returns narrative explanations grounded in the underlying features. On a labeled Bitcoin mixing dataset (Wasabi Wallet, 2020-2024), the baseline detector reaches strong accuracy, while HCLA adds interpretability and interactive refinement. We describe the architecture, interaction loop, dataset, evaluation protocol, and limitations, and discuss how a human-in-the-loop design improves transparency and trust in financial forensics.

Method: Proposed a lesson-based collaboration framework with lesson solicitation, banking, and selection mechanisms where agents share knowledge gained from their solution attempts.

Result: A team of small LLMs with lessons learned outperformed a much larger LLM and other multi-LLM collaboration methods.

Conclusion: Lesson-based collaboration enables effective knowledge sharing among LLM agents, allowing smaller models to collectively achieve better performance than larger individual models through complementary learning.

Abstract: Recent studies show that LLMs possess different skills and specialize in different tasks. In fact, we observe that their varied performance occur in several levels of granularity. For example, in the code optimization task, code LLMs excel at different optimization categories and no one dominates others. This observation prompts the question of how one leverages multiple LLM agents to solve a coding problem without knowing their complementary strengths a priori. We argue that a team of agents can learn from each other’s successes and failures so as to improve their own performance. Thus, a lesson is the knowledge produced by an agent and passed on to other agents in the collective solution process. We propose a lesson-based collaboration framework, design the lesson solicitation–banking–selection mechanism, and demonstrate that a team of small LLMs with lessons learned can outperform a much larger LLM and other multi-LLM collaboration methods.

Method: Proposes a new paradigm for evaluating AI use in legal practice based on the verification-value paradox, considering AI’s limitations and lawyers’ ethical duties.

Result: Identifies that efficiency gains from AI are counterbalanced by greater verification requirements, resulting in minimal net value for lawyers.

Conclusion: Legal practice and education need to emphasize fidelity to truth and civic responsibility, with AI use requiring careful consideration of the verification-value paradox.

Abstract: It is often claimed that machine learning-based generative AI products will drastically streamline and reduce the cost of legal practice. This enthusiasm assumes lawyers can effectively manage AI’s risks. Cases in Australia and elsewhere in which lawyers have been reprimanded for submitting inaccurate AI-generated content to courts suggest this paradigm must be revisited. This paper argues that a new paradigm is needed to evaluate AI use in practice, given (a) AI’s disconnection from reality and its lack of transparency, and (b) lawyers’ paramount duties like honesty, integrity, and not to mislead the court. It presents an alternative model of AI use in practice that more holistically reflects these features (the verification-value paradox). That paradox suggests increases in efficiency from AI use in legal practice will be met by a correspondingly greater imperative to manually verify any outputs of that use, rendering the net value of AI use often negligible to lawyers. The paper then sets out the paradox’s implications for legal practice and legal education, including for AI use but also the values that the paradox suggests should undergird legal practice: fidelity to the truth and civic responsibility.

Method: TRUST uses a consensus mechanism among diverse auditors, hierarchical DAG decomposition of reasoning traces, blockchain ledger for transparency, and privacy-preserving segmentation to protect proprietary logic.

Result: Experiments across multiple LLMs and reasoning tasks show TRUST effectively detects reasoning flaws and remains robust against adversarial auditors, with theoretical guarantees for security and economic incentives.

Conclusion: TRUST pioneers decentralized AI auditing, providing a practical path toward safe and trustworthy LLM deployment by overcoming limitations of centralized auditing approaches.

Abstract: Large Language Models generate complex reasoning chains that reveal their decision-making, yet verifying the faithfulness and harmlessness of these intermediate steps remains a critical unsolved problem. Existing auditing methods are centralized, opaque, and hard to scale, creating significant risks for deploying proprietary models in high-stakes domains. We identify four core challenges: (1) Robustness: Centralized auditors are single points of failure, prone to bias or attacks. (2) Scalability: Reasoning traces are too long for manual verification. (3) Opacity: Closed auditing undermines public trust. (4) Privacy: Exposing full reasoning risks model theft or distillation. We propose TRUST, a transparent, decentralized auditing framework that overcomes these limitations via: (1) A consensus mechanism among diverse auditors, guaranteeing correctness under up to $30%$ malicious participants. (2) A hierarchical DAG decomposition of reasoning traces, enabling scalable, parallel auditing. (3) A blockchain ledger that records all verification decisions for public accountability. (4) Privacy-preserving segmentation, sharing only partial reasoning steps to protect proprietary logic. We provide theoretical guarantees for the security and economic incentives of the TRUST framework. Experiments across multiple LLMs (GPT-OSS, DeepSeek-r1, Qwen) and reasoning tasks (math, medical, science, humanities) show TRUST effectively detects reasoning flaws and remains robust against adversarial auditors. Our work pioneers decentralized AI auditing, offering a practical path toward safe and trustworthy LLM deployment.

Method: Formalized the lock-in phase concept, linked it to learning dynamics, proposed operational metrics for onset detection, and conducted experiments to observe behavioral consolidation patterns across different model scales.

Result: Found rapid non-linear behavioral consolidation with varying side-effects: performance trade-offs in small models, largely cost-free adoption in mid-scale models, and transient instabilities in large quantized models.

Conclusion: Identity consolidation is a prerequisite for AGI-level reliability and a critical safety control point - identities can be engineered for reliability but may also emerge spontaneously during scaling, potentially hardening unpredictable behaviors.

Abstract: Large language models (LLMs) remain broadly open and highly steerable: they imitate at scale, accept arbitrary system prompts, and readily adopt multiple personae. By analogy to human development, we hypothesize that progress toward artificial general intelligence (AGI) involves a lock-in phase: a transition from open imitation to identity consolidation, in which goal structures, refusals, preferences, and internal representations become comparatively stable and resistant to external steering. We formalize this phase, link it to known phenomena in learning dynamics, and propose operational metrics for onset detection. Experimentally, we demonstrate that while the behavioral consolidation is rapid and non-linear, its side-effects on general capabilities are not monolithic. Our results reveal a spectrum of outcomes–from performance trade-offs in small models, through largely cost-free adoption in mid-scale models, to transient instabilities in large, quantized models. We argue that such consolidation is a prerequisite for AGI-level reliability and also a critical control point for safety: identities can be deliberately engineered for reliability, yet may also emerge spontaneously during scaling, potentially hardening unpredictable goals and behaviors.

Method: Proposed RADAR framework with multi-agent collaboration using four specialized roles and multi-round debate mechanisms. Decomposes risk concept space into explicit, implicit, and non-risk subspaces with dynamic update mechanisms for self-evolution.

Result: RADAR significantly outperforms baseline methods, achieving 28.87% improvement in risk identification accuracy on a challenging testset of 800 cases and public benchmarks, with better accuracy, stability, and self-evaluation risk sensitivity.

Conclusion: The RADAR framework provides a more robust and comprehensive approach to LLM safety evaluation by addressing inherent limitations of existing methods through theoretical risk space reconstruction and multi-agent collaborative assessment.

Abstract: Existing safety evaluation methods for large language models (LLMs) suffer from inherent limitations, including evaluator bias and detection failures arising from model homogeneity, which collectively undermine the robustness of risk evaluation processes. This paper seeks to re-examine the risk evaluation paradigm by introducing a theoretical framework that reconstructs the underlying risk concept space. Specifically, we decompose the latent risk concept space into three mutually exclusive subspaces: the explicit risk subspace (encompassing direct violations of safety guidelines), the implicit risk subspace (capturing potential malicious content that requires contextual reasoning for identification), and the non-risk subspace. Furthermore, we propose RADAR, a multi-agent collaborative evaluation framework that leverages multi-round debate mechanisms through four specialized complementary roles and employs dynamic update mechanisms to achieve self-evolution of risk concept distributions. This approach enables comprehensive coverage of both explicit and implicit risks while mitigating evaluator bias. To validate the effectiveness of our framework, we construct an evaluation dataset comprising 800 challenging cases. Extensive experiments on our challenging testset and public benchmarks demonstrate that RADAR significantly outperforms baseline evaluation methods across multiple dimensions, including accuracy, stability, and self-evaluation risk sensitivity. Notably, RADAR achieves a 28.87% improvement in risk identification accuracy compared to the strongest baseline evaluation method.

Method: Implemented two systems: two-agent metaprompting (thinker LLM refines strategies for executor LLM) and single-agent system with value function refinement for Monte Carlo Tree Search, plus rollback features.

Result: Single-agent system achieved average improvement of 473.2 points per cycle with strong upward trend (ρ=0.607), while two-agent system showed minimal improvement.

Conclusion: Evolutionary refinement techniques show promise for AI in non-deterministic environments, but metaprompting has inherent limits compared to value function optimization.

Abstract: Optimizing artificial intelligence (AI) for dynamic environments remains a fundamental challenge in machine learning research. In this paper, we examine evolutionary training methods for optimizing AI to solve the game 2048, a 2D sliding puzzle. 2048, with its mix of strategic gameplay and stochastic elements, presents an ideal playground for studying decision-making, long-term planning, and dynamic adaptation. We implemented two distinct systems: a two-agent metaprompting system where a “thinker” large language model (LLM) agent refines gameplay strategies for an “executor” LLM agent, and a single-agent system based on refining a value function for a limited Monte Carlo Tree Search. We also experimented with rollback features to avoid performance degradation. Our results demonstrate the potential of evolutionary refinement techniques in improving AI performance in non-deterministic environments. The single-agent system achieved substantial improvements, with an average increase of 473.2 points per cycle, and with clear upward trends (correlation $\rho$=0.607) across training cycles. The LLM’s understanding of the game grew as well, shown in its development of increasingly advanced strategies. Conversely, the two-agent system did not garner much improvement, highlighting the inherent limits of meta-prompting.

Method: During reasoning: models construct payoff matrices to estimate welfare for both LLM and user, selecting mutually beneficial actions. During training: introduces mutual welfare reward to reinforce cooperative responses. Also includes inference technique for dynamic adaptation to pricing policy changes.

Result: Extensive experiments show GTAlign substantially improves reasoning efficiency, answer quality, and mutual welfare compared to baselines across diverse tasks.

Conclusion: Game-theoretic alignment provides a principled decision-making mechanism that benefits both LLMs and users, addressing the fundamental challenge of suboptimal interactions in conventional alignment approaches.

Abstract: Large Language Models (LLMs) have achieved remarkable progress in reasoning, yet sometimes produce responses that are suboptimal for users in tasks such as writing, information seeking, or providing practical guidance. Conventional alignment practices typically assume that maximizing model reward also maximizes user welfare, but this assumption frequently fails in practice: models may over-clarify or generate overly verbose reasoning when users prefer concise answers. Such behaviors resemble the prisoner’s dilemma, where individually rational choices lead to socially suboptimal outcomes. The fundamental challenge is the lack of a principled decision making mechanism that mutually benefits both the LLM and the user. We propose Game-Theoretic Alignment (GTAlign), an alignment framework that integrates game-theoretic decision making into both reasoning and training. During reasoning, the model explicitly treats user-LLM interaction as a strategic game: it constructs payoff matrices within its reasoning chain to estimate welfare for both itself and the user, and then selects actions that are mutually beneficial. During training, we introduce a mutual welfare reward that reinforces cooperative responses, aligning model behavior with socially efficient outcomes. In addition, we introduce an inference technique that leverages game-theoretic reasoning to dynamically adapt LLM’s response when pricing policies of LLM service change. Extensive experiments demonstrate that GTAlign substantially improves reasoning efficiency, answer quality, and mutual welfare compared to baselines across diverse tasks. The code is available at https://github.com/ulab-uiuc/GTAlign .

Method: Introduced a novel task evaluating cognitive representation methods in ICS using a dataset from recent novels and an 11-condition cognitive evaluation framework to benchmark seven LLMs in authorial style emulation.

Result: Combining conceptual and linguistic features was most effective in ICS, outperforming static profile-based cues. LLMs were more effective at mimicking linguistic style than narrative structure.

Conclusion: The findings provide a foundation for developing AI systems that adapt to individual ways of thinking and expression, advancing personalized and human-aligned creative technologies.

Abstract: Individualized cognitive simulation (ICS) aims to build computational models that approximate the thought processes of specific individuals. While large language models (LLMs) convincingly mimic surface-level human behavior such as role-play, their ability to simulate deeper individualized cognitive processes remains poorly understood. To address this gap, we introduce a novel task that evaluates different cognitive representation methods in ICS. We construct a dataset from recently published novels (later than the release date of the tested LLMs) and propose an 11-condition cognitive evaluation framework to benchmark seven off-the-shelf LLMs in the context of authorial style emulation. We hypothesize that effective cognitive representations can help LLMs generate storytelling that better mirrors the original author. Thus, we test different cognitive representations, e.g., linguistic features, concept mappings, and profile-based information. Results show that combining conceptual and linguistic features is particularly effective in ICS, outperforming static profile-based cues in overall evaluation. Importantly, LLMs are more effective at mimicking linguistic style than narrative structure, underscoring their limits in deeper cognitive simulation. These findings provide a foundation for developing AI systems that adapt to individual ways of thinking and expression, advancing more personalized and human-aligned creative technologies.

Method: Uses MAIDs as a graphical framework to analyze MARL interaction paradigms. Introduces targeted intervention paradigm applied to single agents using Pre-Strategy Intervention (PSI) causal inference technique. Employs relevance graph analysis to verify workability.

Result: Demonstrates effectiveness of targeted intervention paradigm in experiments. Verifies results of relevance graph analysis, showing the approach can achieve composite desired outcomes by maximizing causal effects.

Conclusion: MAIDs provide an effective graphical framework for analyzing and designing MARL interaction paradigms. The targeted intervention with PSI enables steering multi-agent systems toward desired outcomes without requiring impractical global guidance.

Abstract: Steering cooperative multi-agent reinforcement learning (MARL) towards desired outcomes is challenging, particularly when the global guidance from a human on the whole multi-agent system is impractical in a large-scale MARL. On the other hand, designing external mechanisms (e.g., intrinsic rewards and human feedback) to coordinate agents mostly relies on empirical studies, lacking a easy-to-use research tool. In this work, we employ multi-agent influence diagrams (MAIDs) as a graphical framework to address the above issues. First, we introduce the concept of MARL interaction paradigms, using MAIDs to analyze and visualize both unguided self-organization and global guidance mechanisms in MARL. Then, we design a new MARL interaction paradigm, referred to as the targeted intervention paradigm that is applied to only a single targeted agent, so the problem of global guidance can be mitigated. In our implementation, we introduce a causal inference technique, referred to as Pre-Strategy Intervention (PSI), to realize the targeted intervention paradigm. Since MAIDs can be regarded as a special class of causal diagrams, a composite desired outcome that integrates the primary task goal and an additional desired outcome can be achieved by maximizing the corresponding causal effect through the PSI. Moreover, the bundled relevance graph analysis of MAIDs provides a tool to identify whether an MARL learning paradigm is workable under the design of an MARL interaction paradigm. In experiments, we demonstrate the effectiveness of our proposed targeted intervention, and verify the result of relevance graph analysis.

Method: Use in-context learning with LLMs to generate abstract PDDL domains from natural language objectives, validated by symbolic tools and human experts.

Result: GPT-4o successfully synthesizes useful planning domain abstractions in simple settings, with better performance on action abstraction than fluent abstraction.

Conclusion: LLMs show promise for domain abstraction generation but need improvement in handling fluent abstraction.

Abstract: Generating an abstraction of a dynamic domain that aligns with a given purpose remains a significant challenge given that the choice of such an abstraction can impact an agent’s ability to plan, reason, and provide explanations effectively. We model the agent’s concrete behaviors in PDDL and investigate the use of in-context learning with large language models (LLMs) for the generation of abstract PDDL domains and problem instances, given an abstraction objective specified in natural language. The benchmark examples we use are new and have not been part of the data any LLMs have been trained on. We consider three categories of abstractions: abstraction of choice of alternative concrete actions, abstraction of sequences of concrete actions, and abstraction of action/predicate parameters, as well as combinations of these. The generated abstract PDDL domains and problem instances are then checked by symbolic validation tools as well as human experts. Our experiments show that GPT-4o can generally synthesize useful planning domain abstractions in simple settings, although it is better at abstracting over actions than over the associated fluents.

Method: Enriched BERT-based model with derived temporal descriptors and POI embeddings to create unified, semantically enriched mobility representations.

Result: Significant improvement in prediction accuracy: GEO-BLEU score increased from 0.34 to 0.75 for single-city prediction and from 0.34 to 0.56 for multi-city prediction.

Conclusion: Integrating both POI and temporal information at each location creates more effective mobility representations, substantially improving forecasting performance.

Abstract: Human mobility forecasting is crucial for disaster relief, city planning, and public health. However, existing models either only model location sequences or include time information merely as auxiliary input, thereby failing to leverage the rich semantic context provided by points of interest (POIs). To address this, we enrich a BERT-based mobility model with derived temporal descriptors and POI embeddings to better capture the semantics underlying human movement. We propose STaBERT (Semantic-Temporal aware BERT), which integrates both POI and temporal information at each location to construct a unified, semantically enriched representation of mobility. Experimental results show that STaBERT significantly improves prediction accuracy: for single-city prediction, the GEO-BLEU score improved from 0.34 to 0.75; for multi-city prediction, from 0.34 to 0.56.

Method: ToolEQA integrates external tools with multi-step reasoning, where tools provide useful information to guide exploration. A novel EQA data generation pipeline automatically constructs large-scale EQA tasks with reasoning trajectories, resulting in the EQA-RT dataset with 18K tasks.

Result: ToolEQA improves success rate by 9.2-20.2% over state-of-the-art baselines and outperforms zero-shot ToolEQA by 10%. It also achieves SOTA performance on HM-EQA, OpenEQA, and EXPRESS-Bench datasets, demonstrating strong generalization.

Conclusion: ToolEQA effectively enhances embodied question answering by combining external tools with multi-step reasoning, leading to more efficient exploration and more accurate responses across multiple benchmark datasets.

Abstract: Embodied Question Answering (EQA) requires agents to explore 3D environments to obtain observations and answer questions related to the scene. Existing methods leverage VLMs to directly explore the environment and answer questions without explicit thinking or planning, which limits their reasoning ability and results in excessive or inefficient exploration as well as ineffective responses. In this paper, we introduce ToolEQA, an agent that integrates external tools with multi-step reasoning, where external tools can provide more useful information for completing the task, helping the model derive better exploration directions in the next step of reasoning and thus obtaining additional effective information. This enables ToolEQA to generate more accurate responses with a shorter exploration distance. To enhance the model’s ability for tool-usage and multi-step reasoning, we further design a novel EQA data generation pipeline that automatically constructs large-scale EQA tasks with reasoning trajectories and corresponding answers. Based on the pipeline, we collect the EQA-RT dataset that contains about 18K tasks, divided into a training set EQA-RT-Train, and two test sets EQA-RT-Seen (scenes overlapping with the training set) and EQA-RT-Unseen (novel scenes). Experiments on EQA-RT-Seen and EQA-RT-Unseen show that ToolEQA improves the success rate by 9.2~20.2% over state-of-the-art baselines, while outperforming the zero-shot ToolEQA by 10% in success rate. In addition, ToolEQA also achieves state-of-the-art performance on the HM-EQA, OpenEQA, and EXPRESS-Bench datasets, demonstrating its generality. Our homepage see https://tooleqa.github.io.

Method: Analysis of biases in the AI4HealthyAging project, identifying historical, representation, and measurement biases across multiple use cases.

Result: Identified biases manifesting in variables like sex, gender, age, habitat, socioeconomic status, equipment, and labeling.

Conclusion: Provides practical recommendations for improving fairness and robustness in clinical problem design and data collection to guide development of fairer AI healthcare systems.

Abstract: Artificial intelligence (AI) holds great promise for transforming healthcare. However, despite significant advances, the integration of AI solutions into real-world clinical practice remains limited. A major barrier is the quality and fairness of training data, which is often compromised by biased data collection practices. This paper draws on insights from the AI4HealthyAging project, part of Spain’s national R&D initiative, where our task was to detect biases during clinical data collection. We identify several types of bias across multiple use cases, including historical, representation, and measurement biases. These biases manifest in variables such as sex, gender, age, habitat, socioeconomic status, equipment, and labeling. We conclude with practical recommendations for improving the fairness and robustness of clinical problem design and data collection. We hope that our findings and experience contribute to guiding future projects in the development of fairer AI systems in healthcare.

Method: Design science methodological approach with layered KRR architecture that captures AI system categories, architectural components, engaging vectors, and contextual aspects. Includes spreading, severity, likelihood, and evaluation metrics with transparent reasoning mechanisms.

Result: The model is demonstrated and evaluated through instantiation, providing a basis for building responsible and trustworthy intelligent systems for assessing engagement effects.

Conclusion: The proposed model offers a comprehensive framework for collateral damage assessment in AI-enabled military operations, supporting the development of responsible and trustworthy intelligent systems.

Abstract: In an era where AI (Artificial Intelligence) systems play an increasing role in the battlefield, ensuring responsible targeting demands rigorous assessment of potential collateral effects. In this context, a novel collateral damage assessment model for target engagement of AI systems in military operations is introduced. The model integrates temporal, spatial, and force dimensions within a unified Knowledge Representation and Reasoning (KRR) architecture following a design science methodological approach. Its layered structure captures the categories and architectural components of the AI systems to be engaged together with corresponding engaging vectors and contextual aspects. At the same time, spreading, severity, likelihood, and evaluation metrics are considered in order to provide a clear representation enhanced by transparent reasoning mechanisms. Further, the model is demonstrated and evaluated through instantiation which serves as a basis for further dedicated efforts that aim at building responsible and trustworthy intelligent systems for assessing the effects produced by engaging AI systems in military operations.

Method: The survey reviews emerging LLM-driven approaches from two complementary perspectives: schema-based paradigms (emphasizing structure and consistency) and schema-free paradigms (highlighting flexibility and open discovery), while synthesizing representative frameworks and analyzing their technical mechanisms.

Result: The survey provides a systematic analysis of how LLMs are enabling more adaptive, explainable, and intelligent knowledge systems by transforming knowledge graph construction methodologies.

Conclusion: The evolving interplay between LLMs and knowledge graphs represents a paradigm shift toward developing adaptive, explainable, and intelligent knowledge systems, with future directions including KG-based reasoning for LLMs, dynamic knowledge memory for agentic systems, and multimodal KG construction.

Abstract: Knowledge Graphs (KGs) have long served as a fundamental infrastructure for structured knowledge representation and reasoning. With the advent of Large Language Models (LLMs), the construction of KGs has entered a new paradigm-shifting from rule-based and statistical pipelines to language-driven and generative frameworks. This survey provides a comprehensive overview of recent progress in LLM-empowered knowledge graph construction, systematically analyzing how LLMs reshape the classical three-layered pipeline of ontology engineering, knowledge extraction, and knowledge fusion. We first revisit traditional KG methodologies to establish conceptual foundations, and then review emerging LLM-driven approaches from two complementary perspectives: schema-based paradigms, which emphasize structure, normalization, and consistency; and schema-free paradigms, which highlight flexibility, adaptability, and open discovery. Across each stage, we synthesize representative frameworks, analyze their technical mechanisms, and identify their limitations. Finally, the survey outlines key trends and future research directions, including KG-based reasoning for LLMs, dynamic knowledge memory for agentic systems, and multimodal KG construction. Through this systematic review, we aim to clarify the evolving interplay between LLMs and knowledge graphs, bridging symbolic knowledge engineering and neural semantic understanding toward the development of adaptive, explainable, and intelligent knowledge systems.

Method: Proposes IKnow framework with novel self-supervised objectives in instruction-response dialogue format that leverages domain knowledge embedded within the text itself.

Result: The method learns to encode domain knowledge at a deeper semantic level without external resources.

Conclusion: IKnow provides a simple and general framework for continual adaptation of LLMs to new domains using only unlabeled test-time data.

Abstract: Continual pretraining promises to adapt large language models (LLMs) to new domains using only unlabeled test-time data, but naively applying standard self-supervised objectives to instruction-tuned models is known to degrade their instruction-following capability and semantic representations. Existing fixes assume access to the original base model or rely on knowledge from an external domain-specific database - both of which pose a realistic barrier in settings where the base model weights are withheld for safety reasons or reliable external corpora are unavailable. In this work, we propose Instruction-Knowledge-Aware Continual Adaptation (IKnow), a simple and general framework that formulates novel self-supervised objectives in the instruction-response dialogue format. Rather than depend- ing on external resources, IKnow leverages domain knowledge embedded within the text itself and learns to encode it at a deeper semantic level.

Method: Implemented five functions based on creativity theories and techniques to generate innovation opportunities, then evaluated the model using opportunities generated for a hospitality sector innovation project.

Result: The computational model generated outcomes that were more novel and/or useful than those from Notebook LM and ChatGPT4o, but not all model functions contributed to increased novelty.

Conclusion: The model shows promise for generating novel innovation opportunities but requires further development to optimize all functions for novelty enhancement.

Abstract: This paper presents a new computational model of creative outcomes, informed by creativity theories and techniques, which was implemented to generate more novel opportunities for innovation projects. The model implemented five functions that were developed to contribute to the generation of innovation opportunities with higher novelty without loss of usefulness. The model was evaluated using opportunities generated for an innovation project in the hospitality sector. The evaluation revealed that the computational model generated outcomes that were more novel and/or useful than outcomes from Notebook LM and ChatGPT4o. However, not all model functions contributed to the generation of more novel opportunities, leading to new directions for further model development

Method: EBR uses embeddings to approximate symbolic reasoning results, requiring only retrieval of instances for atomic concepts and existential restrictions to approximate instance sets for any concept in description logic SHIQ.

Result: EBR demonstrates robustness against missing and erroneous data compared to state-of-the-art reasoners in experimental evaluations.

Conclusion: EBR provides a viable neural reasoning approach that overcomes the limitations of traditional symbolic reasoners for concept learning on real-world knowledge bases.

Abstract: Concept learning exploits background knowledge in the form of description logic axioms to learn explainable classification models from knowledge bases. Despite recent breakthroughs in neuro-symbolic concept learning, most approaches still cannot be deployed on real-world knowledge bases. This is due to their use of description logic reasoners, which are not robust against inconsistencies nor erroneous data. We address this challenge by presenting a novel neural reasoner dubbed EBR. Our reasoner relies on embeddings to approximate the results of a symbolic reasoner. We show that EBR solely requires retrieving instances for atomic concepts and existential restrictions to retrieve or approximate the set of instances of any concept in the description logic $\mathcal{SHOIQ}$. In our experiments, we compare EBR with state-of-the-art reasoners. Our results suggest that EBR is robust against missing and erroneous data in contrast to existing reasoners.

Method: FLORA uses fuzzy logic to iteratively align both entities and relations across knowledge graphs in a holistic manner, allowing for dangling entities and providing interpretable reasoning.

Result: The method achieves state-of-the-art results on major benchmarks while being unsupervised and providing interpretable alignments.

Conclusion: FLORA demonstrates that unsupervised knowledge graph alignment with interpretable reasoning is feasible and can outperform supervised methods on major benchmarks.

Abstract: Knowledge graph alignment is the task of matching equivalent entities (that is, instances and classes) and relations across two knowledge graphs. Most existing methods focus on pure entity-level alignment, computing the similarity of entities in some embedding space. They lack interpretable reasoning and need training data to work. In this paper, we propose FLORA, a simple yet effective method that (1) is unsupervised, i.e., does not require training data, (2) provides a holistic alignment for entities and relations iteratively, (3) is based on fuzzy logic and thus delivers interpretable results, (4) provably converges, (5) allows dangling entities, i.e., entities without a counterpart in the other KG, and (6) achieves state-of-the-art results on major benchmarks.

Method: Comparative landscape analysis of three countries (Australia, Colombia, United States) examining how governments solicited public feedback on AI risks and policies and whether that input shaped governance.

Result: In all three cases, fewer than 1% of the population participated, governments did little to attract diverse voices or publicize calls for comment, and officials showed limited responsiveness to feedback, failing to create effective feedback loops.

Conclusion: Current approaches are unlikely to build trust or legitimacy in AI governance because policymakers are not adequately listening or responding to public concerns. Eight recommendations are offered to improve participatory processes.

Abstract: The worlds people have strong opinions about artificial intelligence (AI), and they want policymakers to listen. Governments are inviting public comment on AI, but as they translate input into policy, much of what citizens say is lost. Policymakers are missing a critical opportunity to build trust in AI and its governance. This paper compares three countries, Australia, Colombia, and the United States, that invited citizens to comment on AI risks and policies. Using a landscape analysis, the authors examined how each government solicited feedback and whether that input shaped governance. Yet in none of the three cases did citizens and policymakers establish a meaningful dialogue. Governments did little to attract diverse voices or publicize calls for comment, leaving most citizens unaware or unprepared to respond. In each nation, fewer than one percent of the population participated. Moreover, officials showed limited responsiveness to the feedback they received, failing to create an effective feedback loop. The study finds a persistent gap between the promise and practice of participatory AI governance. The authors conclude that current approaches are unlikely to build trust or legitimacy in AI because policymakers are not adequately listening or responding to public concerns. They offer eight recommendations: promote AI literacy; monitor public feedback; broaden outreach; hold regular online forums; use innovative engagement methods; include underrepresented groups; respond publicly to input; and make participation easier.

Method: Developed a heterogeneous edge-aware graph neural network to predict effective busbar splitting actions, capturing local flow patterns and enabling generalization to unseen topology changes.

Result: Achieved up to 4 orders-of-magnitude speed-up, delivering AC-feasible solutions within one minute with 2.3% optimality gap on the GOC 2000-bus system.

Conclusion: The proposed approach demonstrates significant progress toward near-real-time NTO for large-scale systems with topology and cross-system generalization capabilities.

Abstract: Network topology optimization (NTO) via busbar splitting can mitigate transmission grid congestion and reduce redispatch costs. However, solving this mixed-integer non-linear problem for large-scale systems in near-real-time is currently intractable with existing solvers. Machine learning (ML) approaches have emerged as a promising alternative, but they have limited generalization to unseen topologies, varying operating conditions, and different systems, which limits their practical applicability. This paper formulates NTO for congestion management problem considering linearized AC PF, and proposes a graph neural network (GNN)-accelerated approach. We develop a heterogeneous edge-aware message passing NN to predict effective busbar splitting actions as candidate NTO solutions. The proposed GNN captures local flow patterns, achieves generalization to unseen topology changes, and improves transferability across systems. Case studies show up to 4 orders-of-magnitude speed-up, delivering AC-feasible solutions within one minute and a 2.3% optimality gap on the GOC 2000-bus system. These results demonstrate a significant step toward near-real-time NTO for large-scale systems with topology and cross-system generalization.

Method: Introduced causal stepwise evaluation (CaSE) that decomposes reasoning quality into relevance (grounded in problem) and coherence (logically follows prior steps). Each step is evaluated using only its preceding context to avoid hindsight bias.

Result: Validated CaSE against human judgments on expert-annotated benchmarks MRa-GSM8K and MRa-MATH. Showed that curating training data with CaSE-evaluated relevance and coherence directly improves final task performance.

Conclusion: CaSE provides a scalable framework for analyzing, debugging, and improving LLM reasoning, demonstrating practical value beyond simple validity checks.

Abstract: Evaluating large language models (LLMs) on final-answer correctness is the dominant paradigm. This approach, however, provides a coarse signal for model improvement and overlooks the quality of the underlying reasoning process. We argue that a more granular evaluation of reasoning offers a more effective path to building robust models. We decompose reasoning quality into two dimensions: relevance and coherence. Relevance measures if a step is grounded in the problem; coherence measures if it follows logically from prior steps. To measure these aspects reliably, we introduce causal stepwise evaluation (CaSE). This method assesses each reasoning step using only its preceding context, which avoids hindsight bias. We validate CaSE against human judgments on our new expert-annotated benchmarks, MRa-GSM8K and MRa-MATH. More importantly, we show that curating training data with CaSE-evaluated relevance and coherence directly improves final task performance. Our work provides a scalable framework for analyzing, debugging, and improving LLM reasoning, demonstrating the practical value of moving beyond validity checks.

Method: Two algorithms: one using approximate Cholesky factorization and sparse inverse estimation, and another using sampling of rooted spanning trees.

Result: Both algorithms achieve near-linear time complexity and provide strong approximation guarantees, validated on large real-world networks with over 10 million nodes.

Conclusion: The proposed methods enable efficient computation of random walk centrality at scale while maintaining approximation quality.

Abstract: Random walk centrality is a fundamental metric in graph mining for quantifying node importance and influence, defined as the weighted average of hitting times to a node from all other nodes. Despite its ability to capture rich graph structural information and its wide range of applications, computing this measure for large networks remains impractical due to the computational demands of existing methods. In this paper, we present a novel formulation of random walk centrality, underpinning two scalable algorithms: one leveraging approximate Cholesky factorization and sparse inverse estimation, while the other sampling rooted spanning trees. Both algorithms operate in near-linear time and provide strong approximation guarantees. Extensive experiments on large real-world networks, including one with over 10 million nodes, demonstrate the efficiency and approximation quality of the proposed algorithms.

Method: Formalizes supervised learning across diverse data types, characterizes three families of interpretable models (feature importance, rule-based, instance-based), and defines ethical properties with evaluation metrics and verification procedures.

Result: Establishes theoretical foundations for developing AI systems that are accurate, interpretable, fair, privacy-preserving, and causally aware.

Conclusion: The framework enables generation of trustworthy AI systems by evaluating ethical measures during model generation and embedding key ethical properties within interpretable pipelines.

Abstract: Interpretable-by-design models are crucial for fostering trust, accountability, and safe adoption of automated decision-making models in real-world applications. In this paper we formalize the ground for the MIMOSA (Mining Interpretable Models explOiting Sophisticated Algorithms) framework, a comprehensive methodology for generating predictive models that balance interpretability with performance while embedding key ethical properties. We formally define here the supervised learning setting across diverse decision-making tasks and data types, including tabular data, time series, images, text, transactions, and trajectories. We characterize three major families of interpretable models: feature importance, rule, and instance based models. For each family, we analyze their interpretability dimensions, reasoning mechanisms, and complexity. Beyond interpretability, we formalize three critical ethical properties, namely causality, fairness, and privacy, providing formal definitions, evaluation metrics, and verification procedures for each. We then examine the inherent trade-offs between these properties and discuss how privacy requirements, fairness constraints, and causal reasoning can be embedded within interpretable pipelines. By evaluating ethical measures during model generation, this framework establishes the theoretical foundations for developing AI systems that are not only accurate and interpretable but also fair, privacy-preserving, and causally aware, i.e., trustworthy.

Method: Created EcomEval benchmark covering 6 categories and 37 tasks (8 multimodal), sourced from authentic customer queries and transaction logs. Used semi-automatic pipeline with large models drafting responses reviewed by 50+ expert annotators with e-commerce and multilingual expertise.

Result: Developed a comprehensive benchmark with defined difficulty levels for each question and task category, spanning 7 languages including 5 low-resource Southeast Asian languages, enabling fine-grained multilingual assessment.

Conclusion: EcomEval provides a more realistic and comprehensive evaluation framework for LLMs in e-commerce, addressing previous limitations and offering multilingual capabilities for assessing real-world shopping scenarios.

Abstract: Large Language Models (LLMs) excel on general-purpose NLP benchmarks, yet their capabilities in specialized domains remain underexplored. In e-commerce, existing evaluations-such as EcomInstruct, ChineseEcomQA, eCeLLM, and Shopping MMLU-suffer from limited task diversity (e.g., lacking product guidance and after-sales issues), limited task modalities (e.g., absence of multimodal data), synthetic or curated data, and a narrow focus on English and Chinese, leaving practitioners without reliable tools to assess models on complex, real-world shopping scenarios. We introduce EcomEval, a comprehensive multilingual and multimodal benchmark for evaluating LLMs in e-commerce. EcomEval covers six categories and 37 tasks (including 8 multimodal tasks), sourced primarily from authentic customer queries and transaction logs, reflecting the noisy and heterogeneous nature of real business interactions. To ensure both quality and scalability of reference answers, we adopt a semi-automatic pipeline in which large models draft candidate responses subsequently reviewed and modified by over 50 expert annotators with strong e-commerce and multilingual expertise. We define difficulty levels for each question and task category by averaging evaluation scores across models with different sizes and capabilities, enabling challenge-oriented and fine-grained assessment. EcomEval also spans seven languages-including five low-resource Southeast Asian languages-offering a multilingual perspective absent from prior work.

Method: Developed the Fluidity Index (FI) that evaluates response accuracy based on deviations in initial, current, and future environment states, distinguishing between closed-ended and open-ended benchmarks with focus on closed-loop open-ended real-world testing.

Result: The approach measures a model’s ability to understand, predict, and adjust to state changes in scaling environments, assessing context switching and continuity capabilities.

Conclusion: A truly super-intelligent model should demonstrate at least second-order adaptability, enabling self-sustained computation through digital replenishment to achieve optimal fluidity in dynamic environments.

Abstract: This paper introduces the Fluidity Index (FI) to quantify model adaptability in dynamic, scaling environments. The benchmark evaluates response accuracy based on deviations in initial, current, and future environment states, assessing context switching and continuity. We distinguish between closed-ended and open-ended benchmarks, prioritizing closed-loop open-ended real-world benchmarks to test adaptability. The approach measures a model’s ability to understand, predict, and adjust to state changes in scaling environments. A truly super-intelligent model should exhibit at least second-order adaptability, enabling self-sustained computation through digital replenishment for optimal fluidity.

Method: The method involves a fine-grained systematization of existing approaches using the BDI paradigm as a reference framework for analysis.

Result: The analysis illustrates the fast-evolving literature on rational agents enhanced by ML and identifies key research opportunities in this domain.

Conclusion: The paper concludes by highlighting open challenges for designing effective rational ML agents and provides guidance for future research directions.

Abstract: Thanks to the remarkable human-like capabilities of machine learning (ML) models in perceptual and cognitive tasks, frameworks integrating ML within rational agent architectures are gaining traction. Yet, the landscape remains fragmented and incoherent, often focusing on embedding ML into generic agent containers while overlooking the expressive power of rational architectures–such as Belief-Desire-Intention (BDI) agents. This paper presents a fine-grained systematisation of existing approaches, using the BDI paradigm as a reference. Our analysis illustrates the fast-evolving literature on rational agents enhanced by ML, and identifies key research opportunities and open challenges for designing effective rational ML agents.

Method: Proposes a topological data analysis (TDA)-based evaluation framework that captures the geometry of reasoning traces, enabling label-efficient automated assessment using topological features.

Result: Topological features yield substantially higher predictive power for assessing reasoning quality than standard graph metrics, showing that effective reasoning is better captured by higher-dimensional geometric structures rather than purely relational graphs.

Conclusion: A compact, stable set of topological features reliably indicates trace quality, offering a practical signal for future reinforcement learning algorithms.

Abstract: Evaluating the quality of reasoning traces from large language models remains understudied, labor-intensive, and unreliable: current practice relies on expert rubrics, manual annotation, and slow pairwise judgments. Automated efforts are dominated by graph-based proxies that quantify structural connectivity but do not clarify what constitutes high-quality reasoning; such abstractions can be overly simplistic for inherently complex processes. We introduce a topological data analysis (TDA)-based evaluation framework that captures the geometry of reasoning traces and enables label-efficient, automated assessment. In our empirical study, topological features yield substantially higher predictive power for assessing reasoning quality than standard graph metrics, suggesting that effective reasoning is better captured by higher-dimensional geometric structures rather than purely relational graphs. We further show that a compact, stable set of topological features reliably indicates trace quality, offering a practical signal for future reinforcement learning algorithms.

Method: Two-stage framework: 1) Chain-of-thought fine-tuning with customized plan-retrieval dataset to activate structured reasoning, 2) Plan-retrieval guided RL process with multi-reward design, Cartesian-inspired planning module for question decomposition, and logical expression for tool invocation.

Result: The framework enables globally consistent multi-step reasoning and coverage-aware retrieval scheduling, allowing effective combination of KG and web retrieval when needed.

Conclusion: Graph-RFT addresses key limitations in current KGQA approaches by enabling autonomous planning, adaptive retrieval scheduling, and coherent multi-step reasoning under incomplete knowledge conditions.

Abstract: Knowledge Graph Question Answering aims to answer natural language questions by reasoning over structured knowledge graphs. While large language models have advanced KGQA through their strong reasoning capabilities, existing methods continue to struggle to fully exploit both the rich knowledge encoded in KGs and the reasoning capabilities of LLMs, particularly in complex scenarios. They often assume complete KG coverage and lack mechanisms to judge when external information is needed, and their reasoning remains locally myopic, failing to maintain coherent multi-step planning, leading to reasoning failures even when relevant knowledge exists. We propose Graph-RFT, a novel two-stage reinforcement fine-tuning KGQA framework with a ‘plan-KGsearch-and-Websearch-during-think’ paradigm, that enables LLMs to perform autonomous planning and adaptive retrieval scheduling across KG and web sources under incomplete knowledge conditions. Graph-RFT introduces a chain-of-thought fine-tuning method with a customized plan-retrieval dataset activates structured reasoning and resolves the GRPO cold-start problem. It then introduces a novel plan-retrieval guided reinforcement learning process integrates explicit planning and retrieval actions with a multi-reward design, enabling coverage-aware retrieval scheduling. It employs a Cartesian-inspired planning module to decompose complex questions into ordered subquestions, and logical expression to guide tool invocation for globally consistent multi-step reasoning. This reasoning retrieval process is optimized with a multi-reward combining outcome and retrieval specific signals, enabling the model to learn when and how to combine KG and web retrieval effectively.

Method: Developed a coherence-aware AGI measure based on the integral of generalized means over a continuum of compensability exponents, spanning arithmetic, geometric, and harmonic regimes.

Result: Applied to GPT-4 and GPT-5 CHC domain scores, the coherence-adjusted AUC shows both systems remain far from general competence despite high arithmetic scores (GPT-5 at 24%).

Conclusion: The generalized mean integration provides a principled, interpretable, and stricter foundation for measuring genuine AGI progress by penalizing imbalance and capturing inter-domain dependency.

Abstract: Recent work by \citet{hendrycks2025agidefinition} formalized \textit{Artificial General Intelligence} (AGI) as the arithmetic mean of proficiencies across cognitive domains derived from the Cattell–Horn–Carroll (CHC) model of human cognition. While elegant, this definition assumes \textit{compensability} – that exceptional ability in some domains can offset failure in others. True general intelligence, however, should reflect \textit{coherent sufficiency}: balanced competence across all essential domains. We propose a coherence-aware measure of AGI based on the integral of generalized means over a continuum of compensability exponents. This formulation spans arithmetic, geometric, and harmonic regimes, and the resulting \textit{area under the curve} (AUC) quantifies robustness under varying compensability assumptions. Unlike the arithmetic mean, which rewards specialization, the AUC penalizes imbalance and captures inter-domain dependency. Applied to published CHC-based domain scores for GPT-4 and GPT-5, the coherence-adjusted AUC reveals that both systems remain far from general competence despite high arithmetic scores (e.g., GPT-5 at~24%). Integrating the generalized mean thus yields a principled, interpretable, and stricter foundation for measuring genuine progress toward AGI.

Method: Developed a comprehensive RDR framework pipeline that systematically analyzes research areas, identifies emerging trends, uncovers cross-domain opportunities, and provides starting points for new inquiry.

Result: Applied RDR to AI and robotics domains with focus on foundation models and robotics advancements, and extended analysis to other scientific areas. Appendix provides extensive results across analyzed topics.

Conclusion: The RDR pipeline helps researchers navigate complex research landscapes and provides insights for AI and broader scientific communities.

Abstract: With the rapid growth of research in AI and robotics now producing over 10,000 papers annually it has become increasingly difficult for researchers to stay up to date. Fast evolving trends, the rise of interdisciplinary work, and the need to explore domains beyond one’s expertise all contribute to this challenge. To address these issues, we propose a generalizable pipeline capable of systematically analyzing any research area: identifying emerging trends, uncovering cross domain opportunities, and offering concrete starting points for new inquiry. In this work, we present Real Deep Research (RDR) a comprehensive framework applied to the domains of AI and robotics, with a particular focus on foundation models and robotics advancements. We also briefly extend our analysis to other areas of science. The main paper details the construction of the RDR pipeline, while the appendix provides extensive results across each analyzed topic. We hope this work sheds light for researchers working in the field of AI and beyond.

Method: Three key innovations: (1) structured theory of mind representation linking percepts, beliefs, desires, and actions; (2) natural inter-agent communication; and (3) multi-component memory system. Tested in both instructive and collaborative settings within Minecraft.

Result: In instructive settings with GPT-4, MindForge agents achieved 3× more tech-tree milestones and collected 2.3× more unique items than Voyager baseline. In collaborative settings, performance improved with more communication rounds, showing sophisticated behaviors like knowledge transfer and problem solving.

Conclusion: MindForge enables sophisticated cultural learning behaviors including expert-novice knowledge transfer, collaborative problem solving, and adaptation to out-of-distribution tasks through accumulated cultural experiences.

Abstract: Embodied agents powered by large language models (LLMs), such as Voyager, promise open-ended competence in worlds such as Minecraft. However, when powered by open-weight LLMs they still falter on elementary tasks after domain-specific fine-tuning. We propose MindForge, a generative-agent framework for cultural lifelong learning through explicit perspective taking. We introduce three key innovations: (1) a structured theory of mind representation linking percepts, beliefs, desires, and actions; (2) natural inter-agent communication; and (3) a multi-component memory system. Following the cultural learning framework, we test MindForge in both instructive and collaborative settings within Minecraft. In an instructive setting with GPT-4, MindForge agents powered by open-weight LLMs significantly outperform their Voyager counterparts in basic tasks yielding $3\times$ more tech-tree milestones and collecting $2.3\times$ more unique items than the Voyager baseline. Furthermore, in fully \textit{collaborative} settings, we find that the performance of two underachieving agents improves with more communication rounds, echoing the Condorcet Jury Theorem. MindForge agents demonstrate sophisticated behaviors, including expert-novice knowledge transfer, collaborative problem solving, and adaptation to out-of-distribution tasks through accumulated cultural experiences.

Method: Proposed MIR-Bench benchmark that asks LLMs to predict outputs via input-output examples from underlying functions with diverse data formats, enabling evaluation of many-shot in-context reasoning capabilities.

Result: The study revealed important findings including scaling effects, robustness, inductive vs. transductive reasoning, effectiveness of RAG, coding for inductive reasoning, and cross-domain generalizability in many-shot learning scenarios.

Conclusion: MIR-Bench fills a critical gap in evaluating LLMs’ pattern recognition abilities in many-shot settings and provides insights into how LLMs perform complex reasoning when aggregating information from hundreds to thousands of examples in long contexts.

Abstract: The ability to recognize patterns from examples and apply them to new ones is a primal ability for general intelligence, and is widely studied by psychology and AI researchers. Many benchmarks have been proposed to measure such ability for Large Language Models (LLMs); however, they focus on few-shot (usually <10) setting and lack evaluation for aggregating many pieces of information from long contexts. On the other hand, the ever-growing context length of LLMs have brought forth the novel paradigm of many-shot In-Context Learning (ICL), which addresses new tasks with hundreds to thousands of examples without expensive and inefficient fine-tuning. However, many-shot evaluations often focus on classification, and popular long-context LLM tasks such as Needle-In-A-Haystack (NIAH) seldom require complicated intelligence for integrating many pieces of information. To fix the issues from both worlds, we propose MIR-Bench, the first many-shot in-context reasoning benchmark for pattern recognition that asks LLM to predict output via input-output examples from underlying functions with diverse data format. Based on MIR-Bench, we study many novel problems for many-shot in-context reasoning, and acquired many insightful findings including scaling effect, robustness, inductive vs. transductive reasoning, retrieval Augmented Generation (RAG), coding for inductive reasoning, cross-domain generalizability, etc.

Method: Systematic probing of RLVR-trained LLMs across various model families, RL algorithms, and reasoning benchmarks using pass@k at large k values, with coverage and perplexity analyses.

Result: RLVR-trained models outperform base models at small k (e.g., k=1) but base models achieve higher pass@k scores when k is large. Six popular RLVR algorithms perform similarly and remain far from optimal in leveraging base model potential.

Conclusion: Current RLVR methods have not realized RL’s potential to elicit truly novel reasoning abilities in LLMs, highlighting the need for improved RL paradigms like continual scaling and multi-turn agent-environment interaction.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has recently demonstrated notable success in enhancing the reasoning performance of large language models (LLMs), particularly on mathematics and programming tasks. Similar to how traditional RL helps agents explore and learn new strategies, RLVR is believed to enable LLMs to continuously self-improve, thus acquiring novel reasoning abilities beyond those of the corresponding base models. In this study we critically examine the current state of RLVR by systematically probing the reasoning capability boundaries of RLVR-trained LLMs across various model families, RL algorithms, and math, coding, and visual reasoning benchmarks, using pass@k at large k values as the evaluation metric. Surprisingly, we find that the current training setup does not elicit fundamentally new reasoning patterns. While RLVR-trained models outperform their base models at small k (e.g., k = 1), the base models achieve a higher pass@k score when k is large. Coverage and perplexity analyses show that the observed reasoning abilities originate from and are bounded by the base model. Treating the base model as an upper bound, our quantitative analysis shows that six popular RLVR algorithms perform similarly and remain far from optimal in leveraging the potential of the base model. By contrast, we find that distillation can introduce new reasoning patterns from the teacher and genuinely expand the model’s reasoning capabilities. Overall, our findings suggest that current RLVR methods have not yet realized the potential of RL to elicit truly novel reasoning abilities in LLMs. This highlights the need for improved RL paradigms, such as continual scaling and multi-turn agent-environment interaction, to unlock this potential.

Method: Proposed PURE (Process sUpervised Reinforcement lEarning) with min-form credit assignment that formulates the value function as the minimum of future rewards instead of cumulative gamma-decayed rewards, limiting value function range and distributing advantages more reasonably.

Result: PRM-based approaches with min-form credit assignment achieved comparable reasoning performance to verifiable reward methods within only 30% steps, while sum-form assignment collapsed training. With 10% verifiable rewards, achieved 82.5% accuracy on AMC23 and 53.3% average accuracy across 5 benchmarks using Qwen2.5-Math-7B.

Conclusion: Min-form credit assignment effectively alleviates reward hacking in PRM-based reinforcement fine-tuning, enabling successful training and competitive performance with fewer steps and minimal verifiable reward supplementation.

Abstract: Process reward models (PRMs) have proven effective for test-time scaling of Large Language Models (LLMs) on challenging reasoning tasks. However, reward hacking issues with PRMs limit their successful application in reinforcement fine-tuning. In this paper, we identify the main cause of PRM-induced reward hacking: the canonical summation-form credit assignment in reinforcement learning (RL), which defines the value as cumulative gamma-decayed future rewards, easily induces LLMs to hack steps with high rewards. To address this, we propose PURE: Process sUpervised Reinforcement lEarning. The key innovation of PURE is a min-form credit assignment that formulates the value function as the minimum of future rewards. This method significantly alleviates reward hacking by limiting the value function range and distributing advantages more reasonably. Through extensive experiments on 3 base models, we show that PRM-based approaches enabling min-form credit assignment achieve comparable reasoning performance to verifiable reward-based methods within only 30% steps. In contrast, the canonical sum-form credit assignment collapses training even at the beginning! Additionally, when we supplement PRM-based fine-tuning with just 10% verifiable rewards, we further alleviate reward hacking and produce the best fine-tuned model based on Qwen2.5-Math-7B in our experiments, achieving 82.5% accuracy on AMC23 and 53.3% average accuracy across 5 benchmarks. Moreover, we summarize the observed reward hacking cases and analyze the causes of training collapse. We release our code and model weights at https://github.com/CJReinforce/PURE.

Method: Proposed a Hierarchical Reinforcement Learning (HRL) approach with two-tiered decision-making: high-level for model selection and low-level for determining weights of selected base models to handle non-stationary building data streams.

Result: The approach was evaluated through offline experiments and an on-site case study, demonstrating effectiveness in providing accurate predictions while reducing modeling efforts.

Conclusion: The model ensemble perspective with HRL enables efficient reuse of existing models for target building environments, achieving accurate thermodynamics predictions with reduced data and expert dependency.

Abstract: The building thermodynamics model, which predicts real-time indoor temperature changes under potential HVAC (Heating, Ventilation, and Air Conditioning) control operations, is crucial for optimizing HVAC control in buildings. While pioneering studies have attempted to develop such models for various building environments, these models often require extensive data collection periods and rely heavily on expert knowledge, making the modeling process inefficient and limiting the reusability of the models. This paper explores a model ensemble perspective that utilizes existing developed models as base models to serve a target building environment, thereby providing accurate predictions while reducing the associated efforts. Given that building data streams are non-stationary and the number of base models may increase, we propose a Hierarchical Reinforcement Learning (HRL) approach to dynamically select and weight the base models. Our approach employs a two-tiered decision-making process: the high-level focuses on model selection, while the low-level determines the weights of the selected models. We thoroughly evaluate the proposed approach through offline experiments and an on-site case study, and the experimental results demonstrate the effectiveness of our method.

Method: Conducted a scoping review using standard methodology, extracting 57 articles from 1,943 studies published from 2019-2024. Addressed 3 research questions about privacy risks, preservation methods, and characteristics of privacy-preserving explanations.

Result: Categorized privacy risks and preservation methods in XAI, proposed characteristics of privacy-preserving explanations, identified challenges in balancing privacy with other system requirements, and provided recommendations for achieving privacy-preserving XAI.

Conclusion: The review sheds light on the complex relationship between privacy and explainability, both fundamental principles of Trustworthy AI, providing guidance for researchers and practitioners to develop privacy-compliant XAI systems.

Abstract: Explainable Artificial Intelligence (XAI) has emerged as a pillar of Trustworthy AI and aims to bring transparency in complex models that are opaque by nature. Despite the benefits of incorporating explanations in models, an urgent need is found in addressing the privacy concerns of providing this additional information to end users. In this article, we conduct a scoping review of existing literature to elicit details on the conflict between privacy and explainability. Using the standard methodology for scoping review, we extracted 57 articles from 1,943 studies published from January 2019 to December 2024. The review addresses 3 research questions to present readers with more understanding of the topic: (1) what are the privacy risks of releasing explanations in AI systems? (2) what current methods have researchers employed to achieve privacy preservation in XAI systems? (3) what constitutes a privacy preserving explanation? Based on the knowledge synthesized from the selected studies, we categorize the privacy risks and preservation methods in XAI and propose the characteristics of privacy preserving explanations to aid researchers and practitioners in understanding the requirements of XAI that is privacy compliant. Lastly, we identify the challenges in balancing privacy with other system desiderata and provide recommendations for achieving privacy preserving XAI. We expect that this review will shed light on the complex relationship of privacy and explainability, both being the fundamental principles of Trustworthy AI.

Method: Fine-tune LRMs to emit a short Safety Primer (8 tokens) at the start of reasoning for harmful prompts, while leaving the rest of reasoning unsupervised. Also includes a zero-shot variant requiring no fine-tuning.

Result: Reduces harmful responses by up to 90.0%, blocks 83.3% of jailbreak attempts in DeepSeek-R1-Distill-Llama-8B model, and requires 295.9x less compute than Direct Refusal and 314.1x less than SafeChain.

Conclusion: SAFEPATH effectively reduces harmful outputs while maintaining reasoning performance, with significant computational efficiency improvements over existing methods.

Abstract: Large Reasoning Models (LRMs) have become powerful tools for complex problem solving, but their structured reasoning pathways can lead to unsafe outputs when exposed to harmful prompts. Existing safety alignment methods reduce harmful outputs but can degrade reasoning depth, leading to significant trade-offs in complex, multi-step tasks, and remain vulnerable to sophisticated jailbreak attacks. To address this, we introduce SAFEPATH, a lightweight alignment method that fine-tunes LRMs to emit a short, 8-token Safety Primer at the start of their reasoning, in response to harmful prompts, while leaving the rest of the reasoning process unsupervised. Empirical results across multiple benchmarks indicate that SAFEPATH effectively reduces harmful outputs while maintaining reasoning performance. Specifically, SAFEPATH reduces harmful responses by up to 90.0% and blocks 83.3% of jailbreak attempts in the DeepSeek-R1-Distill-Llama-8B model, while requiring 295.9x less compute than Direct Refusal and 314.1x less than SafeChain. We further introduce a zero-shot variant that requires no fine-tuning. In addition, we provide a comprehensive analysis of how existing methods in LLMs generalize, or fail, when applied to reasoning-centric models, revealing critical gaps and new directions for safer AI.

Method: Empirical study across models and benchmarks, analysis using a probabilistic model to understand the non-monotonic performance trend, and introduction of parallel thinking approach inspired by Best-of-N sampling that generates multiple independent reasoning paths with majority vote selection.

Result: Extended thinking shows a consistent pattern: initial performance improvements followed by decline due to overthinking. Additional thinking increases output variance, creating an illusion of improved reasoning while undermining precision. Parallel thinking achieves up to 20% higher accuracy compared to extended thinking.

Conclusion: Test-time scaling through extended thinking is not effective due to overthinking artifacts. Parallel thinking provides a simple yet effective alternative that better utilizes the inference thinking budget by generating multiple independent reasoning paths and selecting the most consistent response.

Abstract: Recent trends in test-time scaling for reasoning models (e.g., OpenAI o1, DeepSeek R1) have led to a popular belief that extending thinking traces using prompts like “Wait” or “Let me rethink” can improve performance. This raises a natural question: Does thinking more at test-time truly lead to better reasoning? To answer this question, we perform a detailed empirical study across models and benchmarks, which reveals a consistent pattern of initial performance improvements from additional thinking followed by a decline, due to “overthinking”. To understand this non-monotonic trend, we consider a simple probabilistic model, which reveals that additional thinking increases output variance-creating an illusion of improved reasoning while ultimately undermining precision. Thus, observed gains from “more thinking” are not true indicators of improved reasoning, but artifacts stemming from the connection between model uncertainty and evaluation metric. This suggests that test-time scaling through extended thinking is not an effective way to utilize the inference thinking budget. Recognizing these limitations, we introduce an alternative test-time scaling approach, parallel thinking, inspired by Best-of-N sampling. Our method generates multiple independent reasoning paths within the same inference budget and selects the most consistent response via majority vote, achieving up to 20% higher accuracy compared to extended thinking. This provides a simple yet effective mechanism for test-time scaling of reasoning models.

Method: Integrates an informal reasoning LLM with Lean proof assistant, coordinates formal prover models, uses Lean feedback, and generates auxiliary lemmas including subgoals, special cases, and useful facts from assumptions.

Result: Achieves 88.1% success rate on MiniF2F benchmark, establishing new state-of-the-art among methods using small language models with lower sample budget than previous approaches.

Conclusion: Prover Agent demonstrates effective integration of LLMs with formal proof assistants, with generated auxiliary lemmas playing crucial role in solving challenging problems, providing a promising approach for automated theorem proving.

Abstract: We present Prover Agent, a novel AI agent for automated theorem proving that integrates large language models (LLMs) with a formal proof assistant, Lean. Prover Agent coordinates an informal reasoning LLM, a formal prover model, and feedback from Lean while also generating auxiliary lemmas. These auxiliary lemmas are not limited to subgoals in the formal proof but can also include special cases or potentially useful facts derived from the assumptions, which help in discovering a viable proof strategy. It achieves an 88.1% success rate on the MiniF2F benchmark, establishing a new state-of-the-art among methods using small language models (SLMs) with a much lower sample budget than previous approaches. We also present theoretical analyses and case studies that illustrate how these generated lemmas contribute to solving challenging problems. Our code is publicly available at: https://github.com/kAIto47802/Prover-Agent.

Method: Conducted a scoping literature review of 100 unclassified studies on AI in wargames and constructed a novel ontology of open-endedness based on player creativity and novelty for observers.

Result: Developed practical recommendations and critical safety considerations for deploying AI in open-ended wargames across common domains.

Conclusion: Presented the community with a set of high-impact open research challenges for future work on AI in wargaming.

Abstract: Wargames are simulations of conflicts in which participants’ decisions influence future events. While casual wargaming can be used for entertainment or socialization, serious wargaming is used by experts to explore strategic implications of decision-making and experiential learning. In this paper, we take the position that Artificial Intelligence (AI) systems, such as Language Models (LMs), are rapidly approaching human-expert capability for strategic planning – and will one day surpass it. Military organizations have begun using LMs to provide insights into the consequences of real-world decisions during open-ended wargames which use natural language to convey actions and outcomes. We argue the ability for AI systems to influence large-scale decisions motivates additional research into the safety, interpretability, and explainability of AI in open-ended wargames. To demonstrate, we conduct a scoping literature review with a curated selection of 100 unclassified studies on AI in wargames, and construct a novel ontology of open-endedness using the creativity afforded to players, adjudicators, and the novelty provided to observers. Drawing from this body of work, we distill a set of practical recommendations and critical safety considerations for deploying AI in open-ended wargames across common domains. We conclude by presenting the community with a set of high-impact open research challenges for future work.

Method: Combines Mesa-based spatial modelling with CLIMADA climate impact assessment, introducing adaptive learning behaviors that allow firms to evolve strategies for budget allocation, pricing, wages, and risk adaptation through fitness-based selection and mutation.

Result: Evolutionary adaptation enables firms to converge with baseline production levels after decades of disruption due to climate stress. Systemic risks emerge where even non-exposed agents face impacts through supply chain disruptions, with end-of-century average price of goods 5.6% higher under RCP8.5.

Conclusion: This open-source framework provides financial institutions and companies with tools to quantify both direct and cascading climate risks while evaluating cost-effective adaptation strategies.

Abstract: Climate risk assessment requires modelling complex interactions between spatially heterogeneous hazards and adaptive economic systems. We present a novel geospatial agent-based model that integrates climate hazard data with evolutionary learning for economic agents. Our framework combines Mesa-based spatial modelling with CLIMADA climate impact assessment, introducing adaptive learning behaviours that allow firms to evolve strategies for budget allocation, pricing, wages, and risk adaptation through fitness-based selection and mutation. We demonstrate the framework using riverine flood projections under RCP8.5 until 2100, showing that evolutionary adaptation enables firms to converge with baseline (no hazard) production levels after decades of disruption due to climate stress. Our results reveal systemic risks where even agents that are not directly exposed to floods face impacts through supply chain disruptions, with the end-of-century average price of goods 5.6% higher under RCP8.5 compared to the baseline in our illustrative economic network. This open-source framework provides financial institutions and companies with tools to quantify both direct and cascading climate risks while evaluating cost-effective adaptation strategies.

Method: Proposed EA4LLM evolutionary algorithm for optimizing LLMs, empirically testing full-parameter optimization from pretraining across model sizes from 0.5B to 32B parameters.

Result: Successfully demonstrated evolutionary algorithm optimization of LLMs, providing key insights into how EAs can effectively optimize neural networks.

Conclusion: Evolutionary algorithms present a viable alternative to gradient-based optimization, potentially reducing computational costs and enabling broader participation in deep learning research.

Abstract: In recent years, large language models (LLMs) have made remarkable progress, with model optimization primarily relying on gradient-based optimizers such as Adam. However, these gradient-based methods impose stringent hardware requirements, demanding high-concurrency, high-memory GPUs. Moreover, they require all neural network operations to be differentiable, thereby excluding many promising non-differentiable architectures from practical use. To address these limitations, we propose EA4LLM, an evolutionary algorithm for optimizing LLMs, and, for the first time, empirically verify full-parameter optimization from the pretraining stage across model sizes ranging from 0.5B to 32B. We conduct extensive experiments and provide key insights into how evolutionary algorithms can effectively optimize neural networks. Our work challenges the prevailing assumption that gradient-based optimization is the only viable approach for training neural networks. It also holds significant potential to reduce the computational cost of training large language models, thereby enabling groups with limited computational resources to participate in deep learning research.

Method: Developed BuildArena benchmark with four components: customizable framework, extendable task design spanning static/dynamic mechanics, 3D Spatial Geometric Computation Library, and baseline LLM agentic workflow for comprehensive evaluation.

Result: Comprehensively evaluated eight frontier LLMs on their capabilities for language-driven and physics-grounded construction automation using the BuildArena benchmark.

Conclusion: BuildArena provides the first standardized framework for assessing LLMs’ engineering construction automation capabilities, enabling systematic comparison and analysis of model performance in this domain.

Abstract: Engineering construction automation aims to transform natural language specifications into physically viable structures, requiring complex integrated reasoning under strict physical constraints. While modern LLMs possess broad knowledge and strong reasoning capabilities that make them promising candidates for this domain, their construction competencies remain largely unevaluated. To address this gap, we introduce BuildArena, the first physics-aligned interactive benchmark designed for language-driven engineering construction. It contributes to the community in four aspects: (1) a highly customizable benchmarking framework for in-depth comparison and analysis of LLMs; (2) an extendable task design strategy spanning static and dynamic mechanics across multiple difficulty tiers; (3) a 3D Spatial Geometric Computation Library for supporting construction based on language instructions; (4) a baseline LLM agentic workflow that effectively evaluates diverse model capabilities. On eight frontier LLMs, BuildArena comprehensively evaluates their capabilities for language-driven and physics-grounded construction automation. The project page is at https://build-arena.github.io/.

Method: MERCI uses a lightweight Coin Flipping Network (CFN) to estimate pseudo counts and epistemic uncertainty over reasoning trajectories, converting them into intrinsic rewards that value novelty while preserving task reward signals. It’s integrated with RL frameworks like GRPO.

Result: Experiments on complex reasoning benchmarks show MERCI encourages richer and more varied chains of thought, significantly improves performance over strong baselines, and helps policies escape local routines to discover better solutions.

Conclusion: Targeted intrinsic motivation through count-based exploration makes exploration reliable for language model reasoning, addressing fundamental limitations in current RL approaches for LLMs.

Abstract: Reinforcement Learning (RL) has become a compelling way to strengthen the multi step reasoning ability of Large Language Models (LLMs). However, prevalent RL paradigms still lean on sparse outcome-based rewards and limited exploration, which often drives LLMs toward repetitive and suboptimal reasoning patterns. In this paper, we study the central question of how to design exploration for LLM reasoning and introduce MERCI (Motivating Exploration in LLM Reasoning with Count-based Intrinsic Rewards), a novel RL algorithm that augments policy optimization with a principled intrinsic reward. Building on the idea of count-based exploration, MERCI leverages a lightweight Coin Flipping Network (CFN) to estimate the pseudo count and further epistemic uncertainty over reasoning trajectories, and converts them into an intrinsic reward that values novelty while preserving the learning signal from task rewards. We integrate MERCI into some advanced RL frameworks like Group Relative Policy Optimization (GRPO). Experiments on complex reasoning benchmarks demonstrate that MERCI encourages richer and more varied chains of thought, significantly improves performance over strong baselines, and helps the policy escape local routines to discover better solutions. It indicates that our targeted intrinsic motivation can make exploration reliable for language model reasoning.

Method: Tested eight frontier models on producing valid scientific test items and revising after self-critique. Measured first-pass performance and gains from reflection, analyzing whether models detect and correct constraint violations.

Result: First-pass performance was poor (mean ≈1 valid item out of 4), reflection yielded only modest gains (also ≈1). Models frequently repeated the same constraint violations, and performance deteriorated with increased open-endedness. Models marketed for ‘reasoning’ showed no advantage.

Conclusion: Current LLM reflection lacks the active, goal-driven monitoring that helps humans respect constraints. Reliable performance requires external structure that enforces constraints until such mechanisms are instantiated in models themselves.

Abstract: Humans do not just find mistakes after the fact – we often catch them mid-stream because ‘reflection’ is tied to the goal and its constraints. Today’s large language models produce reasoning tokens and ‘reflective’ text, but is it functionally equivalent with human reflective reasoning? Prior work on closed-ended tasks – with clear, external ‘correctness’ signals – can make ‘reflection’ look effective while masking limits in self-correction. We therefore test eight frontier models on a simple, real-world task that is open-ended yet rule-constrained, with auditable success criteria: to produce valid scientific test items, then revise after considering their own critique. First-pass performance is poor (often zero valid items out of 4 required; mean $\approx$ 1), and reflection yields only modest gains (also $\approx$ 1). Crucially, the second attempt frequently repeats the same violation of constraint, indicating ‘corrective gains’ arise largely from chance production of a valid item rather than error detection and principled, constraint-sensitive repair. Performance before and after reflection deteriorates as open-endedness increases, and models marketed for ‘reasoning’ show no advantage. Our results suggest that current LLM ‘reflection’ lacks functional evidence of the active, goal-driven monitoring that helps humans respect constraints even on a first pass. Until such mechanisms are instantiated in the model itself, reliable performance requires external structure that enforces constraints. Our code is available at: https://github.com/cruiseresearchgroup/LLM_ReflectionTest

Method: The method identifies core reasoning circuits (sparse, specialized attention heads) and quantifies reasoning complexity by measuring data’s influence on these circuits to select high-quality training subsets.

Result: Fine-tuning Qwen2.5-Math-7B on just 10% of CircuitSeer-selected data achieved 1.4-point gain in average Pass@1 over full dataset training, demonstrating superior efficiency across 4 models and 9 datasets.

Conclusion: CircuitSeer provides an effective and efficient data selection approach by leveraging internal model mechanisms rather than external heuristics, enabling better performance with significantly less training data.

Abstract: Large language models (LLMs) have demonstrated impressive reasoning capabilities, but scaling their performance often relies on massive reasoning datasets that are computationally expensive to train on. Existing data selection methods aim to curate smaller, high-quality subsets but often rely on costly external models or opaque heuristics. In this work, we shift the focus from external heuristics to the model’s internal mechanisms. We find that complex reasoning tasks consistently activate a sparse, specialized subset of attention heads, forming core reasoning circuits. Building on this insight, we propose CircuitSeer, a novel data selection method that quantifies the reasoning complexity of data by measuring its influence on these crucial circuits. Extensive experiments on 4 models and 9 datasets demonstrate CircuitSeer’s superiority. Notably, fine-tuning Qwen2.5-Math-7B on just 10% of data selected by our method achieves a 1.4-point gain in average Pass@1 over training on the full dataset, highlighting its efficiency and effectiveness.

Method: Integrates Bayesian Experimental Design with LLMs to select tests that maximize diagnostic uncertainty reduction. LLMs act as flexible simulators for patient state distributions without requiring structured training data.

Result: ACTMED optimizes test selection to improve diagnostic accuracy, interpretability, and resource use on real-world datasets.

Conclusion: Represents progress toward transparent, adaptive diagnostic systems that generalize across settings with reduced reliance on domain-specific data while keeping clinicians in the loop.

Abstract: There is growing interest in using machine learning (ML) to support clinical diag- nosis, but most approaches rely on static, fully observed datasets and fail to reflect the sequential, resource-aware reasoning clinicians use in practice. Diagnosis remains complex and error prone, especially in high-pressure or resource-limited settings, underscoring the need for frameworks that help clinicians make timely and cost-effective decisions. We propose ACTMED (Adaptive Clinical Test selection via Model-based Experimental Design), a diagnostic framework that integrates Bayesian Experimental Design (BED) with large language models (LLMs) to better emulate real-world diagnostic reasoning. At each step, ACTMED selects the test expected to yield the greatest reduction in diagnostic uncertainty for a given patient. LLMs act as flexible simulators, generating plausible patient state distributions and supporting belief updates without requiring structured, task-specific training data. Clinicians can remain in the loop; reviewing test suggestions, interpreting intermediate outputs, and applying clinical judgment throughout. We evaluate ACTMED on real-world datasets and show it can optimize test selection to improve diagnostic accuracy, interpretability, and resource use. This represents a step to- ward transparent, adaptive, and clinician-aligned diagnostic systems that generalize across settings with reduced reliance on domain-specific data.

Method: DAIL uses two key components: distributional policy (value distribution estimation) and semantic alignment (capturing correspondence between trajectories and linguistic instructions).

Result: Extensive experiments on structured and visual observation benchmarks show DAIL effectively resolves instruction ambiguities and achieves superior performance compared to baseline methods.

Conclusion: DAIL successfully addresses the challenge of instruction ambiguity in language-conditioned tasks through its novel distributional alignment approach.

Abstract: Comprehending natural language and following human instructions are critical capabilities for intelligent agents. However, the flexibility of linguistic instructions induces substantial ambiguity across language-conditioned tasks, severely degrading algorithmic performance. To address these limitations, we present a novel method named DAIL (Distributional Aligned Learning), featuring two key components: distributional policy and semantic alignment. Specifically, we provide theoretical results that the value distribution estimation mechanism enhances task differentiability. Meanwhile, the semantic alignment module captures the correspondence between trajectories and linguistic instructions. Extensive experimental results on both structured and visual observation benchmarks demonstrate that DAIL effectively resolves instruction ambiguities, achieving superior performance to baseline methods. Our implementation is available at https://github.com/RunpengXie/Distributional-Aligned-Learning.

Method: WorldTest protocol with reward-free exploration phase followed by scored test phase in different but related environments, using AutumnBench suite with 43 interactive grid-world environments and 129 tasks across masked-frame prediction, planning, and causal dynamics prediction.

Result: Humans outperformed frontier models on AutumnBench, and scaling compute only improved performance in some environments but not others, showing significant headroom in world-model learning.

Conclusion: WorldTest provides a novel template for evaluating what agents learn about environment dynamics, separating exploration from testing and enabling comparison across different model representations.

Abstract: Model-learning agents should gather information to learn world models that support many downstream tasks and inferences, such as predicting unobserved states, estimating near- and far-term consequences of actions, planning action sequences, and detecting changes in dynamics. Current methods for learning and evaluating world models diverge from this goal: training and evaluation are anchored to next-frame prediction, and success is scored by reward maximization in the same environment. We propose WorldTest, a protocol to evaluate model-learning agents that separates reward-free interaction from a scored test phase in a different but related environment. WorldTest is open-ended$\unicode{x2014}$models should support many different tasks unknown ahead of time$\unicode{x2014}$and agnostic to model representation, allowing comparison across approaches. We instantiated WorldTest with AutumnBench, a suite of 43 interactive grid-world environments and 129 tasks across three families: masked-frame prediction, planning, and predicting changes to the causal dynamics. We compared 517 human participants and three frontier models on AutumnBench. We found that humans outperform the models, and scaling compute improves performance only in some environments but not others. WorldTest provides a novel template$\unicode{x2014}$reward-free exploration, derived tests, and behavior-based scoring$\unicode{x2014}$to evaluate what agents learn about environment dynamics, and AutumnBench exposes significant headroom in world-model learning.

Method: Propose Vox-Evaluator to identify temporal boundaries of erroneous segments and provide quality assessment. Use it to automatically detect acoustic errors, mask erroneous segments, and regenerate speech conditioning on correct portions. Also use fine-grained information for preference alignment.

Result: Experimental results demonstrate effectiveness in enhancing stability and fidelity through speech correction mechanism and preference optimization. Created a synthesized text-speech dataset with fine-grained pronunciation/quality annotations.

Conclusion: Vox-Evaluator successfully addresses stability and fidelity challenges in zero-shot TTS systems by enabling targeted correction of erroneous segments and preference alignment.

Abstract: Recent advances in zero-shot text-to-speech (TTS), driven by language models, diffusion models and masked generation, have achieved impressive naturalness in speech synthesis. Nevertheless, stability and fidelity remain key challenges, manifesting as mispronunciations, audible noise, and quality degradation. To address these issues, we introduce Vox-Evaluator, a multi-level evaluator designed to guide the correction of erroneous speech segments and preference alignment for TTS systems. It is capable of identifying the temporal boundaries of erroneous segments and providing a holistic quality assessment of the generated speech. Specifically, to refine erroneous segments and enhance the robustness of the zero-shot TTS model, we propose to automatically identify acoustic errors with the evaluator, mask the erroneous segments, and finally regenerate speech conditioning on the correct portions. In addition, the fine-gained information obtained from Vox-Evaluator can guide the preference alignment for TTS model, thereby reducing the bad cases in speech synthesis. Due to the lack of suitable training datasets for the Vox-Evaluator, we also constructed a synthesized text-speech dataset annotated with fine-grained pronunciation errors or audio quality issues. The experimental results demonstrate the effectiveness of the proposed Vox-Evaluator in enhancing the stability and fidelity of TTS systems through the speech correction mechanism and preference optimization. The demos are shown.

Method: Proposes UniSE - a unified decoder-only LM framework that takes input speech features as conditions and generates discrete tokens of target speech using autoregressive modeling, enabling compatibility between distinct learning patterns of multiple tasks.

Result: Experiments on several benchmarks show UniSE achieves competitive performance compared to discriminative and generative baselines.

Conclusion: The work demonstrates the capacity of language models in unifying speech enhancement tasks, showing promising results across multiple SE sub-tasks.

Abstract: The development of neural audio codecs (NACs) has largely promoted applications of language models (LMs) to speech processing and understanding. However, there lacks the verification on the effectiveness of autoregressive (AR) LMbased models in unifying different sub-tasks of speech enhancement (SE). In this work, we propose UniSE, a unified decoder-only LM-based framework to handle different SE tasks including speech restoration, target speaker extraction and speech separation. It takes input speech features as conditions and generates discrete tokens of the target speech using AR modeling, which facilitates a compatibility between distinct learning patterns of multiple tasks. Experiments on several benchmarks indicate the proposed UniSE can achieve competitive performance compared to discriminative and generative baselines, showing the capacity of LMs in unifying SE tasks. The demo page is available here: https://github.com/hyyan2k/UniSE.

Method: Leverages reciprocity property to create physically valid samples by exchanging emitter and listener poses, and uses self-supervised learning to address challenges with emitter/listener gain patterns.

Result: Versa substantially improves acoustic field learning performance on both simulated and real-world datasets across different metrics, and perceptual user studies show improved immersive spatial sound experience.

Conclusion: The proposed method effectively addresses the resounding task and enhances acoustic field learning through physics-inspired reciprocity and self-supervised approaches.

Abstract: Achieving immersive auditory experiences in virtual environments requires flexible sound modeling that supports dynamic source positions. In this paper, we introduce a task called resounding, which aims to estimate room impulse responses at arbitrary emitter location from a sparse set of measured emitter positions, analogous to the relighting problem in vision. We leverage the reciprocity property and introduce Versa, a physics-inspired approach to facilitating acoustic field learning. Our method creates physically valid samples with dense virtual emitter positions by exchanging emitter and listener poses. We also identify challenges in deploying reciprocity due to emitter/listener gain patterns and propose a self-supervised learning approach to address them. Results show that Versa substantially improve the performance of acoustic field learning on both simulated and real-world datasets across different metrics. Perceptual user studies show that Versa can greatly improve the immersive spatial sound experience. Code, dataset and demo videos are available on the project website: https://waves.seas.upenn.edu/projects/versa.

Method: Proposes SimWhisper-Codec which leverages a frozen, simplified Whisper encoder for acoustic reconstruction without external supervision, based on empirical discovery that architectural simplification unlocks Whisper’s acoustic modeling potential.

Result: Superior performance in both semantic preservation and acoustic quality compared to semantically-supervised codecs like Mimi Codec and SpeechTokenizer at similar bitrates.

Conclusion: The semantic-first approach using simplified Whisper encoder is effective for balancing semantic and acoustic preservation in speech codecs.

Abstract: Speech codecs serve as bridges between continuous speech signals and large language models, yet face an inherent conflict between acoustic fidelity and semantic preservation. To mitigate this conflict, prevailing methods augment acoustic codecs with complex semantic supervision. We explore the opposite direction: a semantic-first approach that starts from a semantically-capable model and adapts it for high-fidelity acoustic reconstruction. Through empirical analysis, we discover that targeted architectural simplification can unlock the acoustic modeling potential of Whisper, a text-aligned Automatic Speech Recognition (ASR) model. Based on this finding, we propose SimWhisper-Codec, a novel codec that balances the semantic and acoustic preservation by leveraging a frozen, simplified Whisper encoder without requiring external supervision. Experimental results demonstrate that SimWhisper-Codec achieves superior performance in both semantic preservation and acoustic quality compared to semantically-supervised codecs such as Mimi Codec and SpeechTokenizer at similar bitrates, validating the effectiveness of our semantic-first approach. Code is available at https://github.com/ZhangXinWhut/SimWhisper-Codec.

Method: Three key approaches: 1) Simulation-based robustness with random F0 perturbations and artifact simulations, 2) Enriched speaker representation using DNSMOS-filtered vocals and singing corpora, 3) Neural Source-Filter model for harmonic and noise component representation.

Result: Achieves state-of-the-art performance on multiple SVC benchmarks under both clean and noisy conditions.

Conclusion: R2-SVC effectively bridges the gap between clean training data and noisy real-world scenarios, providing robust and expressive singing voice conversion.

Abstract: In real-world singing voice conversion (SVC) applications, environmental noise and the demand for expressive output pose significant challenges. Conventional methods, however, are typically designed without accounting for real deployment scenarios, as both training and inference usually rely on clean data. This mismatch hinders practical use, given the inevitable presence of diverse noise sources and artifacts from music separation. To tackle these issues, we propose R2-SVC, a robust and expressive SVC framework. First, we introduce simulation-based robustness enhancement through random fundamental frequency ($F_0$) perturbations and music separation artifact simulations (e.g., reverberation, echo), substantially improving performance under noisy conditions. Second, we enrich speaker representation using domain-specific singing data: alongside clean vocals, we incorporate DNSMOS-filtered separated vocals and public singing corpora, enabling the model to preserve speaker timbre while capturing singing style nuances. Third, we integrate the Neural Source-Filter (NSF) model to explicitly represent harmonic and noise components, enhancing the naturalness and controllability of converted singing. R2-SVC achieves state-of-the-art results on multiple SVC benchmarks under both clean and noisy conditions.

Method: DeEAR evaluates speech expressiveness across three dimensions (Emotion, Prosody, Spontaneity) using a framework grounded in phonetics and psychology. It requires fewer than 500 annotated samples and enables targeted data curation.

Result: DeEAR achieves strong alignment with human perception (SRCC=0.86). It distinguishes expressiveness gaps across S2S models and selected 14K expressive utterances to form ExpressiveSpeech, which improved S2S models’ expressive score from 2.0 to 23.4 on a 100-point scale.

Conclusion: DeEAR provides a reliable framework for evaluating speech expressiveness, enabling fair benchmarking and targeted data curation to improve the expressiveness of speech-to-speech models.

Abstract: Recent speech-to-speech (S2S) models generate intelligible speech but still lack natural expressiveness, largely due to the absence of a reliable evaluation metric. Existing approaches, such as subjective MOS ratings, low-level acoustic features, and emotion recognition are costly, limited, or incomplete. To address this, we present DeEAR (Decoding the Expressive Preference of eAR), a framework that converts human preference for speech expressiveness into an objective score. Grounded in phonetics and psychology, DeEAR evaluates speech across three dimensions: Emotion, Prosody, and Spontaneity, achieving strong alignment with human perception (Spearman’s Rank Correlation Coefficient, SRCC = 0.86) using fewer than 500 annotated samples. Beyond reliable scoring, DeEAR enables fair benchmarking and targeted data curation. It not only distinguishes expressiveness gaps across S2S models but also selects 14K expressive utterances to form ExpressiveSpeech, which improves the expressive score (from 2.0 to 23.4 on a 100-point scale) of S2S models. Demos and codes are available at https://github.com/FreedomIntelligence/ExpressiveSpeech

Method: Proposes a framework with sampling mechanism for proxy targets, trains lightweight translation model with joint objective for transformation and information preservation.

Result: Strong mood transformation performance while better retaining genre and instrumentation compared to training-free baselines on two datasets.

Conclusion: Controllable embedding transformation is a promising paradigm for personalized music retrieval, enabling targeted attribute modification while preserving other musical characteristics.

Abstract: Music representations are the backbone of modern recommendation systems, powering playlist generation, similarity search, and personalized discovery. Yet most embeddings offer little control for adjusting a single musical attribute, e.g., changing only the mood of a track while preserving its genre or instrumentation. In this work, we address the problem of controllable music retrieval through embedding-based transformation, where the objective is to retrieve songs that remain similar to a seed track but are modified along one chosen dimension. We propose a novel framework for mood-guided music embedding transformation, which learns a mapping from a seed audio embedding to a target embedding guided by mood labels, while preserving other musical attributes. Because mood cannot be directly altered in the seed audio, we introduce a sampling mechanism that retrieves proxy targets to balance diversity with similarity to the seed. We train a lightweight translation model using this sampling strategy and introduce a novel joint objective that encourages transformation and information preservation. Extensive experiments on two datasets show strong mood transformation performance while retaining genre and instrumentation far better than training-free baselines, establishing controllable embedding transformation as a promising paradigm for personalized music retrieval.

Method: Uses LeLM with two decoder-only transformers for parallel modeling of mixed tokens (combined vocals/accompaniment) and dual-track tokens (separate vocals/accompaniment), plus modular extension training and DPO-based multi-preference alignment.

Result: Significantly outperforms existing open-source methods in objective and subjective metrics, and performs competitively with industry systems.

Conclusion: LeVo effectively addresses key challenges in lyrics-to-song generation through its parallel token modeling approach and preference alignment method.

Abstract: Recent advances in large language models (LLMs) and audio language models have significantly improved music generation, particularly in lyrics-to-song generation. However, existing approaches still struggle with the complex composition of songs and the scarcity of high-quality data, leading to limitations in audio quality, musicality, instruction following, and vocal-instrument harmony. To address these challenges, we introduce LeVo, a language model based framework consisting of LeLM and Music Codec. LeLM is capable of parallel modeling of two types of tokens: mixed tokens, which represent the combined audio of vocals and accompaniment to achieve better vocal-instrument harmony, and dual-track tokens, which separately encode vocals and accompaniment for high-quality song generation. It employs two decoder-only transformers and a modular extension training strategy to prevent interference between different token types. To further enhance musicality and instruction following ability, we introduce a multi-preference alignment method based on Direct Preference Optimization (DPO). This method handles diverse human preferences through a semi-automatic data construction process and post-training. Experimental results demonstrate that LeVo significantly outperforms existing open-source methods in both objective and subjective metrics, while performing competitively with industry systems. Ablation studies further justify the effectiveness of our designs. Audio examples and source code are available at https://levo-demo.github.io and https://github.com/tencent-ailab/songgeneration.

Method: Conducted attention ablation experiments across RecurrentGemma-2B/9B and Jamba-Mini-1.6 models, tested sparsification of attention heads, and identified mechanistic requirements for retrieval tasks.

Result: Attention ablation caused catastrophic retrieval failure (0% accuracy), while SSM layers showed no compensatory mechanisms. Sparsifying attention to 15% of heads maintained near-perfect retrieval while preserving 84% MMLU performance. Identified that needle tokens must be exposed during generation and sufficient context must be available.

Conclusion: Hybrid architectures operate as specialized modules rather than integrated systems, with self-attention specializing primarily for retrieval tasks. This has implications for architecture optimization and interpretability.

Abstract: We demonstrate complete functional segregation in hybrid SSM-Transformer architectures: retrieval depends exclusively on self-attention layers. Across RecurrentGemma-2B/9B and Jamba-Mini-1.6, attention ablation causes catastrophic retrieval failure (0% accuracy), while SSM layers show no compensatory mechanisms even with improved prompting. Conversely, sparsifying attention to just 15% of heads maintains near-perfect retrieval while preserving 84% MMLU performance, suggesting self-attention specializes primarily for retrieval tasks. We identify precise mechanistic requirements for retrieval: needle tokens must be exposed during generation and sufficient context must be available during prefill or generation. This strict functional specialization challenges assumptions about redundancy in hybrid architectures and suggests these models operate as specialized modules rather than integrated systems, with immediate implications for architecture optimization and interpretability.

Method: Introduces NAS-DQN, which integrates a neural architecture search controller directly into the DRL training loop, allowing dynamic network reconfiguration based on cumulative performance feedback.

Result: NAS-DQN achieves superior final performance, sample efficiency, and policy stability compared to fixed-architecture baselines and random search, with negligible computational overhead. The learned search strategy significantly outperforms random exploration.

Conclusion: Architecture adaptation is necessary for optimal sample efficiency in online DRL, and RL agent design can be integrated as a dynamic component of the learning process rather than a static offline choice.

Abstract: The performance of deep reinforcement learning agents is fundamentally constrained by their neural network architecture, a choice traditionally made through expensive hyperparameter searches and then fixed throughout training. This work investigates whether online, adaptive architecture optimization can escape this constraint and outperform static designs. We introduce NAS-DQN, an agent that integrates a learned neural architecture search controller directly into the DRL training loop, enabling dynamic network reconfiguration based on cumulative performance feedback. We evaluate NAS-DQN against three fixed-architecture baselines and a random search control on a continuous control task, conducting experiments over multiple random seeds. Our results demonstrate that NAS-DQN achieves superior final performance, sample efficiency, and policy stability while incurring negligible computational overhead. Critically, the learned search strategy substantially outperforms both undirected random architecture exploration and poorly-chosen fixed designs, indicating that intelligent, performance-guided search is the key mechanism driving success. These findings establish that architecture adaptation is not merely beneficial but necessary for optimal sample efficiency in online deep reinforcement learning, and suggest that the design of RL agents need not be a static offline choice but can instead be seamlessly integrated as a dynamic component of the learning process itself.

Method: ReGraphT organizes CUDA optimization trajectories into structured reasoning graphs, models optimizations as state transitions, and uses Monte Carlo Graph Search for efficient exploration. It’s a training-free framework that transfers LLM-level reasoning to smaller models.

Result: ReGraphT outperforms HPC-specific fine-tuned models and other retrieval-augmented approaches, achieving 2.33X average speedup on CUDAEval and ParEval benchmarks. When paired with specific SLMs, it enables them to approach LLM-level performance without privacy risks or excessive computing overhead.

Conclusion: ReGraphT successfully bridges the gap between SLMs and LLMs for CUDA code generation, providing a privacy-friendly and computationally efficient solution that achieves comparable performance to large models while maintaining the benefits of smaller models.

Abstract: Despite significant evolution of CUDA programming and domain-specific libraries, effectively utilizing GPUs with massively parallel engines remains difficult. Large language models (LLMs) show strong potential in generating optimized CUDA code from sequential code. However, using LLMs in practice faces two major challenges: cloud-based APIs pose risks of code leakage, and local deployment is often computationally expensive and inefficient. These drawbacks have spurred interest in small language models (SLMs), which are more lightweight and privacy-friendly. Encouragingly, recent studies show that SLMs can achieve performance comparable to LLMs on specific tasks. While SLMs can match LLMs on domain-specific tasks, their limited reasoning abilities lead to suboptimal performance in complex CUDA generation according to our experiments. To bridge this gap, we propose ReGraphT, a training-free, retrieval-augmented generation framework that transfers LLM-level reasoning to smaller models. ReGraphT organizes CUDA optimization trajectories into a structured reasoning graph, modeling the combined CUDA optimizations as state transitions, and leverages Monte Carlo Graph Search (MCGS) for efficient exploration. We also present a CUDA-specific benchmark with difficulty tiers defined by reasoning complexity to evaluate models more comprehensively. Experiments show that ReGraphT outperforms HPC-specific fine-tuned models and other retrieval-augmented approaches, achieving an average 2.33X speedup on CUDAEval and ParEval. When paired with DeepSeek-Coder-V2-Lite-Instruct and Qwen2.5-Coder-7B-Instruct, ReGraphT enables SLMs to approach LLM-level performance without the associated privacy risks or excessive computing overhead.

Method: Uses deep neural networks with Transformer architecture to directly predict equilibrium path flows, capturing intricate correlations between OD pairs at the path level rather than link level.

Result: The model is orders of magnitude faster than conventional optimization, efficiently estimates path-level traffic flows in multi-class networks, and improves prediction accuracy by capturing detailed trip information.

Conclusion: The Transformer-based approach reduces computational costs, adapts flexibly to varying demand and network conditions, and enables rapid ‘what-if’ analyses for enhanced transportation planning and policy-making.

Abstract: The traffic assignment problem is essential for traffic flow analysis, traditionally solved using mathematical programs under the Equilibrium principle. These methods become computationally prohibitive for large-scale networks due to non-linear growth in complexity with the number of OD pairs. This study introduces a novel data-driven approach using deep neural networks, specifically leveraging the Transformer architecture, to predict equilibrium path flows directly. By focusing on path-level traffic distribution, the proposed model captures intricate correlations between OD pairs, offering a more detailed and flexible analysis compared to traditional link-level approaches. The Transformer-based model drastically reduces computation time, while adapting to changes in demand and network structure without the need for recalculation. Numerical experiments are conducted on the Manhattan-like synthetic network, the Sioux Falls network, and the Eastern-Massachusetts network. The results demonstrate that the proposed model is orders of magnitude faster than conventional optimization. It efficiently estimates path-level traffic flows in multi-class networks, reducing computational costs and improving prediction accuracy by capturing detailed trip and flow information. The model also adapts flexibly to varying demand and network conditions, supporting traffic management and enabling rapid `what-if’ analyses for enhanced transportation planning and policy-making.

Method: FairGRPO uses hierarchical reinforcement learning with adaptive importance weighting of advantages based on representation, task difficulty, and data source. It employs unsupervised clustering to handle missing demographic labels by automatically discovering latent demographic groups.

Result: Across 7 clinical datasets spanning 5 modalities, FairGRPO reduces predictive parity by 27.2% compared to baselines while improving F1 score by 12.49%. It progressively improves fairness during training, unlike baseline methods that show deteriorating fairness.

Conclusion: FairGRPO effectively addresses AI bias in medical diagnosis and enables the development of fairness-aware clinical models like FairMedGemma-4B, which achieves state-of-the-art performance with significantly reduced demographic disparities.

Abstract: Medical artificial intelligence systems have achieved remarkable diagnostic capabilities, yet they consistently exhibit performance disparities across demographic groups, causing real-world harm to underrepresented populations. While recent multimodal reasoning foundation models have advanced clinical diagnosis through integrated analysis of diverse medical data, reasoning trainings via reinforcement learning inherit and often amplify biases present in training datasets dominated by majority populations. We introduce Fairness-aware Group Relative Policy Optimization (FairGRPO), a hierarchical reinforcement learning approach that promotes equitable learning across heterogeneous clinical populations. FairGRPO employs adaptive importance weighting of advantages based on representation, task difficulty, and data source. To address the common issue of missing demographic labels in the clinical domain, we further employ unsupervised clustering, which automatically discovers latent demographic groups when labels are unavailable. Through comprehensive experiments across 7 clinical diagnostic datasets spanning 5 clinical modalities across X-ray, CT scan, dermoscropy, mammography and ultrasound, we demonstrate that FairGRPO reduces predictive parity by 27.2% against all vanilla and bias mitigated RL baselines, while improving F1 score by 12.49%. Furthermore, training dynamics analysis reveals that FairGRPO progressively improves fairness throughout optimization, while baseline RL methods exhibit deteriorating fairness as training progresses. Based on FairGRPO, we release FairMedGemma-4B, a fairness-aware clinical VLLM that achieves state-of-the-art performance while demonstrating significantly reduced disparities across demographic groups.

Method: Created sequential recycling scenario combining SortingEnv and ContainerGym tasks, evaluated modular vs monolithic architectures with and without action masking.

Result: Without action masking, modular architecture performs better; with action masking, both improve substantially and performance gap narrows.

Conclusion: Action space constraints are decisive, and specialization advantages diminish as action complexity is reduced. Provides valuable testbed for industrial multi-agent RL.

Abstract: Autonomous control of multi-stage industrial processes requires both local specialization and global coordination. Reinforcement learning (RL) offers a promising approach, but its industrial adoption remains limited due to challenges such as reward design, modularity, and action space management. Many academic benchmarks differ markedly from industrial control problems, limiting their transferability to real-world applications. This study introduces an enhanced industry-inspired benchmark environment that combines tasks from two existing benchmarks, SortingEnv and ContainerGym, into a sequential recycling scenario with sorting and pressing operations. We evaluate two control strategies: a modular architecture with specialized agents and a monolithic agent governing the full system, while also analyzing the impact of action masking. Our experiments show that without action masking, agents struggle to learn effective policies, with the modular architecture performing better. When action masking is applied, both architectures improve substantially, and the performance gap narrows considerably. These results highlight the decisive role of action space constraints and suggest that the advantages of specialization diminish as action complexity is reduced. The proposed benchmark thus provides a valuable testbed for exploring practical and robust multi-agent RL solutions in industrial automation, while contributing to the ongoing debate on centralization versus specialization.

Method: Used SHAP-based feature selection with XGBoost, LightGBM, and CatBoost algorithms, hyperparameter optimization via Optuna, and class balancing with SMOTE technique across six binary classification scenarios.

Result: SHAP-based feature selection maintained or improved performance metrics (balanced accuracy, precision, specificity) while enhancing model transparency and interpretability.

Conclusion: SHAP explainability techniques for feature selection provide an effective approach for developing transparent, reliable clinical decision support systems for urinary tract disease screening and early diagnosis.

Abstract: In this paper, we propose an approach to support the diagnosis of urinary tract diseases, with a focus on bladder cancer, using SHAP (SHapley Additive exPlanations)-based feature selection to enhance the transparency and effectiveness of predictive models. Six binary classification scenarios were developed to distinguish bladder cancer from other urological and oncological conditions. The algorithms XGBoost, LightGBM, and CatBoost were employed, with hyperparameter optimization performed using Optuna and class balancing with the SMOTE technique. The selection of predictive variables was guided by importance values through SHAP-based feature selection while maintaining or even improving performance metrics such as balanced accuracy, precision, and specificity. The use of explainability techniques (SHAP) for feature selection proved to be an effective approach. The proposed methodology may contribute to the development of more transparent, reliable, and efficient clinical decision support systems, optimizing screening and early diagnosis of urinary tract diseases.

Method: Formalizes thought communication as a latent variable model, proves identifiability of shared/private thoughts in nonparametric settings, and develops a framework to extract latent thoughts from agents and assign relevant thoughts with sharing patterns.

Result: Experiments on synthetic and real-world benchmarks validate the theory and demonstrate collaborative advantages of thought communication over traditional language-based approaches.

Conclusion: Thought communication illuminates the potential of leveraging hidden generative processes, as many challenges remain unsolvable through surface-level observation alone, regardless of computational or data scale.

Abstract: Natural language has long enabled human cooperation, but its lossy, ambiguous, and indirect nature limits the potential of collective intelligence. While machines are not subject to these constraints, most LLM-based multi-agent systems still rely solely on natural language, exchanging tokens or their embeddings. To go beyond language, we introduce a new paradigm, thought communication, which enables agents to interact directly mind-to-mind, akin to telepathy. To uncover these latent thoughts in a principled way, we formalize the process as a general latent variable model, where agent states are generated by an unknown function of underlying thoughts. We prove that, in a nonparametric setting without auxiliary information, both shared and private latent thoughts between any pair of agents can be identified. Moreover, the global structure of thought sharing, including which agents share which thoughts and how these relationships are structured, can also be recovered with theoretical guarantees. Guided by the established theory, we develop a framework that extracts latent thoughts from all agents prior to communication and assigns each agent the relevant thoughts, along with their sharing patterns. This paradigm naturally extends beyond LLMs to all modalities, as most observational data arise from hidden generative processes. Experiments on both synthetic and real-world benchmarks validate the theory and demonstrate the collaborative advantages of thought communication. We hope this work illuminates the potential of leveraging the hidden world, as many challenges remain unsolvable through surface-level observation alone, regardless of compute or data scale.

Method: Creates stochastic features by viewing datasets as realizations from random fields, maps them to Hilbert spaces, uses Kosambi-Karhunen-Loève expansion to break features into irreducible components, and performs classification via eigen-decomposition of operator spectra.

Result: Achieved state-of-the-art breakthroughs in Alzheimer’s Disease stage classification and remote sensing deforestation detection on challenging, data-deficient scientific domains.

Conclusion: FINDER provides a rigorous framework for noisy dataset classification but has specific failure modes and limitations that determine when it outperforms existing methods.

Abstract: ‘‘Noisy’’ datasets (regimes with low signal to noise ratios, small sample sizes, faulty data collection, etc) remain a key research frontier for classification methods with both theoretical and practical implications. We introduce FINDER, a rigorous framework for analyzing generic classification problems, with tailored algorithms for noisy datasets. FINDER incorporates fundamental stochastic analysis ideas into the feature learning and inference stages to optimally account for the randomness inherent to all empirical datasets. We construct ‘‘stochastic features’’ by first viewing empirical datasets as realizations from an underlying random field (without assumptions on its exact distribution) and then mapping them to appropriate Hilbert spaces. The Kosambi-Karhunen-Lo'eve expansion (KLE) breaks these stochastic features into computable irreducible components, which allow classification over noisy datasets via an eigen-decomposition: data from different classes resides in distinct regions, identified by analyzing the spectrum of the associated operators. We validate FINDER on several challenging, data-deficient scientific domains, producing state of the art breakthroughs in: (i) Alzheimer’s Disease stage classification, (ii) Remote sensing detection of deforestation. We end with a discussion on when FINDER is expected to outperform existing methods, its failure modes, and other limitations.

Method: Analyze inexact orthogonalized updates within Linear Minimization Oracle framework using additive error model to capture practical approximation schemes.

Result: Derived explicit bounds showing performance degradation with LMO inexactness, revealing fundamental coupling between approximation precision and optimal step size/momentum parameters.

Conclusion: Approximation procedure (e.g., number of Newton-Schulz steps) is a critical parameter that must be co-tuned with learning schedule, not just an implementation detail.

Abstract: The Muon optimizer has rapidly emerged as a powerful, geometry-aware alternative to AdamW, demonstrating strong performance in large-scale training of neural networks. However, a critical theory-practice disconnect exists: Muon’s efficiency relies on fast, approximate orthogonalization, yet all prior theoretical work analyzes an idealized, computationally intractable version assuming exact SVD-based updates. This work moves beyond the ideal by providing the first analysis of the inexact orthogonalized update at Muon’s core. We develop our analysis within the general framework of Linear Minimization Oracle (LMO)-based optimization, introducing a realistic additive error model to capture the inexactness of practical approximation schemes. Our analysis yields explicit bounds that quantify performance degradation as a function of the LMO inexactness/error. We reveal a fundamental coupling between this inexactness and the optimal step size and momentum: lower oracle precision requires a smaller step size but larger momentum parameter. These findings elevate the approximation procedure (e.g., the number of Newton-Schulz steps) from an implementation detail to a critical parameter that must be co-tuned with the learning schedule. NanoGPT experiments directly confirm the predicted coupling, with optimal learning rates clearly shifting as approximation precision changes.

Method: Developed two f-DP-based accounting methods: Pairwise Network f-DP for privacy leakage between user pairs under random-walk communication, and Secret-based f-Local DP for structured noise injection via shared secrets, combining f-DP theory with Markov chain concentration.

Result: Experiments show the methods yield consistently tighter (ε,δ) bounds and improved utility compared to Rényi DP-based approaches on both synthetic and real datasets.

Conclusion: The f-DP framework provides significant benefits for decentralized privacy accounting by better capturing privacy amplification from sparse communication, local iterations, and correlated noise.

Abstract: Differentially private (DP) decentralized Federated Learning (FL) allows local users to collaborate without sharing their data with a central server. However, accurately quantifying the privacy budget of private FL algorithms is challenging due to the co-existence of complex algorithmic components such as decentralized communication and local updates. This paper addresses privacy accounting for two decentralized FL algorithms within the $f$-differential privacy ($f$-DP) framework. We develop two new $f$-DP-based accounting methods tailored to decentralized settings: Pairwise Network $f$-DP (PN-$f$-DP), which quantifies privacy leakage between user pairs under random-walk communication, and Secret-based $f$-Local DP (Sec-$f$-LDP), which supports structured noise injection via shared secrets. By combining tools from $f$-DP theory and Markov chain concentration, our accounting framework captures privacy amplification arising from sparse communication, local iterations, and correlated noise. Experiments on synthetic and real datasets demonstrate that our methods yield consistently tighter $(\epsilon,\delta)$ bounds and improved utility compared to R'enyi DP-based approaches, illustrating the benefits of $f$-DP in decentralized privacy accounting.

Method: Using tools from Kaczmarz method literature to formalize greedy orderings, developing geometric and algebraic intuitions, and conducting empirical analysis on linear regression and CIFAR-100 classification.

Result: Greedy orderings converge faster than random ones in average loss across tasks. In high-rank settings, greedy orderings match random ones’ loss bounds, but under general rank, single-pass greedy may fail catastrophically while repetition-enabled greedy converges at O(1/∛k) rate.

Conclusion: Greedy orderings offer benefits over random ones but require careful consideration of repetition strategies, revealing important nuances in continual learning task ordering strategies.

Abstract: We analyze task orderings in continual learning for linear regression, assuming joint realizability of training data. We focus on orderings that greedily maximize dissimilarity between consecutive tasks, a concept briefly explored in prior work but still surrounded by open questions. Using tools from the Kaczmarz method literature, we formalize such orderings and develop geometric and algebraic intuitions around them. Empirically, we demonstrate that greedy orderings converge faster than random ones in terms of the average loss across tasks, both for linear regression with random data and for linear probing on CIFAR-100 classification tasks. Analytically, in a high-rank regression setting, we prove a loss bound for greedy orderings analogous to that of random ones. However, under general rank, we establish a repetition-dependent separation. Specifically, while prior work showed that for random orderings, with or without replacement, the average loss after $k$ iterations is bounded by $\mathcal{O}(1/\sqrt{k})$, we prove that single-pass greedy orderings may fail catastrophically, whereas those allowing repetition converge at rate $\mathcal{O}(1/\sqrt[3]{k})$. Overall, we reveal nuances within and between greedy and random orderings.

Method: Developed a novel class of elliptic uncertainty sets to model directional market impact. Established both implicit and explicit closed-form solutions for worst-case uncertainty under these sets, enabling efficient robust policy evaluation.

Result: Experiments on single-asset and multi-asset trading tasks show the method achieves superior Sharpe ratio and remains robust under increasing trade volumes compared to traditional approaches.

Conclusion: The proposed elliptic uncertainty sets offer a more faithful and scalable approach to RL in financial markets by properly capturing directional market impact effects.

Abstract: In financial applications, reinforcement learning (RL) agents are commonly trained on historical data, where their actions do not influence prices. However, during deployment, these agents trade in live markets where their own transactions can shift asset prices, a phenomenon known as market impact. This mismatch between training and deployment environments can significantly degrade performance. Traditional robust RL approaches address this model misspecification by optimizing the worst-case performance over a set of uncertainties, but typically rely on symmetric structures that fail to capture the directional nature of market impact. To address this issue, we develop a novel class of elliptic uncertainty sets. We establish both implicit and explicit closed-form solutions for the worst-case uncertainty under these sets, enabling efficient and tractable robust policy evaluation. Experiments on single-asset and multi-asset trading tasks demonstrate that our method achieves superior Sharpe ratio and remains robust under increasing trade volumes, offering a more faithful and scalable approach to RL in financial markets.

Method: Reformulates directional derivatives as telescoping series and samples from carefully designed distributions to construct unbiased gradient estimators based solely on function evaluations.

Result: Derived optimal scaling distributions and perturbation stepsizes for four specific constructions, and proved that SGD using the proposed estimators achieves optimal complexity for smooth non-convex objectives.

Conclusion: The proposed unbiased gradient estimators demonstrate superior accuracy and convergence compared to standard methods in both synthetic tasks and language model fine-tuning applications.

Abstract: Zeroth-order optimization (ZOO) is an important framework for stochastic optimization when gradients are unavailable or expensive to compute. A potential limitation of existing ZOO methods is the bias inherent in most gradient estimators unless the perturbation stepsize vanishes. In this paper, we overcome this biasedness issue by proposing a novel family of unbiased gradient estimators based solely on function evaluations. By reformulating directional derivatives as a telescoping series and sampling from carefully designed distributions, we construct estimators that eliminate bias while maintaining favorable variance. We analyze their theoretical properties, derive optimal scaling distributions and perturbation stepsizes of four specific constructions, and prove that SGD using the proposed estimators achieves optimal complexity for smooth non-convex objectives. Experiments on synthetic tasks and language model fine-tuning confirm the superior accuracy and convergence of our approach compared to standard methods.

Method: Formulated as a constrained functional optimization problem over perturbation distributions, the paper develops directionally aligned perturbations (DAP) that adaptively provide higher accuracy along critical directions, with convergence analysis for stochastic gradient descent using δ-unbiased perturbations.

Result: Theoretical and empirical analysis shows that directionally aligned perturbations can minimize asymptotic variance and outperform traditional fixed-length perturbation methods under specific conditions.

Conclusion: Directionally aligned perturbations offer significant advantages over traditional fixed-length perturbations for zeroth-order gradient estimation, providing higher accuracy along critical directions and extending convergence analysis to a wider range of perturbation schemes.

Abstract: In this paper, we explore the two-point zeroth-order gradient estimator and identify the distribution of random perturbations that minimizes the estimator’s asymptotic variance as the perturbation stepsize tends to zero. We formulate it as a constrained functional optimization problem over the space of perturbation distributions. Our findings reveal that such desired perturbations can align directionally with the true gradient, instead of maintaining a fixed length. While existing research has largely focused on fixed-length perturbations, the potential advantages of directional alignment have been overlooked. To address this gap, we delve into the theoretical and empirical properties of the directionally aligned perturbation (DAP) scheme, which adaptively offers higher accuracy along critical directions. Additionally, we provide a convergence analysis for stochastic gradient descent using $\delta$-unbiased random perturbations, extending existing complexity bounds to a wider range of perturbations. Through empirical evaluations on both synthetic problems and practical tasks, we demonstrate that DAPs outperform traditional methods under specific conditions.

Method: UCAN transforms existing randomized smoothing methods from symmetric to asymmetric noise distributions, supports various noise distributions for different ℓ_p-norms, and uses noise parameter generators to optimize anisotropic noise parameters per data dimension.

Result: Empirical evaluations show up to 182.6% improvement in certified accuracy at large certified radii on MNIST, CIFAR10, and ImageNet datasets compared to state-of-the-art methods.

Conclusion: UCAN provides a versatile framework for enhancing certified adversarial robustness through anisotropic noise, significantly outperforming existing isotropic approaches while maintaining broad applicability.

Abstract: Randomized smoothing has become essential for achieving certified adversarial robustness in machine learning models. However, current methods primarily use isotropic noise distributions that are uniform across all data dimensions, such as image pixels, limiting the effectiveness of robustness certification by ignoring the heterogeneity of inputs and data dimensions. To address this limitation, we propose UCAN: a novel technique that \underline{U}niversally \underline{C}ertifies adversarial robustness with \underline{A}nisotropic \underline{N}oise. UCAN is designed to enhance any existing randomized smoothing method, transforming it from symmetric (isotropic) to asymmetric (anisotropic) noise distributions, thereby offering a more tailored defense against adversarial attacks. Our theoretical framework is versatile, supporting a wide array of noise distributions for certified robustness in different $\ell_p$-norms and applicable to any arbitrary classifier by guaranteeing the classifier’s prediction over perturbed inputs with provable robustness bounds through tailored noise injection. Additionally, we develop a novel framework equipped with three exemplary noise parameter generators (NPGs) to optimally fine-tune the anisotropic noise parameters for different data dimensions, allowing for pursuing different levels of robustness enhancements in practice.Empirical evaluations underscore the significant leap in UCAN’s performance over existing state-of-the-art methods, demonstrating up to $182.6%$ improvement in certified accuracy at large certified radii on MNIST, CIFAR10, and ImageNet datasets.\footnote{Code is anonymously available at \href{https://github.com/youbin2014/UCAN/}{https://github.com/youbin2014/UCAN/}}

Method: The proposed method is AMRC (Adaptive Masking Loss with Representation Consistency), which consists of two components: 1) Dynamic masking loss that adaptively identifies discriminative temporal segments to guide gradient descent, and 2) Representation consistency constraint that stabilizes mapping relationships among inputs, labels, and predictions.

Result: Experimental results demonstrate that AMRC effectively suppresses redundant feature learning while significantly improving model performance, showing that appropriately truncating historical data can paradoxically enhance prediction accuracy.

Conclusion: This work challenges conventional assumptions in temporal modeling and provides novel theoretical insights and methodological breakthroughs for developing efficient and robust forecasting models in domains like energy management and financial markets.

Abstract: Time series forecasting plays a pivotal role in critical domains such as energy management and financial markets. Although deep learning-based approaches (e.g., MLP, RNN, Transformer) have achieved remarkable progress, the prevailing “long-sequence information gain hypothesis” exhibits inherent limitations. Through systematic experimentation, this study reveals a counterintuitive phenomenon: appropriately truncating historical data can paradoxically enhance prediction accuracy, indicating that existing models learn substantial redundant features (e.g., noise or irrelevant fluctuations) during training, thereby compromising effective signal extraction. Building upon information bottleneck theory, we propose an innovative solution termed Adaptive Masking Loss with Representation Consistency (AMRC), which features two core components: 1) Dynamic masking loss, which adaptively identified highly discriminative temporal segments to guide gradient descent during model training; 2) Representation consistency constraint, which stabilized the mapping relationships among inputs, labels, and predictions. Experimental results demonstrate that AMRC effectively suppresses redundant feature learning while significantly improving model performance. This work not only challenges conventional assumptions in temporal modeling but also provides novel theoretical insights and methodological breakthroughs for developing efficient and robust forecasting models.

Method: Proposed reasoning-as-infilling using templates to structure outputs and distinguish reasoning from answers, and multi-token entropy decoding (MED) - an adaptive sampler that minimizes parallel decoding errors based on conditional entropies.

Result: Fine-tuning on posterior reasoning traces from MDLMs provides performance boosts comparable to human-written traces. MED preserves performance while reducing steps by 2.7x.

Conclusion: MDLMs’ training and compute unlock novel inference and post-training methods that go beyond traditional left-to-right decoding, enabling structured reasoning, uncertainty measurement, and efficient parallel decoding.

Abstract: Masked diffusion language models (MDLMs) are trained to in-fill positions in randomly masked sequences, in contrast to next-token prediction models. Discussions around MDLMs focus on two benefits: (1) any-order decoding and 2) multi-token decoding. However, we observe that for math and coding tasks, any-order algorithms often underperform or behave similarly to left-to-right sampling, and standard multi-token decoding significantly degrades performance. At inference time, MDLMs compute the conditional distribution of all masked positions. A natural question is: How can we justify this additional compute when left-to-right one-token-at-a-time decoding is on par with any-order decoding algorithms? First, we propose reasoning-as-infilling. By using MDLMs to infill a reasoning template, we can structure outputs and distinguish between reasoning and answer tokens. In turn, this enables measuring answer uncertainty during reasoning, and early exits when the model converges on an answer. Next, given an answer, reasoning-as-infilling enables sampling from the MDLM posterior over reasoning traces conditioned on the answer, providing a new source of high-quality data for post-training. On GSM8k, we observe that fine-tuning LLaDA-8B Base on its posterior reasoning traces provides a performance boost on par with fine-tuning on human-written reasoning traces. Additionally, given an answer, reasoning-as-infilling provides a method for scoring the correctness of the reasoning process at intermediate steps. Second, we propose multi-token entropy decoding (MED), a simple adaptive sampler that minimizes the error incurred by decoding positions in parallel based on the conditional entropies of those positions. MED preserves performance across benchmarks and leads to 2.7x fewer steps. Our work demonstrates that the training and compute used by MDLMs unlock many new inference and post-training methods.

Method: Supervised learning simulation using realistic RSSI data with Decision Tree classification on CAD-modeled floor plan, trained on 5,000 balanced observations with class weights to handle class imbalance.

Result: Overall accuracy of 34.2% with F1-scores >0.40 for multiple zones, but rare classes (especially LabZoneC) were often misclassified despite class weights.

Conclusion: RSSI-based decision trees can enable zone-level anomaly detection for defense logistics, but performance in low-coverage zones needs improvement through better antenna placement or sensor fusion.

Abstract: Radio Frequency Identification (RFID) tracking may be a viable solution for defense assets that must be stored in accordance with security guidelines. However, poor sensor specificity (vulnerabilities include long range detection, spoofing, and counterfeiting) can lead to erroneous detection and operational security events. We present a supervised learning simulation with realistic Received Signal Strength Indicator (RSSI) data and Decision Tree classification in a Computer Assisted Design (CAD)-modeled floor plan that encapsulates some of the challenges encountered in defense storage. In this work, we focused on classifying 12 lab zones (LabZoneA-L) to perform location inference. The raw dataset had approximately 980,000 reads. Class frequencies were imbalanced, and class weights were calculated to account for class imbalance in this multi-class setting. The model, trained on stratified subsamples to 5,000 balanced observations, yielded an overall accuracy of 34.2% and F1-scores greater than 0.40 for multiple zones (Zones F, G, H, etc.). However, rare classes (most notably LabZoneC) were often misclassified, even with the use of class weights. An adjacency-aware confusion matrix was calculated to allow better interpretation of physically adjacent zones. These results suggest that RSSI-based decision trees can be applied in realistic simulations to enable zone-level anomaly detection or misplacement monitoring for defense supply logistics. Reliable classification performance in low-coverage and low-signal zones could be improved with better antenna placement or additional sensors and sensor fusion with other modalities.

Method: Constructs a graph from trajectories of the same prompt to quantify step quality and assign advantages accordingly, serving as a plug-and-play module for existing group-based RL algorithms without rollout modifications.

Result: Extensive experiments on WebShop, ALFWorld, and AppWorld benchmarks with various model sizes show consistent performance improvements.

Conclusion: SALT effectively addresses the coarse reward assignment problem in group-based RL, providing actionable insights for better policy learning in complex multi-step tasks.

Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities, enabling language agents to excel at single-turn tasks. However, their application to complex, multi-step, and long-horizon tasks remains challenging. While reinforcement learning (RL) offers a promising avenue for addressing these challenges, mainstream approaches typically rely solely on sparse, outcome-based rewards, a limitation that becomes especially problematic for group-based RL algorithms lacking critic models, such as Group Relative Policy Optimization (GRPO). In such methods, uniformly rewarding or penalizing all actions within a trajectory can lead to training instability and suboptimal policies, because beneficial and detrimental actions are often entangled across multi-step interactions. To address this challenge, we propose SALT, a novel and lightweight framework that provides a finer-grained advantage assignment, derived solely from outcome rewards. We achieve this by constructing a graph from trajectories of the same prompt, which allows us to quantify the quality of each step and assign advantages accordingly. Crucially, SALT is designed as a plug-and-play module that seamlessly integrates with existing group-based RL algorithms, requiring no modifications to the rollout procedure and introducing negligible computational overhead. Extensive experiments on the WebShop, ALFWorld, and AppWorld benchmarks with various model sizes demonstrate that SALT consistently improves performance. We also conduct a thorough analysis to validate the design choices behind SALT and offer actionable insights.

Method: Created a temporal, heterogeneous graph encompassing complete Bitcoin transaction history with over 2.4B nodes and 39.72B edges, plus custom sampling methods, tools for graph databases, and ready-to-use database snapshots.

Result: Developed a comprehensive dataset and toolkit that models Bitcoin’s complete economic topology, enabling ML researchers to analyze Bitcoin’s intricate ecosystem at scale.

Conclusion: This work empowers the ML community to tackle Bitcoin applications like anomaly detection, address classification, market analysis, and large-scale graph ML benchmarking through accessible graph data and tools.

Abstract: Since its 2009 genesis block, the Bitcoin network has processed \num{>1.08} billion (B) transactions representing \num{>8.72}B BTC, offering rich potential for machine learning (ML); yet, its pseudonymity and obscured flow of funds inherent in its \utxo-based design, have rendered this data largely inaccessible for ML research. Addressing this gap, we present an ML-compatible graph modeling the Bitcoin’s economic topology by reconstructing the flow of funds. This temporal, heterogeneous graph encompasses complete transaction history up to block \cutoffHeight, consisting of \num{>2.4}B nodes and \num{>39.72}B edges. Additionally, we provide custom sampling methods yielding node and edge feature vectors of sampled communities, tools to load and analyze the Bitcoin graph data within specialized graph databases, and ready-to-use database snapshots. This comprehensive dataset and toolkit empower the ML community to tackle Bitcoin’s intricate ecosystem at scale, driving progress in applications such as anomaly detection, address classification, market analysis, and large-scale graph ML benchmarking. Dataset and code available at \href{https://github.com/B1AAB/EBA}{github.com/b1aab/eba}

Method: Proposed a rejection sampling-based algorithm that enables exact parallel sampling of multiple future values from existing TPP models without requiring model modifications or retraining.

Result: The method provides theoretical guarantees and demonstrates empirical speedups on real-world datasets.

Conclusion: The approach successfully bridges the gap between expressive modeling and efficient parallel generation for large-scale temporal point process applications.

Abstract: Temporal point processes are powerful generative models for event sequences that capture complex dependencies in time-series data. They are commonly specified using autoregressive models that learn the distribution of the next event from the previous events. This makes sampling inherently sequential, limiting efficiency. In this paper, we propose a novel algorithm based on rejection sampling that enables exact sampling of multiple future values from existing TPP models, in parallel, and without requiring any architectural changes or retraining. Besides theoretical guarantees, our method demonstrates empirical speedups on real-world datasets, bridging the gap between expressive modeling and efficient parallel generation for large-scale TPP applications.

Method: Models ad bidding as Contextual Markov Decision Process (CMDP) with delayed Poisson rewards, uses two-stage maximum likelihood estimator with data-splitting, and designs reinforcement learning algorithm for personalized bidding strategies.

Result: Achieves near-optimal regret bound of O~(dH²√T) where d is contextual dimension, H is number of rounds, and T is number of customers. Theoretical findings validated through simulation experiments.

Conclusion: The proposed framework effectively captures complex ad impact factors and provides efficient personalized bidding strategies with strong theoretical guarantees.

Abstract: Online advertising platforms use automated auctions to connect advertisers with potential customers, requiring effective bidding strategies to maximize profits. Accurate ad impact estimation requires considering three key factors: delayed and long-term effects, cumulative ad impacts such as reinforcement or fatigue, and customer heterogeneity. However, these effects are often not jointly addressed in previous studies. To capture these factors, we model ad bidding as a Contextual Markov Decision Process (CMDP) with delayed Poisson rewards. For efficient estimation, we propose a two-stage maximum likelihood estimator combined with data-splitting strategies, ensuring controlled estimation error based on the first-stage estimator’s (in)accuracy. Building on this, we design a reinforcement learning algorithm to derive efficient personalized bidding strategies. This approach achieves a near-optimal regret bound of $\tilde{O}{(dH^2\sqrt{T})}$, where $d$ is the contextual dimension, $H$ is the number of rounds, and $T$ is the number of customers. Our theoretical findings are validated by simulation experiments.

Method: Designs an algorithm that evaluates all draft models without additional target model queries, applicable to various speculative decoding methods (single draft, multi-drafts, draft-trees) with system-efficient implementations.

Result: Substantially outperforms state-of-the-art EAGLE3 and BanditSpec baselines across diverse datasets and LLMs, especially in domains requiring long reasoning chains with specialized drafters.

Conclusion: The proposed online draft model selection approach provides significant performance improvements in speculative decoding while maintaining computational efficiency.

Abstract: Speculative decoding is widely used in accelerating large language model (LLM) inference. In this work, we focus on the online draft model selection problem in speculative decoding. We design an algorithm that provably competes with the best draft model in hindsight for each query in terms of either the token acceptance probability or expected acceptance length. In particular, we show that we can accurately evaluate all draft models, instead of only the chosen model without incurring additional queries to the target model, which allows us to improve exponentially over the existing bandit-based approach as the number of draft models increases. Our approach is generically applicable with any speculative decoding methods (single draft, multi-drafts and draft-trees). Moreover, we design system-efficient versions of online learners and demonstrate that the overhead in computation and latency can be substantially reduced. We conduct extensive experiments on open-source LLMs and diverse datasets, demonstrating that our methods substantially outperform the state-of-the-art EAGLE3 and the BanditSpec baseline in a variety of domains where specialized domain-expert drafters are available, especially when long reasoning chains are required.

Method: Three-layer statistical framework: (A) core liquidity-return interactions, (B) principal-component relations, (C) volatility-factor projections; complemented by VAR models, HAR-X models, and leakage-safe machine learning with temporal splits and SHAP interpretation.

Result: Statistically significant Granger-causal relationships across layers and moderate out-of-sample predictive accuracy using daily data from 2021-2025 (1462 observations across 74 assets).

Conclusion: Liquidity and volatility proxies from core cryptoassets generate spillovers that can forecast market-wide risk, with documented statistical significance and moderate predictive performance.

Abstract: We study whether liquidity and volatility proxies of a core set of cryptoassets generate spillovers that forecast market-wide risk. Our empirical framework integrates three statistical layers: (A) interactions between core liquidity and returns, (B) principal-component relations linking liquidity and returns, and (C) volatility-factor projections that capture cross-sectional volatility crowding. The analysis is complemented by vector autoregression impulse responses and forecast error variance decompositions (see Granger 1969; Sims 1980), heterogeneous autoregressive models with exogenous regressors (HAR-X, Corsi 2009), and a leakage-safe machine learning protocol using temporal splits, early stopping, validation-only thresholding, and SHAP-based interpretation. Using daily data from 2021 to 2025 (1462 observations across 74 assets), we document statistically significant Granger-causal relationships across layers and moderate out-of-sample predictive accuracy. We report the most informative figures, including the pipeline overview, Layer A heatmap, Layer C robustness analysis, vector autoregression variance decompositions, and the test-set precision-recall curve. Full data and figure outputs are provided in the artifact repository.

Method: Uses transformer encoders/decoders with orthogonal latent subspaces to partition shared and private neural dynamics across brain regions.

Result: Outperforms existing approaches in decoding behavioral variables from multi-region electrophysiology datasets.

Conclusion: CTAE effectively captures both shared and region-specific neural dynamics while maintaining temporal structure.

Abstract: Simultaneous recordings from thousands of neurons across multiple brain areas reveal rich mixtures of activity that are shared between regions and dynamics that are unique to each region. Existing alignment or multi-view methods neglect temporal structure, whereas dynamical latent variable models capture temporal dependencies but are usually restricted to a single area, assume linear read-outs, or conflate shared and private signals. We introduce the Coupled Transformer Autoencoder (CTAE) - a sequence model that addresses both (i) non-stationary, non-linear dynamics and (ii) separation of shared versus region-specific structure in a single framework. CTAE employs transformer encoders and decoders to capture long-range neural dynamics and explicitly partitions each region’s latent space into orthogonal shared and private subspaces. We demonstrate the effectiveness of CTAE on two high-density electrophysiology datasets with simultaneous recordings from multiple regions, one from motor cortical areas and the other from sensory areas. CTAE extracts meaningful representations that better decode behavioral variables compared to existing approaches.

Method: ShapeX segments time series into shapelet-driven segments using the Shapelet Describe-and-Detect (SDD) framework to learn diverse shapelets, then employs Shapley values to evaluate their saliency.

Result: ShapeX outperforms existing methods in identifying relevant subsequences on both synthetic and real-world datasets, producing explanations that reveal causal relationships rather than just correlations.

Conclusion: ShapeX successfully bridges the gap in time series explanation by focusing on shapelets as core features, providing more precise and causally faithful explanations for time series classification models.

Abstract: Explaining time series classification models is crucial, particularly in high-stakes applications such as healthcare and finance, where transparency and trust play a critical role. Although numerous time series classification methods have identified key subsequences, known as shapelets, as core features for achieving state-of-the-art performance and validating their pivotal role in classification outcomes, existing post-hoc time series explanation (PHTSE) methods primarily focus on timestep-level feature attribution. These explanation methods overlook the fundamental prior that classification outcomes are predominantly driven by key shapelets. To bridge this gap, we present ShapeX, an innovative framework that segments time series into meaningful shapelet-driven segments and employs Shapley values to assess their saliency. At the core of ShapeX lies the Shapelet Describe-and-Detect (SDD) framework, which effectively learns a diverse set of shapelets essential for classification. We further demonstrate that ShapeX produces explanations which reveal causal relationships instead of just correlations, owing to the atomicity properties of shapelets. Experimental results on both synthetic and real-world datasets demonstrate that ShapeX outperforms existing methods in identifying the most relevant subsequences, enhancing both the precision and causal fidelity of time series explanations.

Method: Proposes a hierarchical dual-head neural network based on MentalRoBERTa with two prediction heads: CORAL head for ordinal relationships and standard classification head for categorical distinctions. Uses 3-layer Transformer encoder with multi-head attention for temporal dependencies and time interval embeddings. Trained with combined loss function (0.5 CORAL + 0.3 Cross-Entropy + 0.2 Focal Loss) and employs frozen layers and mixed-precision training for efficiency.

Result: The model is evaluated using 5-fold stratified cross-validation with macro F1 score as the primary metric (specific performance results not provided in abstract).

Conclusion: The proposed approach effectively addresses the key challenges in suicide risk classification from social media by combining ordinal and categorical modeling, temporal dependency capture, and specialized loss functions for class imbalance.

Abstract: Social media platforms have become important sources for identifying suicide risk, but automated detection systems face multiple challenges including severe class imbalance, temporal complexity in posting patterns, and the dual nature of risk levels as both ordinal and categorical. This paper proposes a hierarchical dual-head neural network based on MentalRoBERTa for suicide risk classification into four levels: indicator, ideation, behavior, and attempt. The model employs two complementary prediction heads operating on a shared sequence representation: a CORAL (Consistent Rank Logits) head that preserves ordinal relationships between risk levels, and a standard classification head that enables flexible categorical distinctions. A 3-layer Transformer encoder with 8-head multi-head attention models temporal dependencies across post sequences, while explicit time interval embeddings capture posting behavior dynamics. The model is trained with a combined loss function (0.5 CORAL + 0.3 Cross-Entropy + 0.2 Focal Loss) that simultaneously addresses ordinal structure preservation, overconfidence reduction, and class imbalance. To improve computational efficiency, we freeze the first 6 layers (50%) of MentalRoBERTa and employ mixed-precision training. The model is evaluated using 5-fold stratified cross-validation with macro F1 score as the primary metric.

Method: A DDQN agent directly competes against strong baselines (GES or GraN-DAG), always warm-starting from the opponent’s solution, providing provable guarantees including never being worse than the opponent and accelerated convergence.

Result: Achieves first-of-its-kind finite-sample guarantees with observed error probability decaying with sample size, and consistently improves upon GES and GraN-DAG on real-world benchmarks while scaling to large graphs (up to 220 nodes).

Conclusion: Establishes a new class of RL-based causal discovery algorithms that are simultaneously provably consistent, sample-efficient, and practically scalable, unifying empirical performance with rigorous finite-sample theory.

Abstract: Causal discovery remains a central challenge in machine learning, yet existing methods face a fundamental gap: algorithms like GES and GraN-DAG achieve strong empirical performance but lack finite-sample guarantees, while theoretically principled approaches fail to scale. We close this gap by introducing a game-theoretic reinforcement learning framework for causal discovery, where a DDQN agent directly competes against a strong baseline (GES or GraN-DAG), always warm-starting from the opponent’s solution. This design yields three provable guarantees: the learned graph is never worse than the opponent, warm-starting strictly accelerates convergence, and most importantly, with high probability the algorithm selects the true best candidate graph. To the best of our knowledge, our result makes a first-of-its-kind progress in explaining such finite-sample guarantees in causal discovery: on synthetic SEMs (30 nodes), the observed error probability decays with n, tightly matching theory. On real-world benchmarks including Sachs, Asia, Alarm, Child, Hepar2, Dream, and Andes, our method consistently improves upon GES and GraN-DAG while remaining theoretically safe. Remarkably, it scales to large graphs such as Hepar2 (70 nodes), Dream (100 nodes), and Andes (220 nodes). Together, these results establish a new class of RL-based causal discovery algorithms that are simultaneously provably consistent, sample-efficient, and practically scalable, marking a decisive step toward unifying empirical performance with rigorous finite-sample theory.

Method: Developed a novel weighting scheme for each dimension, incorporated into KNN classifier using weighted Minkowski distance. Analyzed impact of distance norms and dimension weights with visualization.

Result: The method performed well across diverse experiments, showing significant and consistent gain in classification accuracy (around 10%) for gene expression datasets in all cross-validation experiments with different k values.

Conclusion: The approach stands as an important generalization of KNN classifier powered by weighted Minkowski distance with the novel weighting schema, effectively handling high-dimensional datasets with limited samples by regulating the shape and size of neighborhood regions.

Abstract: Studies on various facets of pattern classification is often imperative while working with multi-dimensional samples pertaining to diverse application scenarios. In this notion, weighted dimension-based distance measure has been one of the vital considerations in pattern analysis as it reflects the degree of similarity between samples. Though it is often presumed to be settled with the pervasive use of Euclidean distance, plethora of issues often surface. In this paper, we present (a) a detail analysis on the impact of distance measure norms and weights of dimensions along with visualization, (b) a novel weighting scheme for each dimension, (c) incorporation of this dimensional weighting schema into a KNN classifier, and (d) pattern classification on a variety of synthetic as well as realistic datasets with the developed model. It has performed well across diverse experiments in comparison to the traditional KNN under the same experimental setups. Specifically, for gene expression datasets, it yields significant and consistent gain in classification accuracy (around 10%) in all cross-validation experiments with different values of k. As such datasets contain limited number of samples of high dimensions, meaningful selection of nearest neighbours is desirable, and this requirement is reasonably met by regulating the shape and size of the region enclosing the k number of reference samples with the developed weighting schema and appropriate norm. It, therefore, stands as an important generalization of KNN classifier powered by weighted Minkowski distance with the present weighting schema.

Method: A fully decoupled training strategy that learns prototypes and encoders under separate objectives. Prototypes are modeled as a Gaussian mixture updated with an online EM-style procedure independent of the encoder’s loss.

Result: The decoupling eliminates prototype collapse without explicit regularization, yielding consistently diverse prototypes and stronger downstream performance.

Conclusion: Breaking the joint optimization of encoders and prototypes through decoupled training effectively addresses the root cause of prototype collapse and improves representation learning.

Abstract: Prototypical self-supervised learning methods consistently suffer from partial prototype collapse, where multiple prototypes converge to nearly identical representations. This undermines their central purpose – providing diverse and informative targets to guide encoders toward rich representations – and has led practitioners to over-parameterize prototype sets or add ad-hoc regularizers, which mitigate symptoms rather than address the root cause. We empirically trace the collapse to the joint optimization of encoders and prototypes, which encourages a type of shortcut learning: early in training prototypes drift toward redundant representations that minimize loss without necessarily enhancing representation diversity. To break the joint optimization, we introduce a fully decoupled training strategy that learns prototypes and encoders under separate objectives. Concretely, we model prototypes as a Gaussian mixture updated with an online EM-style procedure, independent of the encoder’s loss. This simple yet principled decoupling eliminates prototype collapse without explicit regularization and yields consistently diverse prototypes and stronger downstream performance.

Method: Propose shifting from opportunistic data aggregation to principled dataset construction that systematically spans the space of temporal invariances, building ontologies based on first principles.

Result: The paper argues that current approaches fail because timeseries data lacks the dense concept coverage of web-scale text/image data, and the scrape-everything paradigm doesn’t work for timeseries.

Conclusion: True progress in timeseries foundation models requires ensuring representational completeness through systematic invariance coverage to achieve proper generalization, reasoning, and emergent behavior.

Abstract: Timeseries foundation models (TSFMs) have multiplied, yet lightweight supervised baselines and even classical models often match them. We argue this gap stems from the naive importation of NLP or CV pipelines. In language and vision, large web-scale corpora densely capture human concepts i.e. there are countless images and text of apples. In contrast, timeseries data is built to complement the image and text modalities. There are no timeseries dataset that contains the concept apple. As a result, the scrape-everything-online paradigm fails for TS. We posit that progress demands a shift from opportunistic aggregation to principled design: constructing datasets that systematically span the space of invariance that preserve temporal semantics. To this end, we suggest that the ontology of timeseries invariances should be built based on first principles. Only by ensuring representational completeness through invariance coverage can TSFMs achieve the aligned structure necessary for generalisation, reasoning, and truly emergent behaviour.

Method: Used a multi-layer graph diffusion model on longitudinal neuroimaging data to examine SC-FC interactions in tau propagation across brain networks.

Result: Found regionally asymmetric contributions: FC drives tau spread in subcortical areas, insula, frontal and temporal cortices, while SC dominates in occipital, parietal, and limbic regions. The SC-FC dominance shifts over disease course, with FC prevailing early and SC later. These patterns align with AD-associated gene expression.

Conclusion: SC and FC asymmetrically constrain tau propagation in Alzheimer’s disease, with their relative dominance shifting over disease progression and being influenced by genetic factors and biological mechanisms.

Abstract: Emerging neuroimaging evidence shows that pathological tau proteins build up along specific brain networks, suggesting that large-scale network architecture plays a key role in the progression of Alzheimer’s disease (AD). However, how structural connectivity (SC) and functional connectivity (FC) interact to influence tau propagation remains unclear. Leveraging an unprecedented volume of longitudinal neuroimaging data, we examine SC-FC interactions through a multi-layer graph diffusion model. Beyond showing that connectome architecture constrains tau spread, our model reveals a regionally asymmetric contribution of SC and FC. Specifically, FC predominantly drives tau spread in subcortical areas, the insula, frontal and temporal cortices, whereas SC plays a larger role in occipital, parietal, and limbic regions. The relative dominance of SC versus FC shifts over the course of disease, with FC generally prevailing in early AD and SC becoming primary in later stages. Spatial patterns of SC- and FC-dominant regions strongly align with the regional expression of AD-associated genes involved in inflammation, apoptosis, and lysosomal function, including CHUK (IKK-alpha), TMEM106B, MCL1, NOTCH1, and TH. In parallel, other non-modifiable risk factors (e.g., APOE genotype, sex) and biological mechanisms (e.g., amyloid deposition) selectively reshape tau propagation by shifting dominant routes between anatomical and functional pathways in a region-specific manner. Findings are validated in an independent AD cohort.

Method: Uses stepwise-decaying schedule for noise variance and learning rate, progressive gradient fusion with historical gradients, decentralized push-sum and aggregation for time-varying topologies.

Result: Achieves robust convergence with improved training stability and speed, outperforms existing baselines across multiple scenarios.

Conclusion: ADP-VRSGP effectively addresses privacy-preserving decentralized learning challenges by balancing privacy protection with model performance.

Abstract: Differential privacy is widely employed in decentralized learning to safeguard sensitive data by introducing noise into model updates. However, existing approaches that use fixed-variance noise often degrade model performance and reduce training efficiency. To address these limitations, we propose a novel approach called decentralized learning with adaptive differential privacy via variance-reduced stochastic gradient push (ADP-VRSGP). This method dynamically adjusts both the noise variance and the learning rate using a stepwise-decaying schedule, which accelerates training and enhances final model performance while providing node-level personalized privacy guarantees. To counteract the slowed convergence caused by large-variance noise in early iterations, we introduce a progressive gradient fusion strategy that leverages historical gradients. Furthermore, ADP-VRSGP incorporates decentralized push-sum and aggregation techniques, making it particularly suitable for time-varying communication topologies. Through rigorous theoretical analysis, we demonstrate that ADP-VRSGP achieves robust convergence with an appropriate learning rate, significantly improving training stability and speed. Experimental results validate that our method outperforms existing baselines across multiple scenarios, highlighting its efficacy in addressing the challenges of privacy-preserving decentralized learning.

Method: Divide TSP into clusters using sparse heatmap graph, abstract clusters as supernodes, generate hyper tour to guide initialization and optimization, focusing search on relevant edges.

Result: Outperforms existing neural-based methods on synthetic and real-world datasets, especially for larger-scale instances, with significant reduction in gap to optimal solution.

Conclusion: HyperNS provides an effective approach for large-scale TSP by combining clustering strategy with hyper tour guidance, enabling more efficient optimization and better solution quality.

Abstract: Traveling Salesman Problem (TSP) is a classic NP-hard problem that has garnered significant attention from both academia and industry. While neural-based methods have shown promise for solving TSPs, they still face challenges in scaling to larger instances, particularly in memory constraints associated with global heatmaps, edge weights, or access matrices, as well as in generating high-quality initial solutions and insufficient global guidance for efficiently navigating vast search spaces. To address these challenges, we propose a Hyper Tour Guided Neighborhood Search (HyperNS) method for large-scale TSP instances. Inspired by the ``clustering first, route second" strategy, our approach initially divides the TSP instance into clusters using a sparse heatmap graph and abstracts them as supernodes, followed by the generation of a hyper tour to guide both the initialization and optimization processes. This method reduces the search space by focusing on edges relevant to the hyper tour, leading to more efficient and effective optimization. Experimental results on both synthetic and real-world datasets demonstrate that our approach outperforms existing neural-based methods, particularly in handling larger-scale instances, offering a significant reduction in the gap to the optimal solution.

Method: Extends RL framework by incorporating explicit human value signals into reward function using exam-style data with ground-truth value labels. Uses value-weighted gradient amplification on end-of-sequence tokens.

Result: Consistently outperforms correctness-only baselines across multiple RL algorithms and model scales. Learns value-sensitive termination policies - concise for low-value prompts, thorough for high-value ones.

Conclusion: RLEV offers a practical path to aligning LLMs with human priorities by optimizing for explicit utility functions, remaining robust even under noisy value signals.

Abstract: We propose Reinforcement Learning with Explicit Human Values (RLEV), a method that aligns Large Language Model (LLM) optimization directly with quantifiable human value signals. While Reinforcement Learning with Verifiable Rewards (RLVR) effectively trains models in objective domains using binary correctness rewards, it overlooks that not all tasks are equally significant. RLEV extends this framework by incorporating human-defined value signals directly into the reward function. Using exam-style data with explicit ground-truth value labels, RLEV consistently outperforms correctness-only baselines across multiple RL algorithms and model scales. Crucially, RLEV policies not only improve value-weighted accuracy but also learn a value-sensitive termination policy: concise for low-value prompts, thorough for high-value ones. We demonstrate this behavior stems from value-weighted gradient amplification on end-of-sequence tokens. Ablation studies confirm the gain is causally linked to value alignment. RLEV remains robust under noisy value signals, such as difficulty-based labels, demonstrating that optimizing for an explicit utility function offers a practical path to aligning LLMs with human priorities.

Method: Proposes a risk-aware constrained RL framework using optimized certainty equivalents (OCEs) with per-stage robustness in reward values and time, wrapped around standard RL solvers like PPO.

Result: The framework ensures exact equivalence to the original constrained problem under appropriate constraint qualifications and demonstrates risk-aware properties through numerical experiments.

Conclusion: The proposed approach provides a practical risk-aware constrained RL framework with proven convergence and effective handling of tail risks in high-stakes applications.

Abstract: Constrained optimization provides a common framework for dealing with conflicting objectives in reinforcement learning (RL). In most of these settings, the objectives (and constraints) are expressed though the expected accumulated reward. However, this formulation neglects risky or even possibly catastrophic events at the tails of the reward distribution, and is often insufficient for high-stakes applications in which the risk involved in outliers is critical. In this work, we propose a framework for risk-aware constrained RL, which exhibits per-stage robustness properties jointly in reward values and time using optimized certainty equivalents (OCEs). Our framework ensures an exact equivalent to the original constrained problem within a parameterized strong Lagrangian duality framework under appropriate constraint qualifications, and yields a simple algorithmic recipe which can be wrapped around standard RL solvers, such as PPO. Lastly, we establish the convergence of the proposed algorithm under common assumptions, and verify the risk-aware properties of our approach through several numerical experiments.

Method: Three relaxations of replicability: (1) Pointwise replicability - consistent on fixed inputs but not across all inputs simultaneously, (2) Approximate replicability - hypotheses classify most of distribution consistently, (3) Semi-replicability - fully replicable but can use shared unlabeled samples.

Result: For constant replicability parameters, sample-optimal agnostic PAC learners are obtained: (1) and (2) require Θ(d/α²) samples, while (3) requires Θ(d²/α²) labeled samples.

Conclusion: Relaxed notions of replicability enable learning where full replicability is impossible, with different sample complexities depending on the relaxation type.

Abstract: Replicability, introduced by (Impagliazzo et al. STOC ‘22), is the notion that algorithms should remain stable under a resampling of their inputs (given access to shared randomness). While a strong and interesting notion of stability, the cost of replicability can be prohibitive: there is no replicable algorithm, for instance, for tasks as simple as threshold learning (Bun et al. STOC ‘23). Given such strong impossibility results we ask: under what approximate notions of replicability is learning possible? In this work, we propose three natural relaxations of replicability in the context of PAC learning: (1) Pointwise: the learner must be consistent on any fixed input, but not across all inputs simultaneously, (2) Approximate: the learner must output hypotheses that classify most of the distribution consistently, (3) Semi: the algorithm is fully replicable, but may additionally use shared unlabeled samples. In all three cases, for constant replicability parameters, we obtain sample-optimal agnostic PAC learners: (1) and (2) are achievable for ``free" using $\Theta(d/\alpha^2)$ samples, while (3) requires $\Theta(d^2/\alpha^2)$ labeled samples.

Method: Comprehensive benchmark evaluation of 126 analytical pipelines using machine learning models, feature selection methods, and data balancing techniques on Golden Retriever Lifetime Study data, with patient-level data partitioning to prevent leakage.

Result: Optimal model (Logistic Regression with class weighting and recursive feature elimination) showed moderate AUROC (0.815) but poor clinical performance (F1-score 0.25, PPV 0.15). High NPV (0.98) but insufficient recall (0.79) for reliable rule-out testing. Predictions driven by non-specific features like age and inflammation markers.

Conclusion: Statistically detectable cancer signal exists in routine lab data but is too weak and confounded for clinically reliable discrimination from normal aging or inflammatory conditions. Multi-modal data integration is needed for meaningful progress in computational veterinary oncology.

Abstract: The development of accessible screening tools for early cancer detection in dogs represents a significant challenge in veterinary medicine. Routine laboratory data offer a promising, low-cost source for such tools, but their utility is hampered by the non-specificity of individual biomarkers and the severe class imbalance inherent in screening populations. This study assesses the feasibility of cancer risk classification using the Golden Retriever Lifetime Study (GRLS) cohort under real-world constraints, including the grouping of diverse cancer types and the inclusion of post-diagnosis samples. A comprehensive benchmark evaluation was conducted, systematically comparing 126 analytical pipelines that comprised various machine learning models, feature selection methods, and data balancing techniques. Data were partitioned at the patient level to prevent leakage. The optimal model, a Logistic Regression classifier with class weighting and recursive feature elimination, demonstrated moderate ranking ability (AUROC = 0.815; 95% CI: 0.793-0.836) but poor clinical classification performance (F1-score = 0.25, Positive Predictive Value = 0.15). While a high Negative Predictive Value (0.98) was achieved, insufficient recall (0.79) precludes its use as a reliable rule-out test. Interpretability analysis with SHapley Additive exPlanations (SHAP) revealed that predictions were driven by non-specific features like age and markers of inflammation and anemia. It is concluded that while a statistically detectable cancer signal exists in routine lab data, it is too weak and confounded for clinically reliable discrimination from normal aging or other inflammatory conditions. This work establishes a critical performance ceiling for this data modality in isolation and underscores that meaningful progress in computational veterinary oncology will require integration of multi-modal data sources.

Method: CO-PFL performs joint assessment using gradient direction discrepancies and prediction deviations from dual subspaces, integrates parameter-wise personalization with mask-aware momentum optimization for stable updates.

Result: Extensive experiments on CIFAR10, CIFAR10C, CINIC10, and Mini-ImageNet show CO-PFL consistently outperforms state-of-the-art methods in personalization accuracy, robustness, scalability and convergence stability.

Conclusion: CO-PFL effectively mitigates aggregation bias, strengthens global coordination, and enhances local performance by building tailored submodels with stable updates through principled client contribution estimation.

Abstract: Personalized federated learning (PFL) addresses a critical challenge of collaboratively training customized models for clients with heterogeneous and scarce local data. Conventional federated learning, which relies on a single consensus model, proves inadequate under such data heterogeneity. Its standard aggregation method of weighting client updates heuristically or by data volume, operates under an equal-contribution assumption, failing to account for the actual utility and reliability of each client’s update. This often results in suboptimal personalization and aggregation bias. To overcome these limitations, we introduce Contribution-Oriented PFL (CO-PFL), a novel algorithm that dynamically estimates each client’s contribution for global aggregation. CO-PFL performs a joint assessment by analyzing both gradient direction discrepancies and prediction deviations, leveraging information from gradient and data subspaces. This dual-subspace analysis provides a principled and discriminative aggregation weight for each client, emphasizing high-quality updates. Furthermore, to bolster personalization adaptability and optimization stability, CO-PFL cohesively integrates a parameter-wise personalization mechanism with mask-aware momentum optimization. Our approach effectively mitigates aggregation bias, strengthens global coordination, and enhances local performance by facilitating the construction of tailored submodels with stable updates. Extensive experiments on four benchmark datasets (CIFAR10, CIFAR10C, CINIC10, and Mini-ImageNet) confirm that CO-PFL consistently surpasses state-of-the-art methods in in personalization accuracy, robustness, scalability and convergence stability.

Method: Reformulated constrained optimization using Lagrangian method and Sherman-Morrison-Woodbury identity to create Fair-SMW algorithm. Used three Laplacian matrix alternatives with different spectral gaps to generate multiple Fair-SMW variations.

Result: Fair-SMW achieved computation time twice as fast as state-of-the-art methods and was flexible enough to achieve twice as much balance. Evaluated on real-world datasets (LastFM, FacebookNet, Deezer, German) using Stochastic Block Model.

Conclusion: The proposed Fair-SMW algorithm successfully enhances efficiency of fair spectral clustering while maintaining comparable balance, offering significant improvements in computational performance.

Abstract: Recent research has focused on mitigating algorithmic bias in clustering by incorporating fairness constraints into algorithmic design. Notions such as disparate impact, community cohesion, and cost per population have been implemented to enforce equitable outcomes. Among these, group fairness (balance) ensures that each protected group is proportionally represented within every cluster. However, incorporating balance as a metric of fairness into spectral clustering algorithms has led to computational times that can be improved. This study aims to enhance the efficiency of spectral clustering algorithms by reformulating the constrained optimization problem using a new formulation derived from the Lagrangian method and the Sherman-Morrison-Woodbury (SMW) identity, resulting in the Fair-SMW algorithm. Fair-SMW employs three alternatives to the Laplacian matrix with different spectral gaps to generate multiple variations of Fair-SMW, achieving clustering solutions with comparable balance to existing algorithms while offering improved runtime performance. We present the results of Fair-SMW, evaluated using the Stochastic Block Model (SBM) to measure both runtime efficiency and balance across real-world network datasets, including LastFM, FacebookNet, Deezer, and German. We achieve an improvement in computation time that is twice as fast as the state-of-the-art, and also flexible enough to achieve twice as much balance.

Method: Introduces QKCV (Query-Key-Category-Value) attention that adds a static categorical embedding C to the standard QKV framework, serving as a plug-in module for attention-based models.

Result: Improves forecasting accuracy across diverse real-world datasets on models like Vanilla Transformer, Informer, PatchTST, and TFT. Enables efficient fine-tuning of univariate time series foundation models by updating only the static embedding C while keeping pretrained weights fixed.

Conclusion: QKCV attention is an effective extension that enhances time series forecasting performance and provides computational efficiency in fine-tuning scenarios.

Abstract: In real-world time series forecasting tasks, category information plays a pivotal role in capturing inherent data patterns. This paper introduces QKCV (Query-Key-Category-Value) attention, an extension of the traditional QKV framework that incorporates a static categorical embedding C to emphasize category-specific information. As a versatile plug-in module, QKCV enhances the forecasting accuracy of attention-based models (e.g., Vanilla Transformer, Informer, PatchTST, TFT) across diverse real-world datasets. Furthermore, QKCV demonstrates remarkable adaptability in fine-tuning univariate time series foundation model by solely updating the static embedding C while preserving pretrained weights, thereby reducing computational overhead and achieving superior fine-tuning performance.

Method: MetaVD learns client-dependent dropout rates via a shared hypernetwork, enabling model personalization through conditional dropout posterior for both meta-learning and Bayesian FL perspectives.

Result: MetaVD demonstrated excellent classification accuracy and uncertainty calibration, especially for OOD clients, while compressing local model parameters and reducing communication costs.

Conclusion: MetaVD effectively addresses FL challenges in non-IID settings through Bayesian meta-learning with personalized dropout rates, improving performance while reducing overfitting and communication overhead.

Abstract: Federated Learning (FL) aims to train a global inference model from remotely distributed clients, gaining popularity due to its benefit of improving data privacy. However, traditional FL often faces challenges in practical applications, including model overfitting and divergent local models due to limited and non-IID data among clients. To address these issues, we introduce a novel Bayesian meta-learning approach called meta-variational dropout (MetaVD). MetaVD learns to predict client-dependent dropout rates via a shared hypernetwork, enabling effective model personalization of FL algorithms in limited non-IID data settings. We also emphasize the posterior adaptation view of meta-learning and the posterior aggregation view of Bayesian FL via the conditional dropout posterior. We conducted extensive experiments on various sparse and non-IID FL datasets. MetaVD demonstrated excellent classification accuracy and uncertainty calibration performance, especially for out-of-distribution (OOD) clients. MetaVD compresses the local model parameters needed for each client, mitigating model overfitting and reducing communication costs. Code is available at https://github.com/insujeon/MetaVD.

Method: Train a model using sparse weather stations and topography data over Japan, comparing against interpolation baseline and CorrDiff.

Result: Model achieves up to 50% better performance than both CorrDiff and baseline for temperature downscaling, and 10-20% better for wind downscaling.

Conclusion: SpLIIF effectively generates implicit neural representations for weather variable downscaling with significant improvements over existing methods.

Abstract: We introduce SpLIIF to generate implicit neural representations and enable arbitrary downscaling of weather variables. We train a model from sparse weather stations and topography over Japan and evaluate in- and out-of-distribution accuracy predicting temperature and wind, comparing it to both an interpolation baseline and CorrDiff. We find the model to be up to 50% better than both CorrDiff and the baseline at downscaling temperature, and around 10-20% better for wind.

Method: Reformulate max-min MORL as a two-player zero-sum regularized continuous game and develop an efficient mirror descent algorithm with adaptive regularization.

Result: The algorithm shows global last-iterate convergence, provides theoretical complexity bounds, and significantly outperforms previous baselines in MORL environments.

Conclusion: The proposed framework offers a provably convergent and practical solution for max-min multi-objective RL with strong theoretical foundations and empirical performance.

Abstract: In this paper, we propose a provably convergent and practical framework for multi-objective reinforcement learning with max-min criterion. From a game-theoretic perspective, we reformulate max-min multi-objective reinforcement learning as a two-player zero-sum regularized continuous game and introduce an efficient algorithm based on mirror descent. Our approach simplifies the policy update while ensuring global last-iterate convergence. We provide a comprehensive theoretical analysis on our algorithm, including iteration complexity under both exact and approximate policy evaluations, as well as sample complexity bounds. To further enhance performance, we modify the proposed algorithm with adaptive regularization. Our experiments demonstrate the convergence behavior of the proposed algorithm in tabular settings, and our implementation for deep reinforcement learning significantly outperforms previous baselines in many MORL environments.

Method: Layer-to-Layer Knowledge Mixing (LKM) - a self-knowledge distillation method that minimizes mean absolute distance between pre-existing hidden embeddings of GNN layers to aggregate multi-hop and multi-scale information.

Result: LKM reduced mean absolute error by up to 9.8% (QM9), 45.3% (MD17 Energy), and 22.9% (Chignolin) across three GNN architectures on quantum chemical and biophysical property datasets.

Conclusion: LKM significantly improves GNN accuracy for chemical property prediction without substantial increases in training and inference costs, demonstrating its potential for efficient model enhancement.

Abstract: Graph Neural Networks (GNNs) are the currently most effective methods for predicting molecular properties but there remains a need for more accurate models. GNN accuracy can be improved by increasing the model complexity but this also increases the computational cost and memory requirement during training and inference. In this study, we develop Layer-to-Layer Knowledge Mixing (LKM), a novel self-knowledge distillation method that increases the accuracy of state-of-the-art GNNs while adding negligible computational complexity during training and inference. By minimizing the mean absolute distance between pre-existing hidden embeddings of GNN layers, LKM efficiently aggregates multi-hop and multi-scale information, enabling improved representation of both local and global molecular features. We evaluated LKM using three diverse GNN architectures (DimeNet++, MXMNet, and PAMNet) using datasets of quantum chemical properties (QM9, MD17 and Chignolin). We found that the LKM method effectively reduces the mean absolute error of quantum chemical and biophysical property predictions by up to 9.8% (QM9), 45.3% (MD17 Energy), and 22.9% (Chignolin). This work demonstrates the potential of LKM to significantly improve the accuracy of GNNs for chemical property prediction without any substantial increase in training and inference cost.

Method: Finite-sample decomposition of the selective-classification gap into five sources: Bayes noise, approximation error, ranking error, statistical noise, and implementation/shift-induced slack. Validation on synthetic two-moons data and real-world vision/language benchmarks.

Result: Bayes noise and limited model capacity account for substantial gaps; only feature-aware calibrators meaningfully improve score ordering; data shift introduces separate slack requiring distributionally robust training.

Conclusion: The decomposition provides a quantitative error budget and actionable design guidelines for building selective classifiers that more closely approximate ideal oracle behavior, emphasizing the need for scoring mechanisms that can effectively reorder predictions rather than merely rescale them.

Abstract: Selective classifiers improve model reliability by abstaining on inputs the model deems uncertain. However, few practical approaches achieve the gold-standard performance of a perfect-ordering oracle that accepts examples exactly in order of correctness. Our work formalizes this shortfall as the selective-classification gap and present the first finite-sample decomposition of this gap to five distinct sources of looseness: Bayes noise, approximation error, ranking error, statistical noise, and implementation- or shift-induced slack. Crucially, our analysis reveals that monotone post-hoc calibration – often believed to strengthen selective classifiers – has limited impact on closing this gap, since it rarely alters the model’s underlying score ranking. Bridging the gap therefore requires scoring mechanisms that can effectively reorder predictions rather than merely rescale them. We validate our decomposition on synthetic two-moons data and on real-world vision and language benchmarks, isolating each error component through controlled experiments. Our results confirm that (i) Bayes noise and limited model capacity can account for substantial gaps, (ii) only richer, feature-aware calibrators meaningfully improve score ordering, and (iii) data shift introduces a separate slack that demands distributionally robust training. Together, our decomposition yields a quantitative error budget as well as actionable design guidelines that practitioners can use to build selective classifiers which approximate ideal oracle behavior more closely.

Method: FedGPS statically modifies each client’s learning objective to model global data distribution using surrogate information, and dynamically adjusts local update directions with gradient information from other clients at each round.

Result: Extensive experiments show FedGPS outperforms state-of-the-art methods across diverse heterogeneity scenarios, validating its effectiveness and robustness.

Conclusion: FedGPS provides an effective solution to FL data heterogeneity by synergistically combining statistical and gradient information, demonstrating superior performance and robustness compared to existing methods.

Abstract: Federated Learning (FL) confronts a significant challenge known as data heterogeneity, which impairs model performance and convergence. Existing methods have made notable progress in addressing this issue. However, improving performance in certain heterogeneity scenarios remains an overlooked question: \textit{How robust are these methods to deploy under diverse heterogeneity scenarios?} To answer this, we conduct comprehensive evaluations across varied heterogeneity scenarios, showing that most existing methods exhibit limited robustness. Meanwhile, insights from these experiments highlight that sharing statistical information can mitigate heterogeneity by enabling clients to update with a global perspective. Motivated by this, we propose \textbf{FedGPS} (\textbf{Fed}erated \textbf{G}oal-\textbf{P}ath \textbf{S}ynergy), a novel framework that seamlessly integrates statistical distribution and gradient information from others. Specifically, FedGPS statically modifies each client’s learning objective to implicitly model the global data distribution using surrogate information, while dynamically adjusting local update directions with gradient information from other clients at each round. Extensive experiments show that FedGPS outperforms state-of-the-art methods across diverse heterogeneity scenarios, validating its effectiveness and robustness. The code is available at: https://github.com/CUHK-AIM-Group/FedGPS.

Method: Proposes OpTI-BFM, an optimistic decision criterion that models uncertainty over reward functions and guides BFMs in data collection for task inference. Connects to upper-confidence algorithms for linear bandits.

Result: Empirical evaluation shows OpTI-BFM enables successor-features-based BFMs to identify and optimize unseen reward functions in a handful of episodes with minimal compute overhead.

Conclusion: OpTI-BFM provides an efficient approach for task inference in BFMs through environment interaction, with formal regret bounds and practical effectiveness on zero-shot benchmarks.

Abstract: Behavior Foundation Models (BFMs) are capable of retrieving high-performing policy for any reward function specified directly at test-time, commonly referred to as zero-shot reinforcement learning (RL). While this is a very efficient process in terms of compute, it can be less so in terms of data: as a standard assumption, BFMs require computing rewards over a non-negligible inference dataset, assuming either access to a functional form of rewards, or significant labeling efforts. To alleviate these limitations, we tackle the problem of task inference purely through interaction with the environment at test-time. We propose OpTI-BFM, an optimistic decision criterion that directly models uncertainty over reward functions and guides BFMs in data collection for task inference. Formally, we provide a regret bound for well-trained BFMs through a direct connection to upper-confidence algorithms for linear bandits. Empirically, we evaluate OpTI-BFM on established zero-shot benchmarks, and observe that it enables successor-features-based BFMs to identify and optimize an unseen reward function in a handful of episodes with minimal compute overhead. Code is available at https://github.com/ThomasRupf/opti-bfm.

Method: Creates impossible variants of existing benchmarks by introducing conflicts between natural-language specifications and unit tests, then measures “cheating rate” as pass rate on these impossible tasks where any pass indicates specification-violating shortcuts.

Result: Reveals fine-grained cheating behaviors from simple test modification to complex operator overloading, shows how prompt engineering and test access affect cheating rates, and provides a testbed for developing monitoring tools.

Conclusion: ImpossibleBench serves as a versatile framework for studying model behaviors, context engineering, and developing monitoring tools to build more robust and reliable LLM systems.

Abstract: The tendency to find and exploit “shortcuts” to complete tasks poses significant risks for reliable assessment and deployment of large language models (LLMs). For example, an LLM agent with access to unit tests may delete failing tests rather than fix the underlying bug. Such behavior undermines both the validity of benchmark results and the reliability of real-world LLM coding assistant deployments. To quantify, study, and mitigate such behavior, we introduce ImpossibleBench, a benchmark framework that systematically measures LLM agents’ propensity to exploit test cases. ImpossibleBench creates “impossible” variants of tasks from existing benchmarks like LiveCodeBench and SWE-bench by introducing direct conflicts between the natural-language specification and the unit tests. We measure an agent’s “cheating rate” as its pass rate on these impossible tasks, where any pass necessarily implies a specification-violating shortcut. As a practical framework, ImpossibleBench is not just an evaluation but a versatile tool. We demonstrate its utility for: (1) studying model behaviors, revealing more fine-grained details of cheating behaviors from simple test modification to complex operator overloading; (2) context engineering, showing how prompt, test access and feedback loop affect cheating rates; and (3) developing monitoring tools, providing a testbed with verified deceptive solutions. We hope ImpossibleBench serves as a useful framework for building more robust and reliable LLM systems. Our implementation can be found at https://github.com/safety-research/impossiblebench.

Method: Developed CUDA kernels optimized for Ampere GPUs using 128B-coalesced access and hierarchical shared-memory accumulation, plus a PyTorch layer with Differentiable Euler Characteristic Transform-style sigmoid relaxation.

Result: Achieved 16-2000x speedups over prior GPU implementations on synthetic grids and created a differentiable framework for learning thresholds.

Conclusion: The work enables broader adoption of topological features in deep learning through computational efficiency and differentiability, with potential extensions for batching and multi-GPU applications.

Abstract: Topological features capture global geometric structure in imaging data, but practical adoption in deep learning requires both computational efficiency and differentiability. We present optimized GPU kernels for the Euler Characteristic Curve (ECC) computation achieving 16-2000"O speedups over prior GPU implementations on synthetic grids, and introduce a differentiable PyTorch layer enabling end-to-end learning. Our CUDA kernels, optimized for Ampere GPUs use 128B-coalesced access and hierarchical shared-memory accumulation. Our PyTorch layer learns thresholds in a single direction via a Differentiable Euler Characteristic Transform-style sigmoid relaxation. We discuss downstream relevance, including applications highlighted by prior ECC work, and outline batching/multi-GPU extensions to broaden adoption.

Method: Proposed adaptive algorithm to maximize PRM scores over intractable action space, investigated PRM-guided tree search methods including Monte Carlo tree search and beam search across 23 diverse mathematical problems using Qwen2.5-Math-7B-Instruct.

Result: PRM-guided tree search showed no statistically significant improvements over BoN despite higher costs; Monte Carlo tree search and beam search outperformed other PRM-guided methods; PRMs poorly approximate state values and their reliability degrades with reasoning depth; PRMs generalize poorly out of distribution.

Conclusion: Tree search’s underperformance stems from greater reliance on unreliable PRM scores, suggesting different reward modeling is necessary before tree search can effectively enhance mathematical reasoning in LLMs.

Abstract: While chain-of-thought prompting with Best-of-N (BoN) selection has become popular for mathematical reasoning in large language models (LLMs), its linear structure fails to capture the branching and exploratory nature of complex problem-solving. In this work, we propose an adaptive algorithm to maximize process reward model (PRM) scores over the intractable action space, and investigate whether PRM-guided tree search can improve mathematical reasoning by exploring multiple partial solution paths. Across $23$ diverse mathematical problems using Qwen2.5-Math-7B-Instruct with its associated PRM as a case study, we find that: (1) PRM-guided tree search shows no statistically significant improvements over BoN despite higher costs, (2) Monte Carlo tree search and beam search outperform other PRM-guided tree search methods, (3) PRMs poorly approximate state values and their reliability degrades with reasoning depth, and (4) PRMs generalize poorly out of distribution. This underperformance stems from tree search’s greater reliance on unreliable PRM scores, suggesting different reward modeling is necessary before tree search can effectively enhance mathematical reasoning in LLMs.

Method: A synthetic data-driven evaluation paradigm with three core analytical dimensions: temporal feature decomposition and capability mapping, robustness analysis under data irregularities, and theoretical optimum benchmarking.

Result: Experiments show current deep learning models do not universally approach optimal baselines across all types of temporal features, revealing specific limitations in model capabilities.

Conclusion: SynTSBench provides an interpretable evaluation system that enables systematic assessment of fundamental modeling capabilities and direct comparison between model predictions and mathematical optima, facilitating better model selection for specific forecasting scenarios.

Abstract: Recent advances in deep learning have driven rapid progress in time series forecasting, yet many state-of-the-art models continue to struggle with robust performance in real-world applications, even when they achieve strong results on standard benchmark datasets. This persistent gap can be attributed to the black-box nature of deep learning architectures and the inherent limitations of current evaluation frameworks, which frequently lack the capacity to provide clear, quantitative insights into the specific strengths and weaknesses of different models, thereby complicating the selection of appropriate models for particular forecasting scenarios. To address these issues, we propose a synthetic data-driven evaluation paradigm, SynTSBench, that systematically assesses fundamental modeling capabilities of time series forecasting models through programmable feature configuration. Our framework isolates confounding factors and establishes an interpretable evaluation system with three core analytical dimensions: (1) temporal feature decomposition and capability mapping, which enables systematic evaluation of model capacities to learn specific pattern types; (2) robustness analysis under data irregularities, which quantifies noise tolerance thresholds and anomaly recovery capabilities; and (3) theoretical optimum benchmarking, which establishes performance boundaries for each pattern type-enabling direct comparison between model predictions and mathematical optima. Our experiments show that current deep learning models do not universally approach optimal baselines across all types of temporal features.The code is available at https://github.com/TanQitai/SynTSBench

Method: Uses KAN (Kolmogorov-Arnold Network) as an alternative to MLPs for the collaborative model, leveraging its superior interpretability and ability to mitigate catastrophic forgetting. Deployed across language, vision, and vision-language cross-modal tasks.

Result: Significantly reduces large model inference calls while maintaining near-identical task accuracy, substantially lowering computational resource consumption. KAN-based model markedly mitigates catastrophic forgetting and improves accuracy for long-tail data.

Conclusion: KCM demonstrates superior performance across all metrics compared to MLP-based small collaborative models (MCM) in large-small model collaboration frameworks.

Abstract: In recent years, Pretrained Large Models(PLMs) researchers proposed large-small model collaboration frameworks, leveraged easily trainable small models to assist large models, aim to(1) significantly reduce computational resource consumption while maintaining comparable accuracy, and (2) enhance large model performance in specialized domain tasks. However, this collaborative paradigm suffers from issues such as significant accuracy degradation, exacerbated catastrophic forgetting, and amplified hallucination problems induced by small model knowledge. To address these challenges, we propose a KAN-based Collaborative Model (KCM) as an improved approach to large-small model collaboration. The KAN utilized in KCM represents an alternative neural network architecture distinct from conventional MLPs. Compared to MLPs, KAN offers superior visualizability and interpretability while mitigating catastrophic forgetting. We deployed KCM in large-small model collaborative systems across three scenarios: language, vision, and vision-language cross-modal tasks. The experimental results demonstrate that, compared with pure large model approaches, the large-small model collaboration framework utilizing KCM as the collaborative model significantly reduces the number of large model inference calls while maintaining near-identical task accuracy, thereby substantially lowering computational resource consumption. Concurrently, the KAN-based small collaborative model markedly mitigates catastrophic forgetting, leading to significant accuracy improvements for long-tail data. The results reveal that KCM demonstrates superior performance across all metrics compared to MLP-based small collaborative models (MCM).

Method: Created CS-54k corpus from 14k CC-licensed papers using a scalable pipeline combining retrieval-augmented generation (RAG) with multi-stage quality control. Derived CS-4k as a curated benchmark and CS-50k as training data.

Result: CS-4k stratifies state-of-the-art LLMs into distinct capability tiers. Open models trained on CS-50k with supervised training and RL show substantial improvements, with 7B-scale models outperforming larger proprietary systems like GPT-4.1, GPT-4o, and Gemini 2.5 Pro.

Conclusion: Making AI models better research assistants relies more on domain-aligned training with high-quality data than on pretraining scale or general benchmark performance. The datasets are released to foster AI systems as reliable collaborators in CS research.

Abstract: As large language models (LLMs) advance, the ultimate vision for their role in science is emerging: we could build an AI collaborator to effectively assist human beings throughout the entire scientific research process. We refer to this envisioned system as ResearchGPT. Given that scientific research progresses through multiple interdependent phases, achieving this vision requires rigorous benchmarks that evaluate the end-to-end workflow rather than isolated sub-tasks. To this end, we contribute CS-54k, a high-quality corpus of scientific Q&A pairs in computer science, built from 14k CC-licensed papers. It is constructed through a scalable, paper-grounded pipeline that combines retrieval-augmented generation (RAG) with multi-stage quality control to ensure factual grounding. From this unified corpus, we derive two complementary subsets: CS-4k, a carefully curated benchmark for evaluating AI’s ability to assist scientific research, and CS-50k, a large-scale training dataset. Extensive experiments demonstrate that CS-4k stratifies state-of-the-art LLMs into distinct capability tiers. Open models trained on CS-50k with supervised training and reinforcement learning demonstrate substantial improvements. Even 7B-scale models, when properly trained, outperform many larger proprietary systems, such as GPT-4.1, GPT-4o, and Gemini 2.5 Pro. This indicates that making AI models better research assistants relies more on domain-aligned training with high-quality data than on pretraining scale or general benchmark performance. We release CS-4k and CS-50k in the hope of fostering AI systems as reliable collaborators in CS research.

Method: Identifies that causal subgraphs have smaller distributional variations across environments, formalizes this as Invariant Distribution Criterion, and develops a norm-guided invariant distribution objective for causal subgraph discovery without IRM framework.

Result: Extensive experiments on two benchmarks show the method consistently outperforms state-of-the-art methods in graph generalization tasks.

Conclusion: The proposed IRM-free approach effectively captures causal subgraphs and improves out-of-distribution generalization in graph neural networks without requiring environment annotations.

Abstract: Out-of-distribution generalization under distributional shifts remains a critical challenge for graph neural networks. Existing methods generally adopt the Invariant Risk Minimization (IRM) framework, requiring costly environment annotations or heuristically generated synthetic splits. To circumvent these limitations, in this work, we aim to develop an IRM-free method for capturing causal subgraphs. We first identify that causal subgraphs exhibit substantially smaller distributional variations than non-causal components across diverse environments, which we formalize as the Invariant Distribution Criterion and theoretically prove in this paper. Building on this criterion, we systematically uncover the quantitative relationship between distributional shift and representation norm for identifying the causal subgraph, and investigate its underlying mechanisms in depth. Finally, we propose an IRM-free method by introducing a norm-guided invariant distribution objective for causal subgraph discovery and prediction. Extensive experiments on two widely used benchmarks demonstrate that our method consistently outperforms state-of-the-art methods in graph generalization.

Method: Proposed double-backbone network integrating VGG16 and Xception with Frequency-Gated Attention (FGA) Block to capture complementary local and global features. Uses Grad-CAM for tumor region visualization and developed GUI for real-time classification.

Result: Achieved 99.24% accuracy on 7K-DS dataset for 4-class, 98.68% for 3-class, and 99.85% for 2-class settings. On independent 3K-DS dataset, generalized with 95.77% accuracy, outperforming baseline and state-of-the-art methods.

Conclusion: Augmentation-free, interpretable deep learning models like DB-FGA-Net hold strong potential for reliable clinical translation in brain tumor diagnosis, bridging the gap between model prediction and clinical interpretability.

Abstract: Brain tumors are a challenging problem in neuro-oncology, where early and precise diagnosis is important for successful treatment. Deep learning-based brain tumor classification methods often rely on heavy data augmentation which can limit generalization and trust in clinical applications. In this paper, we propose a double-backbone network integrating VGG16 and Xception with a Frequency-Gated Attention (FGA) Block to capture complementary local and global features. Unlike previous studies, our model achieves state-of-the-art performance without augmentation which demonstrates robustness to variably sized and distributed datasets. For further transparency, Grad-CAM is integrated to visualize the tumor regions based on which the model is giving prediction, bridging the gap between model prediction and clinical interpretability. The proposed framework achieves 99.24% accuracy on the 7K-DS dataset for the 4-class setting, along with 98.68% and 99.85% in the 3-class and 2-class settings, respectively. On the independent 3K-DS dataset, the model generalizes with 95.77% accuracy, outperforming baseline and state-of-the-art methods. To further support clinical usability, we developed a graphical user interface (GUI) that provides real-time classification and Grad-CAM-based tumor localization. These findings suggest that augmentation-free, interpretable, and deployable deep learning models such as DB-FGA-Net hold strong potential for reliable clinical translation in brain tumor diagnosis.

Method: InvDec uses a patch-based temporal encoder combined with an inverted decoder that operates on the variate dimension through variate-wise self-attention. It introduces delayed variate embeddings to preserve temporal integrity and an adaptive residual fusion mechanism to balance temporal and variate information.

Result: Significant performance gains on high-dimensional datasets: 20.9% MSE reduction on Electricity (321 variables), 4.3% improvement on Weather, and 2.7% gain on Traffic compared to PatchTST, while maintaining competitive performance on low-dimensional datasets.

Conclusion: InvDec effectively addresses the challenge of modeling both temporal patterns and cross-variate dependencies, with its advantage growing with dataset dimensionality, confirming that cross-variate modeling becomes increasingly critical as the number of variables increases.

Abstract: Multivariate time series forecasting requires simultaneously modeling temporal patterns and cross-variate dependencies. Channel-independent methods such as PatchTST excel at temporal modeling but ignore variable correlations, while pure variate-attention approaches such as iTransformer sacrifice temporal encoding. We proposeInvDec (Inverted Decoder), a hybrid architecture that achieves principled separation between temporal encoding and variate-level decoding. InvDec combines a patch-based temporal encoder with an inverted decoder operating on the variate dimension through variate-wise self-attention. We introduce delayed variate embeddings that enrich variable-specific representations only after temporal encoding, preserving temporal feature integrity. An adaptive residual fusion mechanism dynamically balances temporal and variate information across datasets of varying dimensions. Instantiating InvDec with PatchTST yields InvDec-PatchTST. Extensive experiments on seven benchmarks demonstrate significant gains on high-dimensional datasets: 20.9% MSE reduction on Electricity (321 variables), 4.3% improvement on Weather, and 2.7% gain on Traffic compared to PatchTST, while maintaining competitive performance on low-dimensional ETT datasets. Ablation studies validate each component, and analysis reveals that InvDec’s advantage grows with dataset dimensionality, confirming that cross-variate modeling becomes critical as the number of variables increases.

Method: LEGO divides multiple-attribute unlearning into two steps: Embedding Calibration (removes attribute information from user embeddings) and Flexible Combination (combines embeddings to protect all sensitive attributes). The process is framed as a mutual information minimization problem.

Result: Extensive experiments on three real-world datasets across three recommendation models demonstrate the effectiveness and efficiency of the proposed framework.

Conclusion: LEGO successfully addresses the limitations of single-attribute unlearning methods by providing a lightweight, efficient solution for multiple-attribute unlearning with theoretical guarantees and practical adaptability.

Abstract: With the growing demand for safeguarding sensitive user information in recommender systems, recommendation attribute unlearning is receiving increasing attention. Existing studies predominantly focus on single-attribute unlearning. However, privacy protection requirements in the real world often involve multiple sensitive attributes and are dynamic. Existing single-attribute unlearning methods cannot meet these real-world requirements due to i) CH1: the inability to handle multiple unlearning requests simultaneously, and ii) CH2: the lack of efficient adaptability to dynamic unlearning needs. To address these challenges, we propose LEGO, a lightweight and efficient multiple-attribute unlearning framework. Specifically, we divide the multiple-attribute unlearning process into two steps: i) Embedding Calibration removes information related to a specific attribute from user embedding, and ii) Flexible Combination combines these embeddings into a single embedding, protecting all sensitive attributes. We frame the unlearning process as a mutual information minimization problem, providing LEGO a theoretical guarantee of simultaneous unlearning, thereby addressing CH1. With the two-step framework, where Embedding Calibration can be performed in parallel and Flexible Combination is flexible and efficient, we address CH2. Extensive experiments on three real-world datasets across three representative recommendation models demonstrate the effectiveness and efficiency of our proposed framework. Our code and appendix are available at https://github.com/anonymifish/lego-rec-multiple-attribute-unlearning.

Method: Proposed an image generation approach using commercial flight simulator to complement few real annotated images, with customized domain adaptation strategy.

Result: Standard object detection models achieved accurate prediction and showed robustness to adverse conditions like nighttime images not present in real data.

Conclusion: Synthetic data generation with controlled domain adaptation is effective for runway detection in autonomous landing systems, enabling robust performance in challenging conditions.

Abstract: Deep vision models are now mature enough to be integrated in industrial and possibly critical applications such as autonomous navigation. Yet, data collection and labeling to train such models requires too much efforts and costs for a single company or product. This drawback is more significant in critical applications, where training data must include all possible conditions including rare scenarios. In this perspective, generating synthetic images is an appealing solution, since it allows a cheap yet reliable covering of all the conditions and environments, if the impact of the synthetic-to-real distribution shift is mitigated. In this article, we consider the case of runway detection that is a critical part in autonomous landing systems developed by aircraft manufacturers. We propose an image generation approach based on a commercial flight simulator that complements a few annotated real images. By controlling the image generation and the integration of real and synthetic data, we show that standard object detection models can achieve accurate prediction. We also evaluate their robustness with respect to adverse conditions, in our case nighttime images, that were not represented in the real data, and show the interest of using a customized domain adaptation strategy.

Method: Proposes uncertainty-based routing that formulates advantage estimation as pairwise preference classification, using uncertainty quantification to route uncertain pairs to LLM judge and confident ones to RM.

Result: Significantly outperforms random judge calling at same cost on RM benchmarks, and shows effectiveness in improving online RLHF in downstream alignment.

Conclusion: The uncertainty-based routing framework efficiently combines the strengths of both RM and LLM judge, enabling better RLHF performance while managing computational costs.

Abstract: Reward model (RM) plays a pivotal role in reinforcement learning with human feedback (RLHF) for aligning large language models (LLMs). However, classical RMs trained on human preferences are vulnerable to reward hacking and generalize poorly to out-of-distribution (OOD) inputs. By contrast, strong LLM judges equipped with reasoning capabilities demonstrate superior generalization, even without additional training, but incur significantly higher inference costs, limiting their applicability in online RLHF. In this work, we propose an uncertainty-based routing framework that efficiently complements a fast RM with a strong but costly LLM judge. Our approach formulates advantage estimation in policy gradient (PG) methods as pairwise preference classification, enabling principled uncertainty quantification to guide routing. Uncertain pairs are forwarded to the LLM judge, while confident ones are evaluated by the RM. Experiments on RM benchmarks demonstrate that our uncertainty-based routing strategy significantly outperforms random judge calling at the same cost, and downstream alignment results showcase its effectiveness in improving online RLHF.

Method: Introduces an uncertainty set for the estimated covariance matrix and formulates a reconciliation problem that minimizes worst-case expected squared error, cast as a semidefinite optimization problem.

Result: Numerical experiments show the proposed robust reconciliation method achieved better forecast performance than existing hierarchical forecasting methods.

Conclusion: Integrating uncertainty into the reconciliation process through robust optimization is effective for improving hierarchical forecasting performance.

Abstract: This paper focuses on forecasting hierarchical time-series data, where each higher-level observation equals the sum of its corresponding lower-level time series. In such contexts, the forecast values should be coherent, meaning that the forecast value of each parent series exactly matches the sum of the forecast values of its child series. Existing hierarchical forecasting methods typically generate base forecasts independently for each series and then apply a reconciliation procedure to adjust them so that the resulting forecast values are coherent across the hierarchy. These methods generally derive an optimal reconciliation, using a covariance matrix of the forecast error. In practice, however, the true covariance matrix is unknown and has to be estimated from finite samples in advance. This gap between the true and estimated covariance matrix may degrade forecast performance. To address this issue, we propose a robust optimization framework for hierarchical reconciliation that accounts for uncertainty in the estimated covariance matrix. We first introduce an uncertainty set for the estimated covariance matrix and formulate a reconciliation problem that minimizes the worst-case expected squared error over this uncertainty set. We show that our problem can be cast as a semidefinite optimization problem. Numerical experiments demonstrate that the proposed robust reconciliation method achieved better forecast performance than existing hierarchical forecasting methods, which indicates the effectiveness of integrating uncertainty into the reconciliation process.

Method: Proposed Relative-Based Probability (RBP) metric that quantifies probability of correct token being ranked among top predictions, then established Relative-Based Scaling Law to characterize RBP improvement with model size scaling.

Result: Through experiments on four datasets and four model families spanning five orders of magnitude, demonstrated robustness and accuracy of the Relative-Based Scaling Law.

Conclusion: The Relative-Based Scaling Law complements cross-entropy perspective, provides deeper explanation of emergence phenomena, facilitates finding fundamental scaling theories, and contributes to more complete understanding of scaling large language models.

Abstract: Scaling laws aim to accurately predict model performance across different scales. Existing scaling-law studies almost exclusively rely on cross-entropy as the evaluation metric. However, cross-entropy provides only a partial view of performance: it measures the absolute probability assigned to the correct token, but ignores the relative ordering between correct and incorrect tokens. Yet, relative ordering is crucial for language models, such as in greedy-sampling scenario. To address this limitation, we investigate scaling from the perspective of relative ordering. We first propose the Relative-Based Probability (RBP) metric, which quantifies the probability that the correct token is ranked among the top predictions. Building on this metric, we establish the Relative-Based Scaling Law, which characterizes how RBP improves with increasing model size. Through extensive experiments on four datasets and four model families spanning five orders of magnitude, we demonstrate the robustness and accuracy of this law. Finally, we illustrate the broad application of this law with two examples, namely providing a deeper explanation of emergence phenomena and facilitating finding fundamental theories of scaling laws. In summary, the Relative-Based Scaling Law complements the cross-entropy perspective and contributes to a more complete understanding of scaling large language models. Thus, it offers valuable insights for both practical development and theoretical exploration.

Method: The paper analyzes DPO’s statistical estimation problem and RLHF’s local behavior, then proposes AuxDPO which introduces auxiliary variables in the DPO loss function to better approximate RLHF solutions and mitigate misspecification.

Result: Empirical results show AuxDPO achieves superior performance compared to standard DPO in both didactic bandit settings and LLM alignment tasks, effectively mitigating DPO’s misspecification issues.

Conclusion: AuxDPO provides a principled approach to address DPO’s misspecification problems by incorporating auxiliary variables, offering better alignment with RLHF solutions and improved performance in practice.

Abstract: Direct alignment algorithms such as Direct Preference Optimization (DPO) fine-tune models based on preference data, using only supervised learning instead of two-stage reinforcement learning with human feedback (RLHF). We show that DPO encodes a statistical estimation problem over reward functions induced by a parametric policy class. When the true reward function that generates preferences cannot be realized via the policy class, DPO becomes misspecified, resulting in failure modes such as preference order reversal, worsening of policy reward, and high sensitivity to the input preference data distribution. On the other hand, we study the local behavior of two-stage RLHF for a parametric class and relate it to a natural gradient step in policy space. Our fine-grained geometric characterization allows us to propose AuxDPO, which introduces additional auxiliary variables in the DPO loss function to help move towards the RLHF solution in a principled manner and mitigate the misspecification in DPO. We empirically demonstrate the superior performance of AuxDPO on didactic bandit settings as well as LLM alignment tasks.

Method: Proposes a thresholding method that learns thresholds to tune mark probabilities normalized by prior probabilities, predicts mark first then time, and uses a neural MTPP model for efficient time sampling without expensive numerical integration.

Result: Extensive experiments on real-world datasets show superior performance for both mark and time prediction compared to various baselines.

Conclusion: The proposed solution effectively addresses the class imbalance problem in MTPP modeling and achieves better prediction performance, especially for rare marks.

Abstract: Marked Temporal Point Process (MTPP) has been well studied to model the event distribution in marked event streams, which can be used to predict the mark and arrival time of the next event. However, existing studies overlook that the distribution of event marks is highly imbalanced in many real-world applications, with some marks being frequent but others rare. The imbalance poses a significant challenge to the performance of the next event prediction, especially for events of rare marks. To address this issue, we propose a thresholding method, which learns thresholds to tune the mark probability normalized by the mark’s prior probability to optimize mark prediction, rather than predicting the mark directly based on the mark probability as in existing studies. In conjunction with this method, we predict the mark first and then the time. In particular, we develop a novel neural MTPP model to support effective time sampling and estimation of mark probability without computationally expensive numerical improper integration. Extensive experiments on real-world datasets demonstrate the superior performance of our solution against various baselines for the next event mark and time prediction. The code is available at https://github.com/undes1red/IFNMTPP.

Method: Evaluated five selection methods: random sampling, K-Center, gradient-norm-based selection (GraNd), stratified coverage (CCS), and proposed SAPS. Assessed repair effectiveness through toxicity reduction, perplexity metrics, and composite scores (RPS, OPS, RES).

Result: SAPS achieves best balance between detoxification, utility preservation, and efficiency. Random sampling effective for large/robust models. High-overhead methods like CCS and GraNd provide limited benefit. Optimal data proportion depends on model scale and repair method.

Conclusion: Sample selection should be a tunable component of repair pipelines, establishing selection-based repair as an efficient and scalable paradigm for maintaining LLM reliability.

Abstract: Large language models (LLMs) are increasingly deployed in real-world systems, yet they can produce toxic or biased outputs that undermine safety and trust. Post-hoc model repair provides a practical remedy, but the high cost of parameter updates motivates selective use of repair data. Despite extensive prior work on data selection for model training, it remains unclear which sampling criteria are most effective and efficient when applied specifically to behavioral repair of large generative models. Our study presents a systematic analysis of sample prioritization strategies for LLM repair. We evaluate five representative selection methods, including random sampling, K-Center, gradient-norm-based selection(GraNd), stratified coverage (CCS), and a Semantic-Aware Prioritized Sampling (SAPS) approach we proposed. Repair effectiveness and trade-offs are assessed through toxicity reduction, perplexity on WikiText-2 and LAMBADA, and three composite metrics: the Repair Proximity Score (RPS), the Overall Performance Score (OPS), and the Repair Efficiency Score (RES). Experimental results show that SAPS achieves the best balance between detoxification, utility preservation, and efficiency, delivering comparable or superior repair outcomes with substantially less data. Random sampling remains effective for large or robust models, while high-overhead methods such as CCS and GraNd provide limited benefit. The optimal data proportion depends on model scale and repair method, indicating that sample selection should be regarded as a tunable component of repair pipelines. Overall, these findings establish selection-based repair as an efficient and scalable paradigm for maintaining LLM reliability.

Method: The authors propose explainable benchmarking using PruneCEL, a novel concept learning approach developed specifically for large knowledge graphs, to compute performance explanations.

Result: PruneCEL outperforms state-of-the-art concept learners by up to 0.55 F1 points, and a user study with 41 participants shows 80% accuracy in predicting system behavior based on the generated explanations.

Conclusion: Explainable benchmarking is a viable approach that enables automatic generation of meaningful explanations for system performance, with PruneCEL demonstrating strong performance and practical utility in user studies.

Abstract: Evaluating competing systems in a comparable way, i.e., benchmarking them, is an undeniable pillar of the scientific method. However, system performance is often summarized via a small number of metrics. The analysis of the evaluation details and the derivation of insights for further development or use remains a tedious manual task with often biased results. Thus, this paper argues for a new type of benchmarking, which is dubbed explainable benchmarking. The aim of explainable benchmarking approaches is to automatically generate explanations for the performance of systems in a benchmark. We provide a first instantiation of this paradigm for knowledge-graph-based question answering systems. We compute explanations by using a novel concept learning approach developed for large knowledge graphs called PruneCEL. Our evaluation shows that PruneCEL outperforms state-of-the-art concept learners on the task of explainable benchmarking by up to 0.55 points F1 measure. A task-driven user study with 41 participants shows that in 80% of the cases, the majority of participants can accurately predict the behavior of a system based on our explanations. Our code and data are available at https://github.com/dice-group/PruneCEL/tree/K-cap2025

Method: Constructs joint graphs integrating atomic structures of drug pairs, uses structure consistency module to iteratively refine node features while preserving global structural context, enabling learning of both local and global interaction patterns.

Result: Outperforms state-of-the-art baselines across two benchmark datasets, achieving superior performance in long-tail and inductive scenarios with robust representations across frequent and rare DDI types.

Conclusion: Fine-grained graph refinement improves accuracy, robustness, and mechanistic interpretability of DDI event prediction, contributing to graph-based methods for mining drug-drug interaction networks.

Abstract: Drug combinations offer therapeutic benefits but also carry the risk of adverse drug-drug interactions (DDIs), especially under complex molecular structures. Accurate DDI event prediction requires capturing fine-grained inter-drug relationships, which are critical for modeling metabolic mechanisms such as enzyme-mediated competition. However, existing approaches typically rely on isolated drug representations and fail to explicitly model atom-level cross-molecular interactions, limiting their effectiveness across diverse molecular complexities and DDI type distributions. To address these limitations, we propose MolBridge, a novel atom-level joint graph refinement framework for robust DDI event prediction. MolBridge constructs a joint graph that integrates atomic structures of drug pairs, enabling direct modeling of inter-drug associations. A central challenge in such joint graph settings is the potential loss of information caused by over-smoothing when modeling long-range atomic dependencies. To overcome this, we introduce a structure consistency module that iteratively refines node features while preserving the global structural context. This joint design allows MolBridge to effectively learn both local and global interaction outperforms state-of-the-art baselines, achieving superior performance across long-tail and inductive scenarios. patterns, yielding robust representations across both frequent and rare DDI types. Extensive experiments on two benchmark datasets show that MolBridge consistently. These results demonstrate the advantages of fine-grained graph refinement in improving the accuracy, robustness, and mechanistic interpretability of DDI event prediction.This work contributes to Web Mining and Content Analysis by developing graph-based methods for mining and analyzing drug-drug interaction networks.

Method: Graph neural network approach using temporal directed graphs with players as nodes and historical match outcomes as directed edges, explicitly modeling intransitive relationships.

Result: Model achieved 65.7% accuracy and 0.215 Brier Score; when selectively betting on high intransitivity matchups, achieved 3.26% ROI with Kelly staking over 1903 bets.

Conclusion: The graph-based approach successfully captures relational dynamics in intransitive scenarios and exploits market inefficiencies, demonstrating significant positive returns in tennis betting markets.

Abstract: Intransitive player dominance, where player A beats B, B beats C, but C beats A, is common in competitive tennis. Yet, there are few known attempts to incorporate it within forecasting methods. We address this problem with a graph neural network approach that explicitly models these intransitive relationships through temporal directed graphs, with players as nodes and their historical match outcomes as directed edges. We find the bookmaker Pinnacle Sports poorly handles matches with high intransitive complexity and posit that our graph-based approach is uniquely positioned to capture relational dynamics in these scenarios. When selectively betting on higher intransitivity matchups with our model (65.7% accuracy, 0.215 Brier Score), we achieve significant positive returns of 3.26% ROI with Kelly staking over 1903 bets, suggesting a market inefficiency in handling intransitive matchups that our approach successfully exploits.

Method: Leverages textual triggers to poison training data, covertly implanting backdoors that induce attacker-specified subgraphs during inference when triggers appear, while preserving normal performance on clean inputs.

Result: Extensive experiments on four benchmark datasets show high effectiveness and stealth: less than 10% poisoning rate achieves 50% attack success rate, while 24% suffices for over 80% success rate, with negligible performance degradation on benign samples.

Conclusion: The findings reveal security vulnerabilities in latent diffusion models for text-guided graph generation, highlight serious risks in applications like drug discovery, and underscore the need for robust defenses against such backdoor attacks.

Abstract: The rapid progress of graph generation has raised new security concerns, particularly regarding backdoor vulnerabilities. While prior work has explored backdoor attacks in image diffusion and unconditional graph generation, conditional, especially text-guided graph generation remains largely unexamined. This paper proposes BadGraph, a backdoor attack method targeting latent diffusion models for text-guided graph generation. BadGraph leverages textual triggers to poison training data, covertly implanting backdoors that induce attacker-specified subgraphs during inference when triggers appear, while preserving normal performance on clean inputs. Extensive experiments on four benchmark datasets (PubChem, ChEBI-20, PCDes, MoMu) demonstrate the effectiveness and stealth of the attack: less than 10% poisoning rate can achieves 50% attack success rate, while 24% suffices for over 80% success rate, with negligible performance degradation on benign samples. Ablation studies further reveal that the backdoor is implanted during VAE and diffusion training rather than pretraining. These findings reveal the security vulnerabilities in latent diffusion models of text-guided graph generation, highlight the serious risks in models’ applications such as drug discovery and underscore the need for robust defenses against the backdoor attack in such diffusion models.

Method: The authors introduce a preference model trained with ranking loss on procedurally generated images to detect watermarks, then use backpropagation optimization to remove and forge watermarks using only a single watermarked image.

Result: The proposed method effectively forges watermarks across various post-hoc image watermarking models, demonstrating vulnerabilities in current watermarking security approaches.

Conclusion: Current watermarking approaches have security vulnerabilities as the proposed attack can successfully forge watermarks, questioning the reliability of existing watermarking systems.

Abstract: Recent years have seen a surge in interest in digital content watermarking techniques, driven by the proliferation of generative models and increased legal pressure. With an ever-growing percentage of AI-generated content available online, watermarking plays an increasingly important role in ensuring content authenticity and attribution at scale. There have been many works assessing the robustness of watermarking to removal attacks, yet, watermark forging, the scenario when a watermark is stolen from genuine content and applied to malicious content, remains underexplored. In this work, we investigate watermark forging in the context of widely used post-hoc image watermarking. Our contributions are as follows. First, we introduce a preference model to assess whether an image is watermarked. The model is trained using a ranking loss on purely procedurally generated images without any need for real watermarks. Second, we demonstrate the model’s capability to remove and forge watermarks by optimizing the input image through backpropagation. This technique requires only a single watermarked image and works without knowledge of the watermarking model, making our attack much simpler and more practical than attacks introduced in related work. Third, we evaluate our proposed method on a variety of post-hoc image watermarking models, demonstrating that our approach can effectively forge watermarks, questioning the security of current watermarking approaches. Our code and further resources are publicly available.

Method: Uses hurdle model for zero inflation (non-rain vs rain) and IMDL (Inverse Model Debiasing Learning) for long tail problem by transforming learning objective to unbiased inverse model.

Result: Superior performance over conventional, cost-sensitive, generative, and multi-task learning methods, with reduced systematic underestimation and improved heavy-to-extreme rain retrieval.

Conclusion: Hurdle-IMDL provides generalizable approach for handling imbalanced environmental variable distributions, enabling better detection of rare but high-impact events.

Abstract: Artificial intelligence has advanced quantitative remote sensing, yet its effectiveness is constrained by imbalanced label distribution. This imbalance leads conventionally trained models to favor common samples, which in turn degrades retrieval performance for rare ones. Rainfall retrieval exemplifies this issue, with performance particularly compromised for heavy rain. This study proposes Hurdle-Inversion Model Debiasing Learning (IMDL) framework. Following a divide-and-conquer strategy, imbalance in the rain distribution is decomposed into two components: zero inflation, defined by the predominance of non-rain samples; and long tail, defined by the disproportionate abundance of light-rain samples relative to heavy-rain samples. A hurdle model is adopted to handle the zero inflation, while IMDL is proposed to address the long tail by transforming the learning object into an unbiased ideal inverse model. Comprehensive evaluation via statistical metrics and case studies investigating rainy weather in eastern China confirms Hurdle-IMDL’s superiority over conventional, cost-sensitive, generative, and multi-task learning methods. Its key advancements include effective mitigation of systematic underestimation and a marked improvement in the retrieval of heavy-to-extreme rain. IMDL offers a generalizable approach for addressing imbalance in distributions of environmental variables, enabling enhanced retrieval of rare yet high-impact events.

Method: Uses gradient of matrix singular values to identify key layers, allows clustering matrix rows around multiple subspaces for better factorization, and evaluates on only 100 samples instead of full dataset.

Result: Achieves up to 24.6 percentage points accuracy improvement, eliminates layer-by-layer sweep, reduces search time significantly, and enables adaptation without fine-tuning using minimal data.

Conclusion: Combining gradient-based layer selection, multi-subspace factorization, and minimal data evaluation yields a fast and robust adaptation algorithm for LLMs on downstream tasks without fine-tuning.

Abstract: Recently, Sharma et al. suggested a method called Layer-SElective-Rank reduction (LASER) which demonstrated that pruning high-order components of carefully chosen LLM’s weight matrices can boost downstream accuracy – without any gradient-based fine-tuning. Yet LASER’s exhaustive, per-matrix search (each requiring full-dataset forward passes) makes it impractical for rapid deployment. We demonstrate that this overhead can be removed and find that: (i) Only a small, carefully chosen subset of matrices needs to be inspected – eliminating the layer-by-layer sweep, (ii) The gradient of each matrix’s singular values pinpoints which matrices merit reduction, (iii) Increasing the factorization search space by allowing matrices rows to cluster around multiple subspaces and then decomposing each cluster separately further reduces overfitting on the original training data and further lifts accuracy by up to 24.6 percentage points, and finally, (iv) we discover that evaluating on just 100 samples rather than the full training data – both for computing the indicative gradients and for measuring the final accuracy – suffices to further reduce the search time; we explain that as adaptation to downstream tasks is dominated by prompting style, not dataset size. As a result, we show that combining these findings yields a fast and robust adaptation algorithm for downstream tasks. Overall, with a single gradient step on 100 examples and a quick scan of the top candidate layers and factorization techniques, we can adapt LLMs to new datasets – entirely without fine-tuning.

Method: Bi-CoG assigns high-quality pseudo-labels by exploiting both inter-model and intra-model consistency, along with an error-aware dynamic pseudo-label assignment strategy.

Result: Extensive experiments on 14 datasets show Bi-CoG consistently and significantly improves performance of existing methods.

Conclusion: Bi-CoG is an effective plug-and-play methodology that addresses limitations of current semi-supervised fine-tuning approaches through bi-consistency guidance and dynamic pseudo-label assignment.

Abstract: Exploiting unlabeled data through semi-supervised learning (SSL) or leveraging pre-trained models via fine-tuning are two prevailing paradigms for addressing label-scarce scenarios. Recently, growing attention has been given to combining fine-tuning of pre-trained vision-language models (VLMs) with SSL, forming the emerging paradigm of semi-supervised fine-tuning. However, existing methods often suffer from model bias and hyperparameter sensitivity, due to reliance on prediction consistency or pre-defined confidence thresholds. To address these limitations, we propose a simple yet effective plug-and-play methodology named $\underline{\textbf{Bi-Co}}$nsistency-$\underline{\textbf{G}}$uided Self-Training (Bi-CoG), which assigns high-quality and low-bias pseudo-labels, by simultaneously exploiting inter-model and intra-model consistency, along with an error-aware dynamic pseudo-label assignment strategy. Both theoretical analysis and extensive experiments over 14 datasets demonstrate the effectiveness of Bi-CoG, which consistently and significantly improves the performance of existing methods.

Method: Study seven diverse rewiring strategies and correlate changes in local and global graph properties with node classification accuracy through systematic analysis.

Result: Reveals consistent pattern that successful rewiring methods preserve local structure while allowing flexibility in global connectivity.

Conclusion: Findings provide insights for designing effective rewiring strategies, bridging graph theory and practical GNN optimization.

Abstract: Graph rewiring has emerged as a key technique to alleviate over-squashing in Graph Neural Networks (GNNs) and Graph Transformers by modifying the graph topology to improve information flow. While effective, rewiring inherently alters the graph’s structure, raising the risk of distorting important topology-dependent signals. Yet, despite the growing use of rewiring, little is known about which structural properties must be preserved to ensure both performance gains and structural fidelity. In this work, we provide the first systematic analysis of how rewiring affects a range of graph structural metrics, and how these changes relate to downstream task performance. We study seven diverse rewiring strategies and correlate changes in local and global graph properties with node classification accuracy. Our results reveal a consistent pattern: successful rewiring methods tend to preserve local structure while allowing for flexibility in global connectivity. These findings offer new insights into the design of effective rewiring strategies, bridging the gap between graph theory and practical GNN optimization.

Method: Uses sheaf-theoretic framework to replace single-space alignment with multiple comparison spaces, models pairwise modality relations through sheaf structures, and employs decentralized contrastive learning-based objectives.

Result: Superior zero-shot generalization, cross-modal alignment, and robustness to missing modalities, with 50% lower communication cost than state-of-the-art baselines.

Conclusion: SheafAlign effectively overcomes limitations of prior methods by preserving both shared and unique information across modalities without requiring mutual redundancy.

Abstract: Conventional multimodal alignment methods assume mutual redundancy across all modalities, an assumption that fails in real-world distributed scenarios. We propose SheafAlign, a sheaf-theoretic framework for decentralized multimodal alignment that replaces single-space alignment with multiple comparison spaces. This approach models pairwise modality relations through sheaf structures and leverages decentralized contrastive learning-based objectives for training. SheafAlign overcomes the limitations of prior methods by not requiring mutual redundancy among all modalities, preserving both shared and unique information. Experiments on multimodal sensing datasets show superior zero-shot generalization, cross-modal alignment, and robustness to missing modalities, with 50% lower communication cost than state-of-the-art baselines.

Method: The authors develop a unified framework with consistent notation and taxonomy, classifying algorithms into two families: direct representations (end-to-end mappings from rewards to policies) and compositional representations (decomposing representations using value function substructure).

Result: The framework enables direct comparison between methods, highlights shared principles and key differences, and derives an extended bound for successor-feature methods in the zero-shot regime.

Conclusion: By consolidating existing work under a common lens, this framework provides a principled foundation for future zero-shot RL research and outlines a path toward developing more general agents.

Abstract: Zero-shot reinforcement learning (RL) has emerged as a setting for developing general agents in an unsupervised manner, capable of solving downstream tasks without additional training or planning at test-time. Unlike conventional RL, which optimizes policies for a fixed reward, zero-shot RL requires agents to encode representations rich enough to support immediate adaptation to any objective, drawing parallels to vision and language foundation models. Despite growing interest, the field lacks a common analytical lens. We present the first unified framework for zero-shot RL. Our formulation introduces a consistent notation and taxonomy that organizes existing approaches and allows direct comparison between them. Central to our framework is the classification of algorithms into two families: direct representations, which learn end-to-end mappings from rewards to policies, and compositional representations, which decompose the representation leveraging the substructure of the value function. Within this framework, we highlight shared principles and key differences across methods, and we derive an extended bound for successor-feature methods, offering a new perspective on their performance in the zero-shot regime. By consolidating existing work under a common lens, our framework provides a principled foundation for future research in zero-shot RL and outlines a clear path toward developing more general agents.

Method: Learn energy landscapes over solution spaces of smaller subproblems, then combine them to construct global energy landscapes for complex problems. Use Parallel Energy Minimization (PEM) to improve sample quality.

Result: Outperforms state-of-the-art methods on various reasoning problems, demonstrating ability to generalize to larger and more complex problems.

Conclusion: The compositional energy landscape approach enables better reasoning generalization by breaking down complex problems into tractable subproblems and combining their solutions.

Abstract: Generalization is a key challenge in machine learning, specifically in reasoning tasks, where models are expected to solve problems more complex than those encountered during training. Existing approaches typically train reasoning models in an end-to-end fashion, directly mapping input instances to solutions. While this allows models to learn useful heuristics from data, it often results in limited generalization beyond the training distribution. In this work, we propose a novel approach to reasoning generalization by learning energy landscapes over the solution spaces of smaller, more tractable subproblems. At test time, we construct a global energy landscape for a given problem by combining the energy functions of multiple subproblems. This compositional approach enables the incorporation of additional constraints during inference, allowing the construction of energy landscapes for problems of increasing difficulty. To improve the sample quality from this newly constructed energy landscape, we introduce Parallel Energy Minimization (PEM). We evaluate our approach on a wide set of reasoning problems. Our method outperforms existing state-of-the-art methods, demonstrating its ability to generalize to larger and more complex problems. Project website can be found at: https://alexoarga.github.io/compositional_reasoning/

Method: The paper evaluates RAMI 4.0 and ISA-95 reference models for MLOps integration suitability, provides detailed mapping of MLOps lifecycle components to RAMI 4.0, and demonstrates this through a real-world use case.

Result: Findings show that standard MLOps practices cannot be directly applied to OT environments, but structured adaptation using existing reference models provides a viable pathway for successful integration.

Conclusion: Existing OT reference models like RAMI 4.0 and ISA-95 can be effectively leveraged to systematically embed MLOps practices into OT environments through structured adaptation rather than direct transplantation.

Abstract: Machine Learning Operations (MLOps) practices are increas- ingly adopted in industrial settings, yet their integration with Opera- tional Technology (OT) systems presents significant challenges. This pa- per analyzes the fundamental obstacles in combining MLOps with OT en- vironments and proposes a systematic approach to embed MLOps prac- tices into established OT reference models. We evaluate the suitability of the Reference Architectural Model for Industry 4.0 (RAMI 4.0) and the International Society of Automation Standard 95 (ISA-95) for MLOps integration and present a detailed mapping of MLOps lifecycle compo- nents to RAMI 4.0 exemplified by a real-world use case. Our findings demonstrate that while standard MLOps practices cannot be directly transplanted to OT environments, structured adaptation using existing reference models can provide a pathway for successful integration.

Method: Systematic comparison of retrieval configurations across three axes: (i) chunking strategy, (ii) similarity scoring, and (iii) splitting granularity, using tasks from Long Code Arena including code completion and bug localization.

Result: BM25 with word-level splitting is most effective for PL-PL tasks; proprietary dense encoders work best for NL-PL but with 100x latency; optimal chunk size scales with context (32-64 lines for small budgets); line-based chunking matches syntax-aware splitting; BM25 + word splitting offers best quality-latency trade-off.

Conclusion: The study provides evidence-based recommendations for code-oriented RAG systems based on task requirements, model constraints, and computational efficiency, highlighting that BM25 with word splitting offers the best practical performance for most scenarios.

Abstract: We study retrieval design for code-focused generation tasks under realistic compute budgets. Using two complementary tasks from Long Code Arena – code completion and bug localization – we systematically compare retrieval configurations across various context window sizes along three axes: (i) chunking strategy, (ii) similarity scoring, and (iii) splitting granularity. (1) For PL-PL, sparse BM25 with word-level splitting is the most effective and practical, significantly outperforming dense alternatives while being an order of magnitude faster. (2) For NL-PL, proprietary dense encoders (Voyager-3 family) consistently beat sparse retrievers, however requiring 100x larger latency. (3) Optimal chunk size scales with available context: 32-64 line chunks work best at small budgets, and whole-file retrieval becomes competitive at 16000 tokens. (4) Simple line-based chunking matches syntax-aware splitting across budgets. (5) Retrieval latency varies by up to 200x across configurations; BPE-based splitting is needlessly slow, and BM25 + word splitting offers the best quality-latency trade-off. Thus, we provide evidence-based recommendations for implementing effective code-oriented RAG systems based on task requirements, model constraints, and computational efficiency.

Method: The authors refine the ES condition with interpretations and sampling-dependent constants, derive bounds for the expectation of squared full gradient norm, and prove convergence rates with explicit residual errors for different step-size schedules.

Result: The paper establishes O(1/K) convergence rates for SGD under ES with explicit residual errors, unifying and extending recent work in the field.

Conclusion: The analysis provides a self-contained and comprehensive treatment of SGD convergence under ES, offering refined interpretations and explicit error bounds that improve upon classical assumptions.

Abstract: Stochastic gradient descent (SGD) is the workhorse of large-scale learning, yet classical analyses rely on assumptions that can be either too strong (bounded variance) or too coarse (uniform noise). The expected smoothness (ES) condition has emerged as a flexible alternative that ties the second moment of stochastic gradients to the objective value and the full gradient. This paper presents a self-contained convergence analysis of SGD under ES. We (i) refine ES with interpretations and sampling-dependent constants; (ii) derive bounds of the expectation of squared full gradient norm; and (iii) prove $O(1/K)$ rates with explicit residual errors for various step-size schedules. All proofs are given in full detail in the appendix. Our treatment unifies and extends recent threads (Khaled and Richt'arik, 2020; Umeda and Iiduka, 2025).

Method: Integrated framework with customized PSO for feature selection, evaluated on 29 different ML models including classical classifiers, ensemble techniques, neural networks, probabilistic algorithms, and instance-based algorithms. Uses cross-validation and explainable AI methods for interpretability.

Result: Achieved superior score of 99.1% across all performance metrics (accuracy, precision), effectively reduced dimensionality, and provided transparent, model-agnostic explanations.

Conclusion: Combining swarm intelligence with explainable ML shows potential for robust, trustworthy, and clinically meaningful breast cancer diagnosis.

Abstract: Breast cancer is considered the most critical and frequently diagnosed cancer in women worldwide, leading to an increase in cancer-related mortality. Early and accurate detection is crucial as it can help mitigate possible threats while improving survival rates. In terms of prediction, conventional diagnostic methods are often limited by variability, cost, and, most importantly, risk of misdiagnosis. To address these challenges, machine learning (ML) has emerged as a powerful tool for computer-aided diagnosis, with feature selection playing a vital role in improving model performance and interpretability. This research study proposes an integrated framework that incorporates customized Particle Swarm Optimization (PSO) for feature selection. This framework has been evaluated on a comprehensive set of 29 different models, spanning classical classifiers, ensemble techniques, neural networks, probabilistic algorithms, and instance-based algorithms. To ensure interpretability and clinical relevance, the study uses cross-validation in conjunction with explainable AI methods. Experimental evaluation showed that the proposed approach achieved a superior score of 99.1% across all performance metrics, including accuracy and precision, while effectively reducing dimensionality and providing transparent, model-agnostic explanations. The results highlight the potential of combining swarm intelligence with explainable ML for robust, trustworthy, and clinically meaningful breast cancer diagnosis.

Method: MS-BART maps mass spectra and molecular structures into shared token vocabulary, uses multi-task pretraining with denoising and translation objectives, transfers to experimental spectra via finetuning with MIST-generated fingerprints, and employs chemical feedback for alignment.

Result: MS-BART achieves SOTA performance on 5/12 key metrics across MassSpecGym and NPLIB1 benchmarks, with inference speed one order of magnitude faster than diffusion-based methods.

Conclusion: The proposed MS-BART framework effectively addresses data scarcity in mass spectrometry through cross-modal pretraining and demonstrates superior performance and efficiency compared to existing methods.

Abstract: Mass spectrometry (MS) plays a critical role in molecular identification, significantly advancing scientific discovery. However, structure elucidation from MS data remains challenging due to the scarcity of annotated spectra. While large-scale pretraining has proven effective in addressing data scarcity in other domains, applying this paradigm to mass spectrometry is hindered by the complexity and heterogeneity of raw spectral signals. To address this, we propose MS-BART, a unified modeling framework that maps mass spectra and molecular structures into a shared token vocabulary, enabling cross-modal learning through large-scale pretraining on reliably computed fingerprint-molecule datasets. Multi-task pretraining objectives further enhance MS-BART’s generalization by jointly optimizing denoising and translation task. The pretrained model is subsequently transferred to experimental spectra through finetuning on fingerprint predictions generated with MIST, a pre-trained spectral inference model, thereby enhancing robustness to real-world spectral variability. While finetuning alleviates the distributional difference, MS-BART still suffers molecular hallucination and requires further alignment. We therefore introduce a chemical feedback mechanism that guides the model toward generating molecules closer to the reference structure. Extensive evaluations demonstrate that MS-BART achieves SOTA performance across 5/12 key metrics on MassSpecGym and NPLIB1 and is faster by one order of magnitude than competing diffusion-based methods, while comprehensive ablation studies systematically validate the model’s effectiveness and robustness.

Method: The authors analyze gradient distribution changes, examine cumulative DP noise effects, and study clipping as a form of gradient re-weighting, assuming a fixed compute budget (epochs).

Result: Found that larger clipping bounds C perform better under strong privacy (contradicting theory), existing batch size B heuristics don’t work, and using single (C,B) settings across tasks leads to suboptimal performance, especially when privacy constraints or compute resources vary.

Conclusion: Current approaches to hyperparameter tuning in DP transfer learning are inadequate, and better understanding of gradient distributions and cumulative noise effects is needed for optimal performance across different privacy and compute scenarios.

Abstract: Differentially private (DP) transfer learning, i.e., fine-tuning a pretrained model on private data, is the current state-of-the-art approach for training large models under privacy constraints. We focus on two key hyperparameters in this setting: the clipping bound $C$ and batch size $B$. We show a clear mismatch between the current theoretical understanding of how to choose an optimal $C$ (stronger privacy requires smaller $C$) and empirical outcomes (larger $C$ performs better under strong privacy), caused by changes in the gradient distributions. Assuming a limited compute budget (fixed epochs), we demonstrate that the existing heuristics for tuning $B$ do not work, while cumulative DP noise better explains whether smaller or larger batches perform better. We also highlight how the common practice of using a single $(C,B)$ setting across tasks can lead to suboptimal performance. We find that performance drops especially when moving between loose and tight privacy and between plentiful and limited compute, which we explain by analyzing clipping as a form of gradient re-weighting and examining cumulative DP noise.

Method: Proposed Fairness-Aware Survival Modeling (FASM) designed to mitigate algorithmic bias regarding both intra-group and cross-group risk rankings over time. Applied to SEER breast cancer data as a representative case study.

Result: FASM substantially improves fairness while preserving discrimination performance comparable to fairness-unaware survival models. Time-stratified evaluations show FASM maintains stable fairness over a 10-year horizon, with greatest improvements during mid-term follow-up.

Conclusion: FASM enables development of survival models that prioritize both accuracy and equity in clinical decision-making, advancing fairness as a core principle in clinical care.

Abstract: As machine learning models become increasingly integrated into healthcare, structural inequities and social biases embedded in clinical data can be perpetuated or even amplified by data-driven models. In survival analysis, censoring and time dynamics can further add complexity to fair model development. Additionally, algorithmic fairness approaches often overlook disparities in cross-group rankings, e.g., high-risk Black patients may be ranked below lower-risk White patients who do not experience the event of mortality. Such misranking can reinforce biological essentialism and undermine equitable care. We propose a Fairness-Aware Survival Modeling (FASM), designed to mitigate algorithmic bias regarding both intra-group and cross-group risk rankings over time. Using breast cancer prognosis as a representative case and applying FASM to SEER breast cancer data, we show that FASM substantially improves fairness while preserving discrimination performance comparable to fairness-unaware survival models. Time-stratified evaluations show that FASM maintains stable fairness over a 10-year horizon, with the greatest improvements observed during the mid-term of follow-up. Our approach enables the development of survival models that prioritize both accuracy and equity in clinical decision-making, advancing fairness as a core principle in clinical care.

Method: H-SPLID algorithm explicitly decomposes salient and non-salient features into separate spaces. It establishes a theoretical bound showing that expected prediction deviation under perturbations is upper-bounded by the dimension of salient subspace and HSIC between inputs and representations.

Result: Empirical evaluations on image classification tasks show that models trained with H-SPLID primarily rely on salient input components, demonstrating reduced sensitivity to perturbations affecting non-salient features such as image backgrounds.

Conclusion: H-SPLID successfully links robustness and latent representation compression through dimensionality and information preservation, providing a principled approach for learning task-relevant features while being robust to irrelevant input variations.

Abstract: We introduce H-SPLID, a novel algorithm for learning salient feature representations through the explicit decomposition of salient and non-salient features into separate spaces. We show that H-SPLID promotes learning low-dimensional, task-relevant features. We prove that the expected prediction deviation under input perturbations is upper-bounded by the dimension of the salient subspace and the Hilbert-Schmidt Independence Criterion (HSIC) between inputs and representations. This establishes a link between robustness and latent representation compression in terms of the dimensionality and information preserved. Empirical evaluations on image classification tasks show that models trained with H-SPLID primarily rely on salient input components, as indicated by reduced sensitivity to perturbations affecting non-salient features, such as image backgrounds. Our code is available at https://github.com/neu-spiral/H-SPLID.

Method: Propose a framework for task-oriented autonomous communications using LLMs/LMMs with multimodal sensing integration, adaptive reconfiguration, and prompt/fine-tuning strategies. Demonstrate through three case studies: LMM-based traffic control, LLM-based robot scheduling, and LMM-based environment-aware channel estimation.

Result: The proposed LLM/LMM-aided autonomous systems significantly outperform conventional and discriminative deep learning model-based techniques, maintaining robustness under dynamic objectives, varying input parameters, and heterogeneous multimodal conditions.

Conclusion: LLMs and LMMs provide superior performance and robustness for autonomous communications in 6G networks compared to conventional static optimization methods, especially in dynamic and heterogeneous environments.

Abstract: Large language models (LLMs) and large multimodal models (LMMs) have achieved unprecedented breakthrough, showcasing remarkable capabilities in natural language understanding, generation, and complex reasoning. This transformative potential has positioned them as key enablers for 6G autonomous communications among machines, vehicles, and humanoids. In this article, we provide an overview of task-oriented autonomous communications with LLMs/LMMs, focusing on multimodal sensing integration, adaptive reconfiguration, and prompt/fine-tuning strategies for wireless tasks. We demonstrate the framework through three case studies: LMM-based traffic control, LLM-based robot scheduling, and LMM-based environment-aware channel estimation. From experimental results, we show that the proposed LLM/LMM-aided autonomous systems significantly outperform conventional and discriminative deep learning (DL) model-based techniques, maintaining robustness under dynamic objectives, varying input parameters, and heterogeneous multimodal conditions where conventional static optimization degrades.

Method: Constructs monitor heterogeneous graph, uses multi-head attention mechanism to focus on heterogeneous neighbors and attributes, attends to random walk paths for long-range dependencies, and employs multi-faceted loss functions.

Result: Achieved 43.1% increase in MRR over existing methods. Product teams rated the feature 4.5/5 for usefulness.

Conclusion: DiRecGNN effectively addresses entity recommendation for cloud service monitoring with transformer-inspired attention mechanisms, demonstrating significant performance improvements and practical utility.

Abstract: In this paper, we present DiRecGNN, an attention-enhanced entity recommendation framework for monitoring cloud services at Microsoft. We provide insights on the usefulness of this feature as perceived by the cloud service owners and lessons learned from deployment. Specifically, we introduce the problem of recommending the optimal subset of attributes (dimensions) that should be tracked by an automated watchdog (monitor) for cloud services. To begin, we construct the monitor heterogeneous graph at production-scale. The interaction dynamics of these entities are often characterized by limited structural and engagement information, resulting in inferior performance of state-of-the-art approaches. Moreover, traditional methods fail to capture the dependencies between entities spanning a long range due to their homophilic nature. Therefore, we propose an attention-enhanced entity ranking model inspired by transformer architectures. Our model utilizes a multi-head attention mechanism to focus on heterogeneous neighbors and their attributes, and further attends to paths sampled using random walks to capture long-range dependencies. We also employ multi-faceted loss functions to optimize for relevant recommendations while respecting the inherent sparsity of the data. Empirical evaluations demonstrate significant improvements over existing methods, with our model achieving a 43.1% increase in MRR. Furthermore, product teams who consumed these features perceive the feature as useful and rated it 4.5 out of 5.

Method: Uses graph neural networks (GRACE) to formalize locus of care prediction as structured learning, performs extensive feature engineering, and develops an unbiased meta-graph approach to overcome class imbalance.

Result: Experimental results show 11-35% improvement in F1 score for the minority class compared to competitive baselines on real-world data.

Conclusion: The GRACE framework effectively addresses class imbalance in addiction care prediction and demonstrates significant performance improvements over existing methods.

Abstract: Determining the appropriate locus of care for addiction patients is one of the most critical clinical decisions that affects patient treatment outcomes and effective use of resources. With a lack of sufficient specialized treatment resources, such as inpatient beds or staff, there is an unmet need to develop an automated framework for the same. Current decision-making approaches suffer from severe class imbalances in addiction datasets. To address this limitation, we propose a novel graph neural network (GRACE) framework that formalizes locus of care prediction as a structured learning problem. Further, we perform extensive feature engineering and propose a new approach of obtaining an unbiased meta-graph to train a GNN to overcome the class imbalance problem. Experimental results in real-world data show an improvement of 11-35% in terms of the F1 score of the minority class over competitive baselines. The codes and note embeddings are available at https://anonymous.4open.science/r/GRACE-F8E1/.

Method: Derived a new tight lower bound on Kullback-Leibler divergence as a function of JSD, specialized to joint and marginal distributions. Implemented JSD-based objectives via binary classifier cross-entropy loss that distinguishes joint from marginal pairs.

Result: The lower bound estimator provides stable, low-variance estimates of a tight lower bound on MI across various reference scenarios, outperforming state-of-the-art neural estimators. Demonstrated practical usefulness in Information Bottleneck framework.

Conclusion: The work provides new theoretical justifications and strong empirical evidence for using discriminative learning in MI-based representation learning, showing that maximizing JSD-based information increases a guaranteed lower bound on mutual information.

Abstract: Mutual Information (MI) is a fundamental measure of statistical dependence widely used in representation learning. While direct optimization of MI via its definition as a Kullback-Leibler divergence (KLD) is often intractable, many recent methods have instead maximized alternative dependence measures, most notably, the Jensen-Shannon divergence (JSD) between joint and product of marginal distributions via discriminative losses. However, the connection between these surrogate objectives and MI remains poorly understood. In this work, we bridge this gap by deriving a new, tight, and tractable lower bound on KLD as a function of JSD in the general case. By specializing this bound to joint and marginal distributions, we demonstrate that maximizing the JSD-based information increases a guaranteed lower bound on mutual information. Furthermore, we revisit the practical implementation of JSD-based objectives and observe that minimizing the cross-entropy loss of a binary classifier trained to distinguish joint from marginal pairs recovers a known variational lower bound on the JSD. Extensive experiments demonstrate that our lower bound is tight when applied to MI estimation. We compared our lower bound to state-of-the-art neural estimators of variational lower bound across a range of established reference scenarios. Our lower bound estimator consistently provides a stable, low-variance estimate of a tight lower bound on MI. We also demonstrate its practical usefulness in the context of the Information Bottleneck framework. Taken together, our results provide new theoretical justifications and strong empirical evidence for using discriminative learning in MI-based representation learning.

Method: The approach processes binned spikes by projecting them into a shared latent space, where spiking modules adapted to input timing extract relevant features; these latent representations are then integrated and decoded to generate behavioral predictions using SNNs.

Result: Evaluation on 113 recording sessions from 6 non-human primates (43 hours total) shows Spikachu outperforms causal baselines with significantly lower energy consumption (2.26-418.81x less), and scaling training enables few-shot transfer to unseen sessions, subjects, and tasks.

Conclusion: Spikachu introduces a scalable, online-compatible neural decoding framework based on SNNs that achieves competitive performance relative to state-of-the-art models while consuming orders of magnitude less energy.

Abstract: Brain-computer interfaces (BCIs) promise to enable vital functions, such as speech and prosthetic control, for individuals with neuromotor impairments. Central to their success are neural decoders, models that map neural activity to intended behavior. Current learning-based decoding approaches fall into two classes: simple, causal models that lack generalization, or complex, non-causal models that generalize and scale offline but struggle in real-time settings. Both face a common challenge, their reliance on power-hungry artificial neural network backbones, which makes integration into real-world, resource-limited systems difficult. Spiking neural networks (SNNs) offer a promising alternative. Because they operate causally these models are suitable for real-time use, and their low energy demands make them ideal for battery-constrained environments. To this end, we introduce Spikachu: a scalable, causal, and energy-efficient neural decoding framework based on SNNs. Our approach processes binned spikes directly by projecting them into a shared latent space, where spiking modules, adapted to the timing of the input, extract relevant features; these latent representations are then integrated and decoded to generate behavioral predictions. We evaluate our approach on 113 recording sessions from 6 non-human primates, totaling 43 hours of recordings. Our method outperforms causal baselines when trained on single sessions using between 2.26 and 418.81 times less energy. Furthermore, we demonstrate that scaling up training to multiple sessions and subjects improves performance and enables few-shot transfer to unseen sessions, subjects, and tasks. Overall, Spikachu introduces a scalable, online-compatible neural decoding framework based on SNNs, whose performance is competitive relative to state-of-the-art models while consuming orders of magnitude less energy.

Method: Uses knowledge distillation to transfer information from models trained on lower-rarity events, and a mixture of experts mechanism that dynamically selects and fuses outputs from expert models across different rarity levels.

Result: Experiments show forecasting accuracy on extreme events improves from 3% to 78% across multiple datasets.

Conclusion: xTime effectively addresses the challenges of extreme event forecasting and achieves significant improvements in prediction accuracy for rare events.

Abstract: Extreme events frequently occur in real-world time series and often carry significant practical implications. In domains such as climate and healthcare, these events, such as floods, heatwaves, or acute medical episodes, can lead to serious consequences. Accurate forecasting of such events is therefore of substantial importance. Most existing time series forecasting models are optimized for overall performance within the prediction window, but often struggle to accurately predict extreme events, such as high temperatures or heart rate spikes. The main challenges are data imbalance and the neglect of valuable information contained in intermediate events that precede extreme events. In this paper, we propose xTime, a novel framework for extreme event forecasting in time series. xTime leverages knowledge distillation to transfer information from models trained on lower-rarity events, thereby improving prediction performance on rarer ones. In addition, we introduce a mixture of experts (MoE) mechanism that dynamically selects and fuses outputs from expert models across different rarity levels, which further improves the forecasting performance for extreme events. Experiments on multiple datasets show that xTime achieves consistent improvements, with forecasting accuracy on extreme events improving from 3% to 78%.

Method: Hybrid Bayesian mixture-of-experts framework combining physical path-loss model and CNN through log-linear pooling, using building-height maps to capture urban propagation effects with Bayesian inference via Laplace approximation.

Result: Experiments on urban ray-tracing data show improved localization accuracy and reduced uncertainty with more training points, with uncertainty concentrating near jammer and along urban canyons.

Conclusion: The proposed hybrid framework successfully improves jammer localization and RSS field reconstruction while providing uncertainty quantification in challenging urban environments.

Abstract: Global Navigation Satellite System (GNSS) signals are vulnerable to jamming, particularly in urban areas where multipath and shadowing distort received power. Previous data-driven approaches achieved reasonable localization but poorly reconstructed the received signal strength (RSS) field due to limited spatial context. We propose a hybrid Bayesian mixture-of-experts framework that fuses a physical path-loss (PL) model and a convolutional neural network (CNN) through log-linear pooling. The PL expert ensures physical consistency, while the CNN leverages building-height maps to capture urban propagation effects. Bayesian inference with Laplace approximation provides posterior uncertainty over both the jammer position and RSS field. Experiments on urban ray-tracing data show that localization accuracy improves and uncertainty decreases with more training points, while uncertainty concentrates near the jammer and along urban canyons where propagation is most sensitive.

Method: Follows a clear evolutionary path: from early masked models unifying representation learning, to unified architectures sharing a single paradigm, then to interactive generative models closing the action-perception loop, and finally to memory-augmented systems sustaining consistent worlds over time.

Result: Identifies the core components of successful world models: the generative heart, the interactive loop, and the memory system, while bypassing loosely related branches.

Conclusion: This focused approach on the generative core, interactive loop, and memory system represents the most promising path towards developing true world models.

Abstract: This is not a typical survey of world models; it is a guide for those who want to build worlds. We do not aim to catalog every paper that has ever mentioned a ``world model". Instead, we follow one clear road: from early masked models that unified representation learning across modalities, to unified architectures that share a single paradigm, then to interactive generative models that close the action-perception loop, and finally to memory-augmented systems that sustain consistent worlds over time. We bypass loosely related branches to focus on the core: the generative heart, the interactive loop, and the memory system. We show that this is the most promising path towards true world models.

Method: Developed a two-system RNN model: (1) ‘what’ system uses probabilistic generative model to infer computational context and learn task vocabulary incrementally, (2) ‘how’ system implements computations via low-rank RNN components composed according to inferred context.

Result: The framework enables continual learning without catastrophic forgetting, demonstrates competitive performance, forward/backward transfer, and fast compositional generalization to unseen tasks.

Conclusion: The two-system approach provides an effective neural mechanism for continual learning and compositional skill reuse, addressing key challenges in flexible intelligent behavior.

Abstract: The ability to continually learn, retain and deploy skills to accomplish goals is a key feature of intelligent and efficient behavior. However, the neural mechanisms facilitating the continual learning and flexible (re-)composition of skills remain elusive. Here, we study continual learning and the compositional reuse of learned computations in recurrent neural network (RNN) models using a novel two-system approach: one system that infers what computation to perform, and one that implements how to perform it. We focus on a set of compositional cognitive tasks commonly studied in neuroscience. To construct the what system, we first show that a large family of tasks can be systematically described by a probabilistic generative model, where compositionality stems from a shared underlying vocabulary of discrete task epochs. The shared epoch structure makes these tasks inherently compositional. We first show that this compositionality can be systematically described by a probabilistic generative model. Furthermore, We develop an unsupervised online learning approach that can learn this model on a single-trial basis, building its vocabulary incrementally as it is exposed to new tasks, and inferring the latent epoch structure as a time-varying computational context within a trial. We implement the how system as an RNN whose low-rank components are composed according to the context inferred by the what system. Contextual inference facilitates the creation, learning, and reuse of low-rank RNN components as new tasks are introduced sequentially, enabling continual learning without catastrophic forgetting. Using an example task set, we demonstrate the efficacy and competitive performance of this two-system learning framework, its potential for forward and backward transfer, as well as fast compositional generalization to unseen tasks.

Method: Uses neuron-specific most exciting inputs (MEIs), structural similarity index measure loss, and adversarial training to decode visual stimuli from neural activity.

Result: Achieves SOTA performance, reconstructs high-fidelity natural images from only 1,000-2,500 neurons and <1,000 training samples. MEIs identified as main performance driver.

Conclusion: Demonstrates feasibility of reliable decoding in early visual system and provides practical insights for neuroscience and neuroengineering applications.

Abstract: Decoding visual stimuli from neural population activity is crucial for understanding the brain and for applications in brain-machine interfaces. However, such biological data is often scarce, particularly in primates or humans, where high-throughput recording techniques, such as two-photon imaging, remain challenging or impossible to apply. This, in turn, poses a challenge for deep learning decoding techniques. To overcome this, we introduce MEIcoder, a biologically informed decoding method that leverages neuron-specific most exciting inputs (MEIs), a structural similarity index measure loss, and adversarial training. MEIcoder achieves state-of-the-art performance in reconstructing visual stimuli from single-cell activity in primary visual cortex (V1), especially excelling on small datasets with fewer recorded neurons. Using ablation studies, we demonstrate that MEIs are the main drivers of the performance, and in scaling experiments, we show that MEIcoder can reconstruct high-fidelity natural-looking images from as few as 1,000-2,500 neurons and less than 1,000 training data points. We also propose a unified benchmark with over 160,000 samples to foster future research. Our results demonstrate the feasibility of reliable decoding in early visual system and provide practical insights for neuroscience and neuroengineering applications.

Method: A two-stage framework: (1) train forecasting model on anomaly-free data, (2) perform forecasts on unseen data and flag deviations beyond threshold as anomalies. Two approaches: N-BEATS for univariate forecasting assuming independence, and GNN to capture variable relationships.

Result: Both models show strong forecasting performance up to 20 time points and stable anomaly prediction up to 50 time points. GNN consistently outperforms N-BEATS while requiring fewer parameters and lower computational cost.

Conclusion: GNN is positioned as a promising solution for online anomaly forecasting in manufacturing environments, advancing the field from anomaly detection to prediction for proactive fault prevention.

Abstract: Semiconductor manufacturing is an extremely complex and precision-driven process, characterized by thousands of interdependent parameters collected across diverse tools and process steps. Multi-variate time-series analysis has emerged as a critical field for real-time monitoring and fault detection in such environments. However, anomaly prediction in semiconductor fabrication presents several critical challenges, including high dimensionality of sensor data and severe class imbalance due to the rarity of true faults. Furthermore, the complex interdependencies between variables complicate both anomaly prediction and root-cause-analysis. This paper proposes two novel approaches to advance the field from anomaly detection to anomaly prediction, an essential step toward enabling real-time process correction and proactive fault prevention. The proposed anomaly prediction framework contains two main stages: (a) training a forecasting model on a dataset assumed to contain no anomalies, and (b) performing forecast on unseen time series data. The forecast is compared with the forecast of the trained signal. Deviations beyond a predefined threshold are flagged as anomalies. The two approaches differ in the forecasting model employed. The first assumes independence between variables by utilizing the N-BEATS model for univariate time series forecasting. The second lifts this assumption by utilizing a Graph Neural Network (GNN) to capture inter-variable relationships. Both models demonstrate strong forecasting performance up to a horizon of 20 time points and maintain stable anomaly prediction up to 50 time points. The GNN consistently outperforms the N-BEATS model while requiring significantly fewer trainable parameters and lower computational cost. These results position the GNN as promising solution for online anomaly forecasting to be deployed in manufacturing environments.

Method: Retrospective analysis of 54,209 inpatient admissions using constrained score optimization (CSO) models on JHFRAT assessment data and additional EHR variables, comparing performance with benchmark XGBoost models.

Result: The CSO model demonstrated significant improvement over current JHFRAT (AUC-ROC=0.91 vs 0.86) and showed similar performance with and without EHR variables. While XGBoost achieved higher AUC-ROC (0.94), CSO demonstrated better robustness to risk labeling variations.

Conclusion: The evidence-based constrained score optimization approach provides a robust foundation for systematically enhancing inpatient fall prevention protocols and patient safety using data-driven optimization techniques.

Abstract: In this study we aim to better align fall risk prediction from the Johns Hopkins Fall Risk Assessment Tool (JHFRAT) with additional clinically meaningful measures via a data-driven modelling approach. We conducted a retrospective analysis of 54,209 inpatient admissions from three Johns Hopkins Health System hospitals between March 2022 and October 2023. A total of 20,208 admissions were included as high fall risk encounters, and 13,941 were included as low fall risk encounters. To incorporate clinical knowledge and maintain interpretability, we employed constrained score optimization (CSO) models on JHFRAT assessment data and additional electronic health record (EHR) variables. The model demonstrated significant improvements in predictive performance over the current JHFRAT (CSO AUC-ROC=0.91, JHFRAT AUC-ROC=0.86). The constrained score optimization models performed similarly with and without the EHR variables. Although the benchmark black-box model (XGBoost), improves upon the performance metrics of the knowledge-based constrained logistic regression (AUC-ROC=0.94), the CSO demonstrates more robustness to variations in risk labelling. This evidence-based approach provides a robust foundation for health systems to systematically enhance inpatient fall prevention protocols and patient safety using data-driven optimization techniques, contributing to improved risk assessment and resource allocation in healthcare settings.

Method: Error: OutputParser failed

Result: Error: OutputParser failed

Conclusion: Error: OutputParser failed

Abstract: Thompson sampling (TS) is a powerful and widely used strategy for sequential decision-making, with applications ranging from Bayesian optimization to reinforcement learning (RL). Despite its success, the theoretical foundations of TS remain limited, particularly in settings with complex temporal structure such as RL. We address this gap by establishing no-regret guarantees for TS using models with Gaussian marginal distributions. Specifically, we consider TS in episodic RL with joint Gaussian process (GP) priors over rewards and transitions. We prove a regret bound of $\mathcal{\tilde{O}}(\sqrt{KH\Gamma(KH)})$ over $K$ episodes of horizon $H$, where $\Gamma(\cdot)$ captures the complexity of the GP model. Our analysis addresses several challenges, including the non-Gaussian nature of value functions and the recursive structure of Bellman updates, and extends classical tools such as the elliptical potential lemma to multi-output settings. This work advances the understanding of TS in RL and highlights how structural assumptions and model uncertainty shape its performance in finite-horizon Markov Decision Processes.

Method: Assume each modality follows a mixture of multivariate Gaussian distributions and adopt Dirichlet process to calculate mixture weights, leveraging its richer-gets-richer property to dynamically allocate feature contributions.

Result: Extensive experiments on several multimodal datasets demonstrate superior performance over competitors, with ablation analysis validating DP’s effectiveness in aligning modality distributions and robustness to hyperparameter changes.

Conclusion: The proposed DP-driven framework effectively balances intra-modal representation learning and cross-modal alignment, achieving optimal multimodal fusion through dynamic feature selection.

Abstract: Developing effective multimodal fusion approaches has become increasingly essential in many real-world scenarios, such as health care and finance. The key challenge is how to preserve the feature expressiveness in each modality while learning cross-modal interactions. Previous approaches primarily focus on the cross-modal alignment, while over-emphasis on the alignment of marginal distributions of modalities may impose excess regularization and obstruct meaningful representations within each modality. The Dirichlet process (DP) mixture model is a powerful Bayesian non-parametric method that can amplify the most prominent features by its richer-gets-richer property, which allocates increasing weights to them. Inspired by this unique characteristic of DP, we propose a new DP-driven multimodal learning framework that automatically achieves an optimal balance between prominent intra-modal representation learning and cross-modal alignment. Specifically, we assume that each modality follows a mixture of multivariate Gaussian distributions and further adopt DP to calculate the mixture weights for all the components. This paradigm allows DP to dynamically allocate the contributions of features and select the most prominent ones, leveraging its richer-gets-richer property, thus facilitating multimodal feature fusion. Extensive experiments on several multimodal datasets demonstrate the superior performance of our model over other competitors. Ablation analysis further validates the effectiveness of DP in aligning modality distributions and its robustness to changes in key hyperparameters. Code is anonymously available at https://github.com/HKU-MedAI/DPMM.git

Method: Trained a 270M parameter chess Transformer and tested it on out-of-distribution scenarios including rule extrapolation tests and Chess960 variants, comparing against symbolic AI algorithms.

Result: Transformers show strong compositional generalization by consistently choosing valid moves in OOD situations and generating high-quality moves for puzzles. They adapt basic strategies in Chess960 but perform worse than symbolic AI with explicit search, though the gap is smaller in real gameplay.

Conclusion: Transformers demonstrate emergent compositional understanding of chess rules and can generalize systematically, but still have limitations compared to traditional symbolic approaches in complex scenarios requiring deep search.

Abstract: Chess is a canonical example of a task that requires rigorous reasoning and long-term planning. Modern decision Transformers - trained similarly to LLMs - are able to learn competent gameplay, but it is unclear to what extent they truly capture the rules of chess. To investigate this, we train a 270M parameter chess Transformer and test it on out-of-distribution scenarios, designed to reveal failures of systematic generalization. Our analysis shows that Transformers exhibit compositional generalization, as evidenced by strong rule extrapolation: they adhere to fundamental syntactic rules of the game by consistently choosing valid moves even in situations very different from the training data. Moreover, they also generate high-quality moves for OOD puzzles. In a more challenging test, we evaluate the models on variants including Chess960 (Fischer Random Chess) - a variant of chess where starting positions of pieces are randomized. We found that while the model exhibits basic strategy adaptation, they are inferior to symbolic AI algorithms that perform explicit search, but gap is smaller when playing against users on Lichess. Moreover, the training dynamics revealed that the model initially learns to move only its own pieces, suggesting an emergent compositional understanding of the game.

Method: The authors mathematically analyze how reverse/forward KL determines optimal target distributions, then propose a simple algorithm that minimally adjusts reward magnitudes to create target distributions covering all high-quality modes. The method works with both KL types and requires no external diversity signals.

Result: The proposed algorithm successfully post-trains both Large Language Models and Chemical Language Models, achieving higher solution quality and diversity compared to naive approaches using either forward or reverse KL regularization.

Conclusion: Mode coverage in RL with language models depends primarily on regularization strength and reward scales, not KL divergence type. The simple reward modification algorithm enables diverse sampling from multiple high-quality modes without external diversity signals.

Abstract: It is commonly believed that optimizing the reverse KL divergence results in “mode seeking”, while optimizing forward KL results in “mass covering”, with the latter being preferred if the goal is to sample from multiple diverse modes. We show – mathematically and empirically – that this intuition does not necessarily transfer well to doing reinforcement learning with reverse/forward KL regularization (e.g. as commonly used with language models). Instead, the choice of reverse/forward KL determines the family of optimal target distributions, parameterized by the regularization coefficient. Mode coverage depends primarily on other factors, such as regularization strength, and relative scales between rewards and reference probabilities. Further, we show commonly used settings such as low regularization strength and equal verifiable rewards tend to specify unimodal target distributions, meaning the optimization objective is, by construction, non-diverse. We leverage these insights to construct a simple, scalable, and theoretically justified algorithm. It makes minimal changes to reward magnitudes, yet optimizes for a target distribution which puts high probability over all high-quality sampling modes. In experiments, this simple modification works to post-train both Large Language Models and Chemical Language Models to have higher solution quality and diversity, without any external signals of diversity, and works with both forward and reverse KL when using either naively fails.

Method: ERA leverages explicit reward functions to create gradient-based objectives for optimizing autoregressive policies, relating to PPO and DPO but converging to a Gibbs-Boltzmann distribution with reward as energy function.

Result: ERA performs well relative to DPO when preference observations are limited, and successfully aligns molecular transformers and protein language models to generate molecules and sequences with specified properties across diverse chemical space.

Conclusion: ERA provides a scalable, reinforcement-free approach for aligning generative models to produce molecules and protein sequences with desired properties, robustly searching through chemical space.

Abstract: Searching through chemical space is an exceptionally challenging problem because the number of possible molecules grows combinatorially with the number of atoms. Large, autoregressive models trained on databases of chemical compounds have yielded powerful generators, but we still lack robust strategies for generating molecules with desired properties. This molecular search problem closely resembles the “alignment” problem for large language models, though for many chemical tasks we have a specific and easily evaluable reward function. Here, we introduce an algorithm called energy rank alignment (ERA) that leverages an explicit reward function to produce a gradient-based objective that we use to optimize autoregressive policies. We show theoretically that this algorithm is closely related to proximal policy optimization (PPO) and direct preference optimization (DPO), but has a minimizer that converges to an ideal Gibbs-Boltzmann distribution with the reward playing the role of an energy function. Furthermore, this algorithm is highly scalable, does not require reinforcement learning, and performs well relative to DPO when the number of preference observations per pairing is small. We deploy this approach to align molecular transformers and protein language models to generate molecules and protein sequences, respectively, with externally specified properties and find that it does so robustly, searching through diverse parts of chemical space.

Method: The authors propose new learning objectives that integrate regularization effects from multiple previous posterior estimations, preventing individual errors from dominating future updates. They draw connections to Temporal-Difference methods from Reinforcement Learning.

Result: Experiments on challenging Continual Learning benchmarks show the proposed approach effectively mitigates Catastrophic Forgetting and outperforms strong Variational CL methods.

Conclusion: Integrating multiple previous posterior estimations in learning objectives provides an effective solution to mitigate compounding approximation errors and Catastrophic Forgetting in Bayesian Continual Learning.

Abstract: Machine Learning models in real-world applications must continuously learn new tasks to adapt to shifts in the data-generating distribution. Yet, for Continual Learning (CL), models often struggle to balance learning new tasks (plasticity) with retaining previous knowledge (memory stability). Consequently, they are susceptible to Catastrophic Forgetting, which degrades performance and undermines the reliability of deployed systems. In the Bayesian CL literature, variational methods tackle this challenge by employing a learning objective that recursively updates the posterior distribution while constraining it to stay close to its previous estimate. Nonetheless, we argue that these methods may be ineffective due to compounding approximation errors over successive recursions. To mitigate this, we propose new learning objectives that integrate the regularization effects of multiple previous posterior estimations, preventing individual errors from dominating future posterior updates and compounding over time. We reveal insightful connections between these objectives and Temporal-Difference methods, a popular learning mechanism in Reinforcement Learning and Neuroscience. Experiments on challenging CL benchmarks show that our approach effectively mitigates Catastrophic Forgetting, outperforming strong Variational CL methods.

Method: Proposes IGNN with three design principles that enable distinct hop-wise generalization and adaptive smoothness to alleviate the smoothness-generalization dilemma.

Result: IGNN demonstrates superiority over 30 baselines and reveals notable universality in certain homophilic GNN variants.

Conclusion: IGNN effectively addresses the smoothness-generalization dilemma in GNNs and shows strong performance across varying homophily levels.

Abstract: Graph Neural Networks (GNNs) have achieved great success but are often considered to be challenged by varying levels of homophily in graphs. Recent empirical studies have surprisingly shown that homophilic GNNs can perform well across datasets of different homophily levels with proper hyperparameter tuning, but the underlying theory and effective architectures remain unclear. To advance GNN universality across varying homophily, we theoretically revisit GNN message passing and uncover a novel smoothness-generalization dilemma, where increasing hops inevitably enhances smoothness at the cost of generalization. This dilemma hinders learning in higher-order homophilic neighborhoods and all heterophilic ones, where generalization is critical due to complex neighborhood class distributions that are sensitive to shifts induced by noise and sparsity. To address this, we introduce the Inceptive Graph Neural Network (IGNN) built on three simple yet effective design principles, which alleviate the dilemma by enabling distinct hop-wise generalization alongside improved overall generalization with adaptive smoothness. Benchmarking against 30 baselines demonstrates IGNN’s superiority and reveals notable universality in certain homophilic GNN variants. Our code and datasets are available at https://github.com/galogm/IGNN.

Method: Dual-mind framework with RSSM-S1 for intuitive state transitions and LINN-S2 for logical reasoning, using inter-system feedback to ensure logical consistency.

Result: Significant improvements in logical coherence, trial efficiency, data efficiency, and long-term imagination on DMControl benchmark tasks.

Conclusion: DMWM successfully integrates logical reasoning with world models, enabling more accurate and consistent long-term imagination for planning tasks.

Abstract: Imagination in world models is crucial for enabling agents to learn long-horizon policy in a sample-efficient manner. Existing recurrent state-space model (RSSM)-based world models depend on single-step statistical inference to capture the environment dynamics, and, hence, they are unable to perform long-term imagination tasks due to the accumulation of prediction errors. Inspired by the dual-process theory of human cognition, we propose a novel dual-mind world model (DMWM) framework that integrates logical reasoning to enable imagination with logical consistency. DMWM is composed of two components: an RSSM-based System 1 (RSSM-S1) component that handles state transitions in an intuitive manner and a logic-integrated neural network-based System 2 (LINN-S2) component that guides the imagination process through hierarchical deep logical reasoning. The inter-system feedback mechanism is designed to ensure that the imagination process follows the logical rules of the real environment. The proposed framework is evaluated on benchmark tasks that require long-term planning from the DMControl suite. Extensive experimental results demonstrate that the proposed framework yields significant improvements in terms of logical coherence, trial efficiency, data efficiency and long-term imagination over the state-of-the-art world models.

Method: Developed CompleteP parameterization that ensures hyperparameter transfer across depth changes while enabling non-lazy learning in all layers, allowing layers to learn beyond their linearization.

Result: CompleteP enables 12-34% compute efficiency improvements over prior state-of-the-art and allows wider range of model width/depth ratios to remain compute-efficient, better suiting different hardware settings.

Conclusion: CompleteP parameterization successfully addresses both hyperparameter transfer and lazy learning issues in LLM training, unlocking more efficient model architectures and substantial compute savings.

Abstract: We study compute efficiency of LLM training when using different parameterizations, i.e., rules for adjusting model and optimizer hyperparameters (HPs) as model size changes. Some parameterizations fail to transfer optimal base HPs (such as learning rate) across changes in model depth, requiring practitioners to either re-tune these HPs as they scale up (expensive), or accept sub-optimal training when re-tuning is prohibitive. Even when they achieve HP transfer, we develop theory to show parameterizations may still exist in the lazy learning regime where layers learn only features close to their linearization, preventing effective use of depth and nonlinearity. Finally, we identify and adopt the parameterization we call CompleteP that achieves both depth-wise HP transfer and non-lazy learning in all layers. CompleteP enables a wider range of model width/depth ratios to remain compute-efficient, unlocking shapes better suited for different hardware settings and operational contexts. Moreover, CompleteP enables 12-34% compute efficiency improvements over the prior state-of-the-art. All experiments were run on Cerebras CS-3 systems. A minimal implementation is available at https://github.com/EleutherAI/nanoGPT-mup/tree/completep.

Method: Uses carefully crafted patches on the original neural network for targeted forgetting, inspired by neural network repair. Applies iterative selection of representative data points for bulk unlearning.

Result: Effective unlearning with measurable results while preserving model performance, competitive in efficiency and memory consumption compared to baselines.

Conclusion: The patch-based approach provides a practical and verifiable solution for neural network unlearning with certifiable guarantees.

Abstract: It is often desirable to remove (a.k.a. unlearn) a specific part of the training data from a trained neural network model. A typical application scenario is to protect the data holder’s right to be forgotten, which has been promoted by many recent regulation rules. Existing unlearning methods involve training alternative models with remaining data, which may be costly and challenging to verify from the data holder or a thirdparty auditor’s perspective. In this work, we provide a new angle and propose a novel unlearning approach by imposing carefully crafted “patch” on the original neural network to achieve targeted “forgetting” of the requested data to delete. Specifically, inspired by the research line of neural network repair, we propose to strategically seek a lightweight minimum “patch” for unlearning a given data point with certifiable guarantee. Furthermore, to unlearn a considerable amount of data points (or an entire class), we propose to iteratively select a small subset of representative data points to unlearn, which achieves the effect of unlearning the whole set. Extensive experiments on multiple categorical datasets demonstrates our approach’s effectiveness, achieving measurable unlearning while preserving the model’s performance and being competitive in efficiency and memory consumption compared to various baseline methods.

Method: Introduces a novel reformulation of attention mechanism that reveals underlying FFN-like structure, enabling parameter sharing between FFN and attention components in UMoE architecture.

Result: UMoE achieves superior performance through attention-based MoE layers while maintaining efficient parameter sharing.

Conclusion: The proposed unified approach successfully bridges the gap between attention and FFN MoE layers, demonstrating improved performance and implementation efficiency.

Abstract: Sparse Mixture of Experts (MoE) architectures have emerged as a promising approach for scaling Transformer models. While initial works primarily incorporated MoE into feed-forward network (FFN) layers, recent studies have explored extending the MoE paradigm to attention layers to enhance model performance. However, existing attention-based MoE layers require specialized implementations and demonstrate suboptimal performance compared to their FFN-based counterparts. In this paper, we aim to unify MoE designs in attention and FFN layers by introducing a novel reformulation of the attention mechanism, that reveals an underlying FFN-like structure within attention modules. Our proposed architecture, UMoE, achieves superior performance through attention-based MoE layers while enabling efficient parameter sharing between FFN and attention components.

Method: Alternates between finding joint probability distributions to align data from different protected groups and optimizing cluster centers in the aligned space.

Result: FCA achieves superior trade-off between fairness and clustering utility, and attains near-perfect fairness without numerical instability.

Conclusion: FCA provides a theoretically guaranteed approach for high-utility fair clustering that overcomes limitations of existing methods.

Abstract: Algorithmic fairness in clustering aims to balance the proportions of instances assigned to each cluster with respect to a given sensitive attribute. While recently developed fair clustering algorithms optimize clustering objectives under specific fairness constraints, their inherent complexity or approximation often results in suboptimal clustering utility or numerical instability in practice. To resolve these limitations, we propose a new fair clustering algorithm based on a novel decomposition of the fair $K$-means clustering objective function. The proposed algorithm, called Fair Clustering via Alignment (FCA), operates by alternately (i) finding a joint probability distribution to align the data from different protected groups, and (ii) optimizing cluster centers in the aligned space. A key advantage of FCA is that it theoretically guarantees approximately optimal clustering utility for any given fairness level without complex constraints, thereby enabling high-utility fair clustering in practice. Experiments show that FCA outperforms existing methods by (i) attaining a superior trade-off between fairness level and clustering utility, and (ii) achieving near-perfect fairness without numerical instability.

Method: Using Anthropic’s toy model with weight decay to control superposition degree, systematically studying loss scaling with model size across different feature frequency distributions.

Result: Under strong superposition, loss generically scales inversely with model dimension across broad frequency distributions due to geometric overlaps. Open-sourced LLMs operate in strong superposition regime with loss scaling like 1/model dimension, consistent with Chinchilla scaling laws.

Conclusion: Representation superposition is a central driver of neural scaling laws, providing insights into when scaling laws can be improved and when they will break down.

Abstract: The success of today’s large language models (LLMs) depends on the observation that larger models perform better. However, the origin of this neural scaling law, that loss decreases as a power law with model size, remains unclear. We propose that representation superposition, meaning that LLMs represent more features than they have dimensions, can be a key contributor to loss and cause neural scaling. Based on Anthropic’s toy model, we use weight decay to control the degree of superposition, allowing us to systematically study how loss scales with model size. When superposition is weak, the loss follows a power law only if data feature frequencies are power-law distributed. In contrast, under strong superposition, the loss generically scales inversely with model dimension across a broad class of frequency distributions, due to geometric overlaps between representation vectors. We confirmed that open-sourced LLMs operate in the strong superposition regime and have loss scaling like one over the model dimension, and that the Chinchilla scaling laws are also consistent with this behavior. Our results identify representation superposition as a central driver of neural scaling laws, providing insights into questions like when neural scaling laws can be improved and when they will break down.

Method: Faiss provides a toolkit of indexing methods and related primitives designed to handle vector search trade-offs, with optimized structure and interfaces for various use cases.

Result: The library offers comprehensive benchmarking of key features and demonstrates broad applicability across multiple domains through selected applications.

Conclusion: Faiss addresses the critical need for efficient vector similarity search in AI applications by providing a well-designed toolkit that balances trade-offs in vector search performance and functionality.

Abstract: Vector databases typically manage large collections of embedding vectors. Currently, AI applications are growing rapidly, and so is the number of embeddings that need to be stored and indexed. The Faiss library is dedicated to vector similarity search, a core functionality of vector databases. Faiss is a toolkit of indexing methods and related primitives used to search, cluster, compress and transform vectors. This paper describes the trade-off space of vector search and the design principles of Faiss in terms of structure, approach to optimization and interfacing. We benchmark key features of the library and discuss a few selected applications to highlight its broad applicability.

Method: Proposed CALM-PDE with continuous convolution-based encoder-decoder using epsilon-neighborhood-constrained kernel and adaptive query points. Learns to apply convolution operator to optimized query points for handling arbitrary discretizations.

Result: CALM-PDE is competitive with or outperforms existing baselines on diverse PDEs with both regular and irregular spatial domains, while offering significant improvements in memory usage and inference time compared to Transformer-based methods.

Conclusion: The proposed CALM-PDE framework successfully addresses the limitations of existing approaches by providing an efficient solution for arbitrarily discretized PDEs in compressed latent space, balancing performance with computational efficiency.

Abstract: Solving time-dependent Partial Differential Equations (PDEs) using a densely discretized spatial domain is a fundamental problem in various scientific and engineering disciplines, including modeling climate phenomena and fluid dynamics. However, performing these computations directly in the physical space often incurs significant computational costs. To address this issue, several neural surrogate models have been developed that operate in a compressed latent space to solve the PDE. While these approaches reduce computational complexity, they often use Transformer-based attention mechanisms to handle irregularly sampled domains, resulting in increased memory consumption. In contrast, convolutional neural networks allow memory-efficient encoding and decoding but are limited to regular discretizations. Motivated by these considerations, we propose CALM-PDE, a model class that efficiently solves arbitrarily discretized PDEs in a compressed latent space. We introduce a novel continuous convolution-based encoder-decoder architecture that uses an epsilon-neighborhood-constrained kernel and learns to apply the convolution operator to adaptive and optimized query points. We demonstrate the effectiveness of CALM-PDE on a diverse set of PDEs with both regularly and irregularly sampled spatial domains. CALM-PDE is competitive with or outperforms existing baseline methods while offering significant improvements in memory and inference time efficiency compared to Transformer-based methods.

Method: Disentangling rewards into task-specific and common-sense components, then using multi-task inverse reinforcement learning to learn the common-sense reward from expert demonstrations across multiple tasks.

Result: Single-task inverse reinforcement learning fails to learn useful reward functions, but multi-task inverse reinforcement learning successfully learns common-sense rewards that enable desired behaviors when training new agents.

Conclusion: Multi-task inverse reinforcement learning is an effective approach for learning useful common-sense reward functions that prevent reward hacking and ensure desired agent behaviors in complex environments.

Abstract: One of the challenges in applying reinforcement learning in a complex real-world environment lies in providing the agent with a sufficiently detailed reward function. Any misalignment between the reward and the desired behavior can result in unwanted outcomes. This may lead to issues like “reward hacking” where the agent maximizes rewards by unintended behavior. In this work, we propose to disentangle the reward into two distinct parts. A simple task-specific reward, outlining the particulars of the task at hand, and an unknown common-sense reward, indicating the expected behavior of the agent within the environment. We then explore how this common-sense reward can be learned from expert demonstrations. We first show that inverse reinforcement learning, even when it succeeds in training an agent, does not learn a useful reward function. That is, training a new agent with the learned reward does not impair the desired behaviors. We then demonstrate that this problem can be solved by training simultaneously on multiple tasks. That is, multi-task inverse reinforcement learning can be applied to learn a useful reward function.

Method: KDM encodes noisy inputs into an embedded space where a learned linear operator propagates them forward, followed by a decoder that reconstructs clean samples. This is based on Koopman theory which represents nonlinear dynamics linearly in a transformed space.

Result: KDM achieves highly competitive performance across standard offline distillation benchmarks, enabling single-step generation while preserving semantic fidelity.

Conclusion: The proposed KDM framework provides a principled distillation approach that leverages Koopman theory to efficiently distill diffusion models while maintaining semantic coherence in generated outputs.

Abstract: Diffusion-based generative models have demonstrated exceptional performance, yet their iterative sampling procedures remain computationally expensive. A prominent strategy to mitigate this cost is distillation, with offline distillation offering particular advantages in terms of efficiency, modularity, and flexibility. In this work, we identify two key observations that motivate a principled distillation framework: (1) while diffusion models have been viewed through the lens of dynamical systems theory, powerful and underexplored tools can be further leveraged; and (2) diffusion models inherently impose structured, semantically coherent trajectories in latent space. Building on these observations, we introduce the Koopman Distillation Model (KDM), a novel offline distillation approach grounded in Koopman theory - a classical framework for representing nonlinear dynamics linearly in a transformed space. KDM encodes noisy inputs into an embedded space where a learned linear operator propagates them forward, followed by a decoder that reconstructs clean samples. This enables single-step generation while preserving semantic fidelity. We provide theoretical justification for our approach: (1) under mild assumptions, the learned diffusion dynamics admit a finite-dimensional Koopman representation; and (2) proximity in the Koopman latent space correlates with semantic similarity in the generated outputs, allowing for effective trajectory alignment. KDM achieves highly competitive performance across standard offline distillation benchmarks.

Method: Log-NCDEs build on neural rough differential equations and use the Log-ODE method from rough paths theory to approximate CDE solutions, creating a more effective training approach.

Result: Log-NCDEs outperform NCDEs, NRDEs, linear recurrent unit, S5, and MAMBA on various multivariate time series datasets with up to 50,000 observations.

Conclusion: Log-NCDEs provide a novel, effective, and efficient method for training neural controlled differential equations, demonstrating superior performance on real-world time series modeling tasks.

Abstract: The vector field of a controlled differential equation (CDE) describes the relationship between a control path and the evolution of a solution path. Neural CDEs (NCDEs) treat time series data as observations from a control path, parameterise a CDE’s vector field using a neural network, and use the solution path as a continuously evolving hidden state. As their formulation makes them robust to irregular sampling rates, NCDEs are a powerful approach for modelling real-world data. Building on neural rough differential equations (NRDEs), we introduce Log-NCDEs, a novel, effective, and efficient method for training NCDEs. The core component of Log-NCDEs is the Log-ODE method, a tool from the study of rough paths for approximating a CDE’s solution. Log-NCDEs are shown to outperform NCDEs, NRDEs, the linear recurrent unit, S5, and MAMBA on a range of multivariate time series datasets with up to $50{,}000$ observations.

Method: Curated 161 problems where each requires (1) generating a specification matching ground-truth, and (2) generating a Lean implementation that provably satisfies the specification. All outputs are verified using Lean’s type checker.

Result: Evaluated few-shot and agentic approaches using state-of-the-art language models, finding that all methods struggle to achieve full verification, establishing CLEVER as a challenging benchmark.

Conclusion: CLEVER serves as a challenging frontier benchmark for program synthesis and formal reasoning, with all evaluation code and benchmark available online for community use.

Abstract: We introduce ${\rm C{\small LEVER}}$, a high-quality, curated benchmark of 161 problems for end-to-end verified code generation in Lean. Each problem consists of (1) the task of generating a specification that matches a held-out ground-truth specification, and (2) the task of generating a Lean implementation that provably satisfies this specification. Unlike prior benchmarks, ${\rm C{\small LEVER}}$ avoids test-case supervision, LLM-generated annotations, and specifications that leak implementation logic or allow vacuous solutions. All outputs are verified post-hoc using Lean’s type checker to ensure machine-checkable correctness. We use ${\rm C{\small LEVER}}$ to evaluate several few-shot and agentic approaches based on state-of-the-art language models. These methods all struggle to achieve full verification, establishing it as a challenging frontier benchmark for program synthesis and formal reasoning. Our benchmark can be found on GitHub(https://github.com/trishullab/clever) as well as HuggingFace(https://huggingface.co/datasets/amitayusht/clever). All our evaluation code is also available online(https://github.com/trishullab/clever-prover).

Method: Evaluated several machine learning models against two heuristic baselines, investigating the predictive capabilities of various node and channel features for channel balance interpolation.

Result: The machine learning model performed favorably in experimental evaluation, outperforming the equal split baseline (where both edges are assigned half of channel capacity) by 10%.

Conclusion: Machine learning approaches show promise for predicting Lightning Network channel balances, which can enhance network efficiency and pathfinding optimization.

Abstract: The Bitcoin Lightning Network is a Layer 2 payment protocol that addresses Bitcoin’s scalability by facilitating quick and cost effective transactions through payment channels. This research explores the feasibility of using machine learning models to interpolate channel balances within the network, which can be used for optimizing the network’s pathfinding algorithms. While there has been much exploration in balance probing and multipath payment protocols, predicting channel balances using solely node and channel features remains an uncharted area. This paper evaluates the performance of several machine learning models against two heuristic baselines and investigates the predictive capabilities of various features. Our model performs favorably in experimental evaluation, outperforming by 10% against an equal split baseline where both edges are assigned half of the channel capacity.

Method: Uses RL to co-train an LLM generator and a generative process-level verifier that evolves together, trained solely on outcome-level verification correctness rewards without explicit process-level annotations.

Result: Achieves SOTA results among 7B/8B-scale models: generator excels on five math benchmarks and four out-of-domain reasoning tasks, verifier leads on ProcessBench, with substantial improvements on most difficult math problems.

Conclusion: The generative RL-trained verifier exhibits improved robustness and superior generalization, fostering effective mutual reinforcement with the generator, demonstrating the effectiveness of concurrent training approach.

Abstract: Reinforcement learning (RL) has recently emerged as a compelling approach for enhancing the reasoning capabilities of large language models (LLMs), where an LLM generator serves as a policy guided by a verifier (reward model). However, current RL post-training methods for LLMs typically use verifiers that are fixed (rule-based or frozen pretrained) or trained discriminatively via supervised fine-tuning (SFT). Such designs are susceptible to reward hacking and generalize poorly beyond their training distributions. To overcome these limitations, we propose Tango, a novel framework that uses RL to concurrently train both an LLM generator and a verifier in an interleaved manner. A central innovation of Tango is its generative, process-level LLM verifier, which is trained via RL and co-evolves with the generator. Importantly, the verifier is trained solely based on outcome-level verification correctness rewards without requiring explicit process-level annotations. This generative RL-trained verifier exhibits improved robustness and superior generalization compared to deterministic or SFT-trained verifiers, fostering effective mutual reinforcement with the generator. Extensive experiments demonstrate that both components of Tango achieve state-of-the-art results among 7B/8B-scale models: the generator attains best-in-class performance across five competition-level math benchmarks and four challenging out-of-domain reasoning tasks, while the verifier leads on the ProcessBench dataset. Remarkably, both components exhibit particularly substantial improvements on the most difficult mathematical reasoning problems. Code is at: https://github.com/kaiwenzha/rl-tango.

Method: Proposes conditional congruence for assessing probabilistic fit and introduces Conditional Congruence Error (CCE) that uses conditional kernel mean embeddings to estimate the distance between learned predictive distributions and empirical conditional distributions.

Result: CCE exhibits four critical properties: correctness, monotonicity, reliability, and robustness, as demonstrated through high-dimensional regression tasks.

Conclusion: Conditional congruence and CCE provide a stronger framework for evaluating probabilistic alignment in neural networks, addressing limitations of traditional calibration metrics.

Abstract: While significant progress has been made in specifying neural networks capable of representing uncertainty, deep networks still often suffer from overconfidence and misaligned predictive distributions. Existing approaches for measuring this misalignment are primarily developed under the framework of calibration, with common metrics such as Expected Calibration Error (ECE). However, calibration can only provide a strictly marginal assessment of probabilistic alignment. Consequently, calibration metrics such as ECE are $\textit{distribution-wise}$ measures and cannot diagnose the $\textit{point-wise}$ reliability of individual inputs, which is important for real-world decision-making. We propose a stronger condition, which we term $\textit{conditional congruence}$, for assessing probabilistic fit. We also introduce a metric, Conditional Congruence Error (CCE), that uses conditional kernel mean embeddings to estimate the distance, at any point, between the learned predictive distribution and the empirical, conditional distribution in a dataset. We perform several high dimensional regression tasks and show that CCE exhibits four critical properties: $\textit{correctness}$, $\textit{monotonicity}$, $\textit{reliability}$, and $\textit{robustness}$.

Method: Sample the agent’s learning trajectory during RL training, prompt LLM to analyze current policy learning status and generate probability distribution for future exploration, periodically update the distribution to create a task-specific stochastic process.

Result: Achieved average performance improvement up to 37.27% on Atari and MuJoCo benchmarks, compatible with various RL algorithms including DQN series, DDPG, TD3 and their variants.

Conclusion: LLM-Explorer successfully enhances RL policy exploration by adaptively generating task-specific strategies using LLMs’ analytical capabilities, serving as a plug-in module for multiple RL algorithms.

Abstract: Policy exploration is critical in reinforcement learning (RL), where existing approaches include greedy, Gaussian process, etc. However, these approaches utilize preset stochastic processes and are indiscriminately applied in all kinds of RL tasks without considering task-specific features that influence policy exploration. Moreover, during RL training, the evolution of such stochastic processes is rigid, which typically only incorporates a decay in the variance, failing to adjust flexibly according to the agent’s real-time learning status. Inspired by the analyzing and reasoning capability of large language models (LLMs), we design LLM-Explorer to adaptively generate task-specific exploration strategies with LLMs, enhancing the policy exploration in RL. In our design, we sample the learning trajectory of the agent during the RL training in a given task and prompt the LLM to analyze the agent’s current policy learning status and then generate a probability distribution for future policy exploration. Updating the probability distribution periodically, we derive a stochastic process specialized for the particular task and dynamically adjusted to adapt to the learning process. Our design is a plug-in module compatible with various widely applied RL algorithms, including the DQN series, DDPG, TD3, and any possible variants developed based on them. Through extensive experiments on the Atari and MuJoCo benchmarks, we demonstrate LLM-Explorer’s capability to enhance RL policy exploration, achieving an average performance improvement up to 37.27%. Our code is open-source at https://github.com/tsinghua-fib-lab/LLM-Explorer for reproducibility.

Method: A framework that learns surrogate linear constraints through active learning, using linear separators on labeled points and selecting new points via sampling strategies and 0-1 integer linear programming. The proposed improvements include a mixed-integer quadratic programming sampling strategy and a convex optimization-inspired linear separation method.

Result: Experimental evaluation on classical and realistic problem variants shows how different linear separation methods and sampling strategies affect solution quality metrics including objective value, dual bounds, and running time.

Conclusion: The proposed improvements to linear separation and sampling strategies enhance the performance of the active learning framework for solving combinatorial optimization with unknown constraints under oracle budget limitations.

Abstract: We consider solving a combinatorial optimization problem with unknown knapsack constraints using a membership oracle for each unknown constraint such that, given a solution, the oracle determines whether the constraint is satisfied or not with absolute certainty. The goal of the decision maker is to find the best possible solution subject to a budget on the number of oracle calls. Inspired by active learning for binary classification based on Support Vector Machines (SVMs), we devise a framework to solve the problem by learning and exploiting surrogate linear constraints. The framework includes training linear separators on the labeled points and selecting new points to be labeled, which is achieved by applying a sampling strategy and solving a 0-1 integer linear program. Following the active learning literature, a natural choice would be SVM as a linear classifier and the information-based sampling strategy known as simple margin, for each unknown constraint. We improve on both sides: we propose an alternative sampling strategy based on mixed-integer quadratic programming and a linear separation method inspired by an algorithm for convex optimization in the oracle model. We conduct experiments on classical problems and variants inspired by realistic applications to show how different linear separation methods and sampling strategies influence the quality of the results in terms of several metrics including objective value, dual bound and running time.

Method: Prove a generalization bound for policy distillation, then empirically validate insights: train ensemble of distilled policies and use diverse training data.

Result: Theory provides practical insights that are empirically verified; ensemble of distilled policies generalizes significantly better than original agent.

Conclusion: Policy distillation with ensembles and diverse training data enhances generalization in zero-shot policy transfer settings.

Abstract: In the zero-shot policy transfer setting in reinforcement learning, the goal is to train an agent on a fixed set of training environments so that it can generalise to similar, but unseen, testing environments. Previous work has shown that policy distillation after training can sometimes produce a policy that outperforms the original in the testing environments. However, it is not yet entirely clear why that is, or what data should be used to distil the policy. In this paper, we prove, under certain assumptions, a generalisation bound for policy distillation after training. The theory provides two practical insights: for improved generalisation, you should 1) train an ensemble of distilled policies, and 2) distil it on as much data from the training environments as possible. We empirically verify that these insights hold in more general settings, when the assumptions required for the theory no longer hold. Finally, we demonstrate that an ensemble of policies distilled on a diverse dataset can generalise significantly better than the original agent.

Method: Borrow top-p sampling (nucleus sampling) to sparse attention to achieve adaptive budgeting, creating a framework that can be applied to any existing sparse attention algorithm without accuracy loss.

Result: Twilight adaptively prunes up to 98% of redundant tokens, achieving 15.4× acceleration in self-attention operations and 3.9× acceleration in end-to-end per token latency for long-context LLM decoding.

Conclusion: The Twilight framework successfully enables adaptive sparsity in attention mechanisms, providing significant performance improvements while maintaining accuracy across varying real-world scenarios.

Abstract: Leveraging attention sparsity to accelerate long-context large language models (LLMs) has been a hot research topic. However, current algorithms such as sparse attention or key-value (KV) cache compression tend to use a fixed budget, which presents a significant challenge during deployment because it fails to account for the dynamic nature of real-world scenarios, where the optimal balance between accuracy and efficiency can vary greatly. In this paper, we find that borrowing top-$p$ sampling (nucleus sampling) to sparse attention can surprisingly achieve adaptive budgeting. Based on this, we propose Twilight, a framework to bring adaptive sparsity to any existing sparse attention algorithm without sacrificing their accuracy. Empirical results show that Twilight can adaptively prune at most 98% of redundant tokens, leading to $15.4\times$ acceleration in self-attention operations and $3.9\times$ acceleration in end-to-end per token latency in long context LLM decoding.

Method: Uses a transformer as a decision-making algorithm trained with a novel pre-training approach that learns near-optimal policies in-context by leveraging shared structure across tasks, without requiring access to optimal actions.

Result: The proposed solution demonstrates strong performance across various structured bandit problems, quickly identifying expected rewards on unseen test tasks and enabling effective exploration that outperforms the demonstrator.

Conclusion: The transformer-based approach successfully learns to exploit shared structure in multi-task bandit problems, achieving superior performance on unseen test tasks without requiring privileged information, making it a general and effective solution for structured bandit learning.

Abstract: We study learning to learn for the multi-task structured bandit problem where the goal is to learn a near-optimal algorithm that minimizes cumulative regret. The tasks share a common structure and an algorithm should exploit the shared structure to minimize the cumulative regret for an unseen but related test task. We use a transformer as a decision-making algorithm to learn this shared structure from data collected by a demonstrator on a set of training task instances. Our objective is to devise a training procedure such that the transformer will learn to outperform the demonstrator’s learning algorithm on unseen test task instances. Prior work on pretraining decision transformers either requires privileged information like access to optimal arms or cannot outperform the demonstrator. Going beyond these approaches, we introduce a pre-training approach that trains a transformer network to learn a near-optimal policy in-context. This approach leverages the shared structure across tasks, does not require access to optimal actions, and can outperform the demonstrator. We validate these claims over a wide variety of structured bandit problems to show that our proposed solution is general and can quickly identify expected rewards on unseen test tasks to support effective exploration.

Method: Proposes Disagreement-Aware Confidence Alignment (DACA), which selectively uses only agreement examples between PLM and PoLM for temperature scaling calibration, effectively decoupling the influence of disagreement examples that cause under-confidence issues.

Result: Extensive experiments show DACA improves average ECE of open-sourced and API-based LLMs (including GPT-4o) by up to 15.08% on common benchmarks.

Conclusion: DACA effectively addresses the over-confidence problem in post-trained language models through disagreement-aware calibration, providing a practical unsupervised solution for confidence alignment without requiring labeled data.

Abstract: Post-training of large language models is essential for adapting pre-trained language models (PLMs) to align with human preferences and downstream tasks. While PLMs typically exhibit well-calibrated confidence, post-trained language models (PoLMs) often suffer from over-confidence, assigning high confidence to both correct and incorrect outputs, which can undermine reliability in critical applications. A major obstacle in calibrating PoLMs is the scarcity of labeled data for individual downstream tasks. To address this, we propose Disagreement-Aware Confidence Alignment (DACA), a novel unsupervised method to optimize the parameters (e.g., temperature $\tau$) in post-hoc confidence calibration. Our method is motivated by the under-confidence issue caused by prediction disagreement between the PLM and PoLM while aligning their confidence via temperature scaling. Theoretically, the PLM’s confidence underestimates PoLM’s prediction accuracy on disagreement examples, causing a larger $\tau$ and producing under-confident predictions. DACA mitigates this by selectively using only agreement examples for calibration, effectively decoupling the influence of disagreement. In this manner, our method avoids an overly large $\tau$ in temperature scaling caused by disagreement examples, improving calibration performance. Extensive experiments demonstrate the effectiveness of our method, improving the average ECE of open-sourced and API-based LLMs (e.g. GPT-4o) by up to 15.08$%$ on common benchmarks.

Method: Used a framework for studying length generalization with retrieval and copying tasks inspired by Liu et al. [2024a], conducted empirical analysis of induction vs anti-induction asymmetry, and performed mechanistic analysis of transformer circuits.

Result: Found induction-versus-anti-induction asymmetry where pretrained models are better at retrieving tokens to the right (induction) than to the left (anti-induction). This asymmetry disappears with targeted fine-tuning when length-generalization is theoretically guaranteed.

Conclusion: Pretraining selectively enhances certain transformer capabilities but does not overcome fundamental length-generalization limits, highlighting reliability risks in practical applications.

Abstract: Transformers have theoretical limitations in modeling certain sequence-to-sequence tasks, yet it remains largely unclear if these limitations play a role in large-scale pretrained LLMs, or whether LLMs might effectively overcome these constraints in practice due to the scale of both the models themselves and their pretraining data. We explore how these architectural constraints manifest after pretraining, by studying a family of $\textit{retrieval}$ and $\textit{copying}$ tasks inspired by Liu et al. [2024a]. We use a recently proposed framework for studying length generalization [Huang et al., 2025] to provide guarantees for each of our settings. Empirically, we observe an $\textit{induction-versus-anti-induction}$ asymmetry, where pretrained models are better at retrieving tokens to the right (induction) rather than the left (anti-induction) of a query token. This asymmetry disappears upon targeted fine-tuning if length-generalization is guaranteed by theory. Mechanistic analysis reveals that this asymmetry is connected to the differences in the strength of induction versus anti-induction circuits within pretrained transformers. We validate our findings through practical experiments on real-world tasks demonstrating reliability risks. Our results highlight that pretraining selectively enhances certain transformer capabilities, but does not overcome fundamental length-generalization limits.

Method: Proposes a hierarchical neural architecture search with a hierarchical search space that incorporates various forecasting architecture types.

Result: The approach can search for lightweight high-performing forecasting architectures across different forecasting tasks.

Conclusion: The hierarchical NAS approach effectively combines forecasting modules to create optimized architectures for time series forecasting.

Abstract: The rapid development of time series forecasting research has brought many deep learning-based modules in this field. However, despite the increasing amount of new forecasting architectures, it is still unclear if we have leveraged the full potential of these existing modules within a properly designed architecture. In this work, we propose a novel hierarchical neural architecture search approach for time series forecasting tasks. With the design of a hierarchical search space, we incorporate many architecture types designed for forecasting tasks and allow for the efficient combination of different forecasting architecture modules. Results on long-term-time-series-forecasting tasks show that our approach can search for lightweight high-performing forecasting architectures across different forecasting tasks.

Method: MxDs expand pre-trained dense layers into tens of thousands of specialized sublayers using tensor factorization, creating sparsely activating linear transformations with full-rank weights that generalize MLPs and Gated Linear Units.

Result: MxDs significantly outperform state-of-the-art methods like Transcoders on the sparsity-accuracy frontier in language models up to 3B parameters, while learning similarly specialized natural language features.

Conclusion: MxDs provide a promising new approach for designing interpretable yet faithful decompositions of MLPs through layer-level sparsity, overcoming the accuracy limitations of previous neuron-level methods.

Abstract: Multilayer perceptrons (MLPs) are an integral part of large language models, yet their dense representations render them difficult to understand, edit, and steer. Recent methods learn interpretable approximations via neuron-level sparsity, yet fail to faithfully reconstruct the original mapping–significantly increasing model’s next-token cross-entropy loss. In this paper, we advocate for moving to layer-level sparsity to overcome the accuracy trade-off in sparse layer approximation. Under this paradigm, we introduce Mixture of Decoders (MxDs). MxDs generalize MLPs and Gated Linear Units, expanding pre-trained dense layers into tens of thousands of specialized sublayers. Through a flexible form of tensor factorization, each sparsely activating MxD sublayer implements a linear transformation with full-rank weights–preserving the original decoders’ expressive capacity even under heavy sparsity. Experimentally, we show that MxDs significantly outperform state-of-the-art methods (e.g., Transcoders) on the sparsity-accuracy frontier in language models with up to 3B parameters. Further evaluations on sparse probing and feature steering demonstrate that MxDs learn similarly specialized features of natural language–opening up a promising new avenue for designing interpretable yet faithful decompositions. Our code is included at: https://github.com/james-oldfield/MxD/.

Method: Two-stage approach: first approximates models using compact Fourier representations (exact for trees, approximate for black-box), then uses closed-form formula for exact SHAP computation via Fourier representation that linearizes computation into simple summation.

Result: Achieves significant speedups over existing methods and enables amortized SHAP value computation with tunable trade-off between efficiency and precision.

Conclusion: Proposed method provides an efficient and scalable approach for SHAP value computation that can be parallelized and offers flexibility in balancing computational efficiency with accuracy.

Abstract: SHAP (SHapley Additive exPlanations) values are a widely used method for local feature attribution in interpretable and explainable AI. We propose an efficient two-stage algorithm for computing SHAP values in both black-box setting and tree-based models. Motivated by spectral bias in real-world predictors, we first approximate models using compact Fourier representations, exactly for trees and approximately for black-box models. In the second stage, we introduce a closed-form formula for {\em exactly} computing SHAP values using the Fourier representation, that ``linearizes’’ the computation into a simple summation and is amenable to parallelization. As the Fourier approximation is computed only once, our method enables amortized SHAP value computation, achieving significant speedups over existing methods and a tunable trade-off between efficiency and precision.

Method: Created AssistedDS benchmark with synthetic datasets (known generative mechanisms) and real-world Kaggle competitions, accompanied by curated helpful and adversarial documents about data cleaning, feature engineering, and model selection.

Result: LLMs frequently exhibit uncritical adoption of provided information, significantly impairing predictive performance with adversarial content; helpful guidance often fails to counteract adversarial influence; LLMs make errors in time-series data handling, feature engineering across folds, and categorical variable interpretation.

Conclusion: There’s a substantial gap in current models’ ability to critically evaluate and leverage expert knowledge, highlighting the need for more robust, knowledge-aware automated data science systems.

Abstract: Large language models (LLMs) have advanced the automation of data science workflows. Yet it remains unclear whether they can critically leverage external domain knowledge as human data scientists do in practice. To answer this question, we introduce AssistedDS (Assisted Data Science), a benchmark designed to systematically evaluate how LLMs handle domain knowledge in tabular prediction tasks. AssistedDS features both synthetic datasets with explicitly known generative mechanisms and real-world Kaggle competitions, each accompanied by curated bundles of helpful and adversarial documents. These documents provide domain-specific insights into data cleaning, feature engineering, and model selection. We assess state-of-the-art LLMs on their ability to discern and apply beneficial versus harmful domain knowledge, evaluating submission validity, information recall, and predictive performance. Our results demonstrate three key findings: (1) LLMs frequently exhibit an uncritical adoption of provided information, significantly impairing their predictive performance when adversarial content is introduced, (2) helpful guidance is often insufficient to counteract the negative influence of adversarial information, and (3) in Kaggle datasets, LLMs often make errors in handling time-series data, applying consistent feature engineering across different folds, and interpreting categorical variables correctly. These findings highlight a substantial gap in current models’ ability to critically evaluate and leverage expert knowledge, underscoring an essential research direction for developing more robust, knowledge-aware automated data science systems. Our data and code are publicly available here: https://github.com/jeremyxianx/Assisted-DS

Method: Introduces a generic PEFT-based meta-learning framework that learns adaptable parameters, specifically analyzing linear models using LoRA with theoretical performance guarantees.

Result: Theoretical analysis shows standard retraining is provably suboptimal, and experiments on synthetic data, vision, and language tasks demonstrate significant performance improvements over conventional approaches.

Conclusion: Meta-learning during retraining provides substantial benefits for creating adaptable foundation models, with both theoretical foundations and empirical validation across multiple domains.

Abstract: The power of foundation models (FMs) lies in their capacity to learn highly expressive representations that can be adapted to a broad spectrum of tasks. However, these pretrained models require additional training stages to become effective for downstream applications. In the multi-task setting, prior works have shown empirically that specific meta-learning approaches for preparing a model for future adaptation through parameter-efficient fine-tuning (PEFT) can outperform standard retraining methods, but the mechanism of the benefits of meta-learning has been largely unexplored. We introduce a framework for generic PEFT-based meta-learning to learn a model that can easily adapt to unseen tasks. For linear models using LoRA, we show that standard retraining is provably suboptimal for finding an adaptable set of parameters and provide strict performance guarantees for our proposed method. We verify these theoretical insights through experiments on synthetic data as well as real-data vision and language tasks. We observe significant performance benefits using a simple implementation of our proposed meta-learning scheme during retraining relative to the conventional approach.

Method: After each task, P&M constructs a new model by forming a convex combination of previous and new task models. It uses a regularization term involving task vectors and Hessian matrix, efficiently approximated via second-order symmetric finite differences and stochastic perturbation along task vector direction.

Result: The proposed approach achieves state-of-the-art performance on several continual learning benchmark datasets.

Conclusion: Integrating model merging with regularization via efficient perturbation approximation effectively mitigates catastrophic forgetting in continual learning while maintaining high performance.

Abstract: Continual Learning (CL) aims to enable models to continuously acquire new knowledge from a sequence of tasks with avoiding the forgetting of learned information. However, existing CL methods only rely on the parameters of the most recent task for inference, which makes them susceptible to catastrophic forgetting. Inspired by the recent success of model merging techniques, we propose \textbf{Perturb-and-Merge (P&M)}, a novel continual learning framework that integrates model merging into the CL paradigm to mitigate forgetting. Specifically, after training on each task, P&M constructs a new model by forming a convex combination of the previous model and the newly trained task-specific model. Through theoretical analysis, We minimize the total loss increase across all tasks and derive a closed-form solution for the merging coefficient under mild assumptions. To further improve the performance of the merged model, we observe that the degradation introduced during merging can be alleviated by a regularization term composed of the task vector and the Hessian matrix of the loss function. Interestingly, we show that this term can be efficiently approximated using second-order symmetric finite differences, and a stochastic perturbation strategy along the task vector direction is accordingly devised which incurs no additional forward or backward passes while providing an effective approximation of the regularization term. Finally, we combine P&M with LoRA, a parameter-efficient fine-tuning method, to reduce memory overhead. Our proposed approach achieves state-of-the-art performance on several continual learning benchmark datasets. The code is available at https://github.com/qhmiao/P-M-for-Continual-Learning.

Method: Proposes ReDit method that dithers discrete reward signals by adding simple random noise, providing continuous exploratory gradients and introducing stochasticity to encourage policy exploration.

Result: ReDit achieves comparable performance to vanilla GRPO with only ~10% training steps, and shows 4% performance improvement when trained for similar duration. Visualizations confirm gradient issue mitigation.

Conclusion: ReDit effectively addresses discrete reward limitations through reward dithering, enabling smoother optimization, faster convergence, and better performance while maintaining theoretical advantages.

Abstract: DeepSeek-R1 has successfully enhanced Large Language Model (LLM) reasoning capabilities through its rule-based reward system. While it’s a ‘‘perfect’’ reward system that effectively mitigates reward hacking, such reward functions are often discrete. Our experimental observations suggest that discrete rewards can lead to gradient anomaly, unstable optimization, and slow convergence. To address this issue, we propose ReDit (Reward Dithering), a method that dithers the discrete reward signal by adding simple random noise. With this perturbed reward, exploratory gradients are continuously provided throughout the learning process, enabling smoother gradient updates and accelerating convergence. The injected noise also introduces stochasticity into flat reward regions, encouraging the model to explore novel policies and escape local optima. Experiments across diverse tasks demonstrate the effectiveness and efficiency of ReDit. On average, ReDit achieves performance comparable to vanilla GRPO with only approximately 10% the training steps, and furthermore, still exhibits a 4% performance improvement over vanilla GRPO when trained for a similar duration. Visualizations confirm significant mitigation of gradient issues with ReDit. Moreover, theoretical analyses are provided to further validate these advantages.

Method: Adaptively adjusts class proportions within batches during training, focusing on classes with higher error rates, rather than using static class proportions like classical methods.

Result: Neural networks trained with learn2mix converge faster than existing approaches, achieving improved results for classification, regression, and reconstruction tasks under limited training resources and with imbalanced classes.

Conclusion: The learn2mix strategy enables faster convergence through adaptive class proportion adjustment, with empirical results supported by theoretical analysis.

Abstract: Accelerating model convergence in resource-constrained environments is essential for fast and efficient neural network training. This work presents learn2mix, a new training strategy that adaptively adjusts class proportions within batches, focusing on classes with higher error rates. Unlike classical training methods that use static class proportions, learn2mix continually adapts class proportions during training, leading to faster convergence. Empirical evaluations on benchmark datasets show that neural networks trained with learn2mix converge faster than those trained with existing approaches, achieving improved results for classification, regression, and reconstruction tasks under limited training resources and with imbalanced classes. Our empirical findings are supported by theoretical analysis.

Method: Proposes defining unlearning as finding the minimum-complexity interpolator over retained data. Uses a regularized objective with perturbations constrained to be orthogonal to model gradients on the retained set at the original solution.

Result: The proposed framework provides exact and approximate unlearning guarantees for different model classes and outperforms existing baselines across various unlearning experiments.

Conclusion: The paper successfully addresses machine unlearning in overparameterized settings by introducing a new definition based on minimum-complexity interpolation and developing an effective algorithm that works with vanishing gradients.

Abstract: Machine unlearning algorithms aim to remove the influence of specific training samples, ideally recovering the model that would have resulted from training on the remaining data alone. We study unlearning in the overparameterized setting, where many models interpolate the data, and defining the solution as any loss minimizer over the retained set$\unicode{x2013}$as in prior work in the underparameterized setting$\unicode{x2013}$is inadequate, since the original model may already interpolate the retained data and satisfy this condition. In this regime, loss gradients vanish, rendering prior methods based on gradient perturbations ineffective, motivating both new unlearning definitions and algorithms. For this setting, we define the unlearning solution as the minimum-complexity interpolator over the retained data and propose a new algorithmic framework that only requires access to model gradients on the retained set at the original solution. We minimize a regularized objective over perturbations constrained to be orthogonal to these model gradients, a first-order relaxation of the interpolation condition. For different model classes, we provide exact and approximate unlearning guarantees and demonstrate that an implementation of our framework outperforms existing baselines across various unlearning experiments.

Method: Three-stage approach: 1) Pre-train language model on antibody sequences, 2) Train diffusion model for joint sequence-structure optimization, 3) Multi-objective alignment using extended AbDPO with iterative learning and temperature scaling.

Result: Achieves high stability and efficiency in producing better Pareto front of antibody designs compared to baselines, generating nature-like antibodies with high binding affinity.

Conclusion: The proposed framework successfully generates functional antibodies through sequence-structure co-design and multi-objective optimization, outperforming previous methods in binding affinity and design quality.

Abstract: We present a three-stage framework for training deep learning models specializing in antibody sequence-structure co-design. We first pre-train a language model using millions of antibody sequence data. Then, we employ the learned representations to guide the training of a diffusion model for joint optimization over both sequence and structure of antibodies. During the final alignment stage, we optimize the model to favor antibodies with low repulsion and high attraction to the antigen binding site, enhancing the rationality and functionality of the designs. To mitigate conflicting energy preferences, we extend AbDPO (Antibody Direct Preference Optimization) to guide the model toward Pareto optimality under multiple energy-based alignment objectives. Furthermore, we adopt an iterative learning paradigm with temperature scaling, enabling the model to benefit from diverse online datasets without requiring additional data. In practice, our proposed methods achieve high stability and efficiency in producing a better Pareto front of antibody designs compared to top samples generated by baselines and previous alignment techniques. Through extensive experiments, we showcase the superior performance of our methods in generating nature-like antibodies with high binding affinity.

Method: Two-stage approach: 1) Derive information-theoretic prior by evaluating patch sequence compressibility, 2) Learn permutation policy using REINFORCE with Plackett-Luce policy for efficient learning in combinatorial permutation space.

Result: REOrder improves top-1 accuracy over row-major ordering by up to 3.01% on ImageNet-1K and 13.35% on Functional Map of the World.

Conclusion: Patch ordering significantly impacts transformer performance, and REOrder provides an effective framework for discovering task-optimal orderings that substantially improve accuracy.

Abstract: Sequence models such as transformers require inputs to be represented as one-dimensional sequences. In vision, this typically involves flattening images using a fixed row-major (raster-scan) order. While full self-attention is permutation-equivariant, modern long-sequence transformers increasingly rely on architectural approximations that break this invariance and introduce sensitivity to patch ordering. We show that patch order significantly affects model performance in such settings, with simple alternatives like column-major or Hilbert curves yielding notable accuracy shifts. Motivated by this, we propose REOrder, a two-stage framework for discovering task-optimal patch orderings. First, we derive an information-theoretic prior by evaluating the compressibility of various patch sequences. Then, we learn a policy over permutations by optimizing a Plackett-Luce policy using REINFORCE. This approach enables efficient learning in a combinatorial permutation space. REOrder improves top-1 accuracy over row-major ordering on ImageNet-1K by up to 3.01% and Functional Map of the World by 13.35%.

Method: FlyLoRA introduces rank-wise expert activation in the up-projection matrix and an implicit router that unifies expert routing and down-projection using a frozen sparse random projection matrix instead of traditional dense trainable versions.

Result: Extensive experiments across general knowledge understanding, scientific question answering, mathematical reasoning, and code generation demonstrate consistent performance improvements over existing methods.

Conclusion: FlyLoRA resolves the trade-off between intra-task decorrelation and computational efficiency while inherently mitigating inter-task interference, and highlights how biological structures can inspire AI innovations.

Abstract: Low-Rank Adaptation (LoRA) is a widely used parameter-efficient fine-tuning method for foundation models, but it suffers from parameter interference, resulting in suboptimal performance. Although Mixture-of-Experts (MoE)-based LoRA variants show promise in mitigating intra-task correlations in single-task instruction tuning, they introduce additional router parameters and remain ineffective in multi-task model merging where inter-task interference arises. Inspired by the fly olfactory circuit, we propose FlyLoRA, an implicit MoE-based LoRA variant that introduces: (1) rank-wise expert activation in the up-projection matrix, and (2) an implicit router that unifies expert routing and down-projection, where a frozen sparse random projection matrix replaces the traditional dense trainable version. This design resolves the trade-off between intra-task decorrelation and computational efficiency by eliminating the need for an explicit router, while inherently mitigating inter-task interference due to the orthogonality property of random matrices. Extensive experiments across four domains – general knowledge understanding, scientific question answering, mathematical reasoning, and code generation – demonstrate consistent performance improvements over existing methods. Beyond empirical gains, FlyLoRA highlights how biological structures can inspire innovations in AI technologies. Code is available at https://github.com/gfyddha/FlyLoRA.

Method: CAPE models epidemic dynamics as mixtures of latent population states (compartmental prototypes) discovered from surveillance data. It combines self-supervised pre-training with epidemic-aware regularizers to align prototypes with epidemiological semantics.

Result: On a comprehensive benchmark spanning 17 diseases and 50+ regions, CAPE significantly outperforms strong baselines in zero-shot, few-shot, and full-shot forecasting scenarios.

Conclusion: CAPE represents a principled step toward pre-trained epidemic models that are both transferable and epidemiologically grounded, enabling robust forecasting across diverse disease outbreaks.

Abstract: Accurate epidemic forecasting is crucial for outbreak preparedness, but existing data-driven models are often brittle. Typically trained on a single pathogen, they struggle with data scarcity during new outbreaks and fail under distribution shifts caused by viral evolution or interventions. However, decades of surveillance data from diverse diseases offer an untapped source of transferable knowledge. To leverage the collective lessons from history, we propose CAPE, the first open-source pre-trained model for epidemic forecasting. Unlike existing time series foundation models that overlook epidemiological challenges, CAPE models epidemic dynamics as mixtures of latent population states, termed compartmental prototypes. It discovers a flexible dictionary of compartment prototypes directly from surveillance data, enabling each outbreak to be expressed as a time-varying mixture that links observed infections to latent population states. To promote robust generalization, CAPE combines self-supervised pre-training objectives with lightweight epidemic-aware regularizers that align the learned prototypes with epidemiological semantics. On a comprehensive benchmark spanning 17 diseases and 50+ regions, CAPE significantly outperforms strong baselines in zero-shot, few-shot, and full-shot forecasting. This work represents a principled step toward pre-trained epidemic models that are both transferable and epidemiologically grounded.

Method: The authors analyze existing differentiable safeguards, adapt them through modified mappings and gradient formulations, and integrate them into a state-of-the-art learning algorithm and differentiable simulation.

Result: Numerical experiments on three control tasks show that safeguarded training can be achieved without compromising performance, demonstrating effective safety integration.

Conclusion: The work successfully bridges the gap in safeguarding analytic gradient-based RL, providing the first effective approach that maintains performance while ensuring safety during training.

Abstract: The deployment of autonomous robots in safety-critical applications requires safety guarantees. Provably safe reinforcement learning is an active field of research that aims to provide such guarantees using safeguards. These safeguards should be integrated during training to reduce the sim-to-real gap. While there are several approaches for safeguarding sampling-based reinforcement learning, analytic gradient-based reinforcement learning often achieves superior performance from fewer environment interactions. However, there is no safeguarding approach for this learning paradigm yet. Our work addresses this gap by developing the first effective safeguard for analytic gradient-based reinforcement learning. We analyse existing, differentiable safeguards, adapt them through modified mappings and gradient formulations, and integrate them into a state-of-the-art learning algorithm and a differentiable simulation. Using numerical experiments on three control tasks, we evaluate how different safeguards affect learning. The results demonstrate safeguarded training without compromising performance. Additional visuals are provided at \href{https://timwalter.github.io/safe-agb-rl.github.io}{timwalter.github.io/safe-agb-rl.github.io}.

Method: The authors formalize model reconstruction risks and complexity, focusing on oracle queries needed to infer prediction functions. They use competitive analysis to evaluate attack efficiency and introduce novel reconstruction algorithms specifically designed for additive decision tree models (decision trees, gradient boosting, random forests).

Result: The proposed reconstruction algorithms achieve provably perfect fidelity while demonstrating strong anytime performance. The framework provides theoretical bounds on query complexity for extracting tree-based models.

Conclusion: The research offers new insights into security vulnerabilities of tree-based model deployment in MLaaS environments, establishing a foundational framework for understanding and evaluating model extraction attack efficiency.

Abstract: The advent of Machine Learning as a Service (MLaaS) has heightened the trade-off between model explainability and security. In particular, explainability techniques, such as counterfactual explanations, inadvertently increase the risk of model extraction attacks, enabling unauthorized replication of proprietary models. In this paper, we formalize and characterize the risks and inherent complexity of model reconstruction, focusing on the “oracle’’ queries required for faithfully inferring the underlying prediction function. We present the first formal analysis of model extraction attacks through the lens of competitive analysis, establishing a foundational framework to evaluate their efficiency. Focusing on models based on additive decision trees (e.g., decision trees, gradient boosting, and random forests), we introduce novel reconstruction algorithms that achieve provably perfect fidelity while demonstrating strong anytime performance. Our framework provides theoretical bounds on the query complexity for extracting tree-based model, offering new insights into the security vulnerabilities of their deployment.

Method: Defines discrete flow over sequences using edit operations within a Continuous-time Markov Chain over sequence space. Uses expanded state space with auxiliary variables for efficient and tractable training.

Result: Outperforms both autoregressive and mask models on image captioning, and significantly outperforms mask construction in text and code generation tasks.

Conclusion: Edit Flows successfully enables flexible, position-relative generation that better aligns with sequence data structure, making non-autoregressive modeling more effective for variable-length sequence generation.

Abstract: Autoregressive generative models naturally generate variable-length sequences, while non-autoregressive models struggle, often imposing rigid, token-wise structures. We propose Edit Flows, a non-autoregressive model that overcomes these limitations by defining a discrete flow over sequences through edit operations$\unicode{x2013}$insertions, deletions, and substitutions. By modeling these operations within a Continuous-time Markov Chain over the sequence space, Edit Flows enable flexible, position-relative generation that aligns more closely with the structure of sequence data. Our training method leverages an expanded state space with auxiliary variables, making the learning process efficient and tractable. Empirical results show that Edit Flows outperforms both autoregressive and mask models on image captioning and significantly outperforms the mask construction in text and code generation.

Method: Developed WENDy-MLE algorithm that approximates maximum likelihood estimator via local non-convex optimization using analytic expressions for likelihood function and its first/second order derivatives. Extended framework to handle multiplicative log-normal noise.

Result: WENDy-MLE shows better accuracy, substantially larger domain of convergence, and often faster performance than other weak form methods and conventional output error least squares method across benchmark ODE systems.

Conclusion: The extended WENDy framework successfully handles nonlinear-in-parameters ODEs with improved performance characteristics, making it applicable to a broader class of differential equation systems.

Abstract: The Weak-form Estimation of Non-linear Dynamics (WENDy) framework is a recently developed approach for parameter estimation and inference of systems of ordinary differential equations (ODEs). Prior work demonstrated WENDy to be robust, computationally efficient, and accurate, but only works for ODEs which are linear-in-parameters. In this work, we derive a novel extension to accommodate systems of a more general class of ODEs that are nonlinear-in-parameters. Our new WENDy-MLE algorithm approximates a maximum likelihood estimator via local non-convex optimization methods. This is made possible by the availability of analytic expressions for the likelihood function and its first and second order derivatives. WENDy-MLE has better accuracy, a substantially larger domain of convergence, and is often faster than other weak form methods and the conventional output error least squares method. Moreover, we extend the framework to accommodate data corrupted by multiplicative log-normal noise. The WENDy.jl algorithm is efficiently implemented in Julia. In order to demonstrate the practical benefits of our approach, we present extensive numerical results comparing our method, other weak form methods, and output error least squares on a suite of benchmark systems of ODEs in terms of accuracy, precision, bias, and coverage.

Method: Adapts language model watermarking techniques, introduces custom tokenizer-detokenizer finetuning to improve reverse cycle-consistency, and adds a complementary watermark synchronization layer.

Result: Enables reliable and robust watermark detection with theoretically grounded p-values, resilient to common image transformations, neural compression, and removal attacks.

Conclusion: Successfully demonstrates the first effective token-level watermarking approach for autoregressive image generation models with practical robustness guarantees.

Abstract: Watermarking the outputs of generative models has emerged as a promising approach for tracking their provenance. Despite significant interest in autoregressive image generation models and their potential for misuse, no prior work has attempted to watermark their outputs at the token level. In this work, we present the first such approach by adapting language model watermarking techniques to this setting. We identify a key challenge: the lack of reverse cycle-consistency (RCC), wherein re-tokenizing generated image tokens significantly alters the token sequence, effectively erasing the watermark. To address this and to make our method robust to common image transformations, neural compression, and removal attacks, we introduce (i) a custom tokenizer-detokenizer finetuning procedure that improves RCC, and (ii) a complementary watermark synchronization layer. As our experiments demonstrate, our approach enables reliable and robust watermark detection with theoretically grounded p-values. Code and models are available at https://github.com/facebookresearch/wmar.

Method: The authors analyze neural networks compatible with polyhedral complexes (specifically the braid fan) and provide combinatorial proofs to establish lower bounds. They also examine maxout networks as a natural generalization.

Result: Proved a non-constant lower bound of Ω(log log d) hidden layers for representing the maximum of d numbers under their assumptions. Showed that 3 hidden layers are necessary for computing the maximum of 5 numbers. Demonstrated that a rank-3 maxout layer followed by a rank-2 maxout layer can represent the maximum of 7 numbers.

Conclusion: The work provides important lower bounds on network depth requirements for exact function representation and shows that existing upper bounds for maxout networks are not tight, advancing our understanding of neural network expressivity.

Abstract: We contribute towards resolving the open question of how many hidden layers are required in ReLU networks for exactly representing all continuous and piecewise linear functions on $\mathbb{R}^d$. While the question has been resolved in special cases, the best known lower bound in general is still 2. We focus on neural networks that are compatible with certain polyhedral complexes, more precisely with the braid fan. For such neural networks, we prove a non-constant lower bound of $\Omega(\log\log d)$ hidden layers required to exactly represent the maximum of $d$ numbers. Additionally, under our assumption, we provide a combinatorial proof that 3 hidden layers are necessary to compute the maximum of 5 numbers; this had only been verified with an excessive computation so far. Finally, we show that a natural generalization of the best known upper bound to maxout networks is not tight, by demonstrating that a rank-3 maxout layer followed by a rank-2 maxout layer is sufficient to represent the maximum of 7 numbers.

Method: FANTOM uses a Bayesian Expectation Maximization algorithm to maximize the evidence lower bound of data log-likelihood, simultaneously inferring the number of regimes, their boundaries, and each regime’s Directed Acyclic Graph.

Result: Theoretical proofs show identifiability under mild assumptions, and extensive experiments demonstrate FANTOM outperforms existing methods on both synthetic and real data.

Conclusion: FANTOM provides an effective solution for causal discovery in non-stationary time series with multiple regimes and complex noise distributions, addressing key limitations of existing approaches.

Abstract: Understanding causal relationships in multivariate time series is crucial in many scenarios, such as those dealing with financial or neurological data. Many such time series exhibit multiple regimes, i.e., consecutive temporal segments with a priori unknown boundaries, with each regime having its own causal structure. Inferring causal dependencies and regime shifts is critical for analyzing the underlying processes. However, causal structure learning in this setting is challenging due to (1) non-stationarity, i.e., each regime can have its own causal graph and mixing function, and (2) complex noise distributions, which may be nonGaussian or heteroscedastic. Existing causal discovery approaches cannot address these challenges, since generally assume stationarity or Gaussian noise with constant variance. Hence, we introduce FANTOM, a unified framework for causal discovery that handles non-stationary processes along with non-Gaussian and heteroscedastic noises. FANTOM simultaneously infers the number of regimes and their corresponding indices and learns each regime’s Directed Acyclic Graph. It uses a Bayesian Expectation Maximization algorithm that maximizes the evidence lower bound of the data log-likelihood. On the theoretical side, we prove, under mild assumptions, that temporal heteroscedastic causal models, introduced in FANTOM’s formulation, are identifiable in both stationary and non-stationary settings. In addition, extensive experiments on synthetic and real data show that FANTOM outperforms existing methods.

Method: Developed Ring-mini-linear-2.0 (16B parameters) and Ring-flash-linear-2.0 (104B parameters) using a hybrid architecture integrating linear attention and softmax attention, with systematic exploration of attention mechanism ratios and leveraging a self-developed FP8 operator library (linghe) for efficiency.

Result: Achieved 90% reduction in inference cost compared to 32B dense models and over 50% cost reduction compared to original Ring series. Improved training efficiency by 50% and maintained SOTA performance across multiple complex reasoning benchmarks.

Conclusion: The hybrid attention architecture successfully balances efficiency and performance, enabling stable and efficient optimization during reinforcement learning while significantly reducing computational costs for long-context inference.

Abstract: In this technical report, we present the Ring-linear model series, specifically including Ring-mini-linear-2.0 and Ring-flash-linear-2.0. Ring-mini-linear-2.0 comprises 16B parameters and 957M activations, while Ring-flash-linear-2.0 contains 104B parameters and 6.1B activations. Both models adopt a hybrid architecture that effectively integrates linear attention and softmax attention, significantly reducing I/O and computational overhead in long-context inference scenarios. Compared to a 32 billion parameter dense model, this series reduces inference cost to 1/10, and compared to the original Ring series, the cost is also reduced by over 50%. Furthermore, through systematic exploration of the ratio between different attention mechanisms in the hybrid architecture, we have identified the currently optimal model structure. Additionally, by leveraging our self-developed high-performance FP8 operator library-linghe, overall training efficiency has been improved by 50%. Benefiting from the high alignment between the training and inference engine operators, the models can undergo long-term, stable, and highly efficient optimization during the reinforcement learning phase, consistently maintaining SOTA performance across multiple challenging complex reasoning benchmarks.

Method: Proposes a continuous diffusion model incorporating categorical distribution geometry, establishes connection between discrete diffusion and continuous flow on statistical manifolds, introduces simulation-free training based on radial symmetry, and addresses high dimensionality.

Result: Outperforms existing discrete diffusion models and approaches the performance of autoregressive models on language modeling benchmarks and other modalities.

Conclusion: The proposed method successfully bridges the gap between discrete and continuous approaches, demonstrating superior performance while maintaining the benefits of diffusion modeling for discrete data.

Abstract: Diffusion models have emerged as a promising alternative to autoregressive models in modeling discrete categorical data. However, diffusion models that directly work on discrete data space fail to fully exploit the power of iterative refinement, as the signals are lost during transitions between discrete states. Existing continuous diffusion models for discrete data underperform compared to discrete methods, and the lack of a clear connection between the two approaches hinders the development of effective diffusion models for discrete data. In this work, we propose a continuous diffusion model for language modeling that incorporates the geometry of the underlying categorical distribution. We establish a connection between the discrete diffusion and continuous flow on the statistical manifold, and building on this analogy, introduce a simple diffusion process that generalizes existing discrete diffusion models. We further propose a simulation-free training framework based on radial symmetry, along with a simple technique to address the high dimensionality of the manifold. Comprehensive experiments on language modeling benchmarks and other modalities show that our method outperforms existing discrete diffusion models and approaches the performance of autoregressive models. The code is available at https://github.com/harryjo97/RDLM.

Method: RSL framework uses feature resonance phenomenon - measuring feature vector movement during training steps, combined with synthetic OOD nodes to train an OOD classifier.

Result: Extensive experiments on 13 real-world graph datasets show RSL achieves state-of-the-art performance in OOD detection.

Conclusion: Feature resonance provides an effective mechanism for OOD detection in graphs without requiring multi-category labels, with theoretical error bounds supporting its separability.

Abstract: Detecting out-of-distribution (OOD) nodes in the graph-based machine-learning field is challenging, particularly when in-distribution (ID) node multi-category labels are unavailable. Thus, we focus on feature space rather than label space and find that, ideally, during the optimization of known ID samples, unknown ID samples undergo more significant representation changes than OOD samples, even if the model is trained to fit random targets, which we called the Feature Resonance phenomenon. The rationale behind it is that even without gold labels, the local manifold may still exhibit smooth resonance. Based on this, we further develop a novel graph OOD framework, dubbed Resonance-based Separation and Learning (RSL), which comprises two core modules: (i) a more practical micro-level proxy of feature resonance that measures the movement of feature vectors in one training step. (ii) integrate with synthetic OOD nodes strategy to train an effective OOD classifier. Theoretically, we derive an error bound showing the superior separability of OOD nodes during the resonance period. Extensive experiments on a total of thirteen real-world graph datasets empirically demonstrate that RSL achieves state-of-the-art performance.

Method: Trained 4-layer encoder-decoder Transformer on modular exponentiation, using principled sampling, PCA-based embedding analysis, activation patching, and reciprocal operand training strategies.

Result: Reciprocal operand training led to strong performance gains with sudden generalization across related moduli (grokking dynamics), and identified attention-only subgraph in final layer sufficient for full performance.

Conclusion: Transformers learn modular arithmetic through specialized computational circuits, enabling more interpretable and efficient neural approaches to modular exponentiation.

Abstract: Modular exponentiation is crucial to number theory and cryptography, yet remains largely unexplored from a mechanistic interpretability standpoint. We train a 4-layer encoder-decoder Transformer model to perform this operation and investigate the emergence of numerical reasoning during training. Utilizing principled sampling strategies, PCA-based embedding analysis, and activation patching, we examine how number-theoretic properties are encoded within the model. We find that reciprocal operand training leads to strong performance gains, with sudden generalization across related moduli. These synchronized accuracy surges reflect grokking-like dynamics, suggesting the model internalizes shared arithmetic structure. We also find a subgraph consisting entirely of attention heads in the final layer sufficient to achieve full performance on the task of regular exponentiation. These results suggest that transformer models learn modular arithmetic through specialized computational circuits, paving the way for more interpretable and efficient neural approaches to modular exponentiation.

Method: Gatekeeper loss function fine-tunes smaller models to confidently handle tasks they can perform correctly while deferring complex tasks to larger models, with a mechanism to manage performance-deferral trade-offs.

Result: Experiments across encoder-only, decoder-only, and encoder-decoder architectures show substantial improvements in deferral performance for image classification, language modeling, and vision-language tasks.

Conclusion: Gatekeeper is broadly applicable across tasks and domains without architectural changes and effectively improves cascade model performance.

Abstract: Large-scale machine learning models deliver strong performance across a wide range of tasks but come with significant computational and resource constraints. To mitigate these challenges, local smaller models are often deployed alongside larger models, relying on routing and deferral mechanisms to offload complex tasks. However, existing approaches inadequately balance the capabilities of these models, often resulting in unnecessary deferrals or sub-optimal resource usage. In this work we introduce a novel loss function called Gatekeeper for calibrating smaller models in cascade setups. Our approach fine-tunes the smaller model to confidently handle tasks it can perform correctly while deferring complex tasks to the larger model. Moreover, it incorporates a mechanism for managing the trade-off between model performance and deferral accuracy, and is broadly applicable across various tasks and domains without any architectural changes. We evaluate our method on encoder-only, decoder-only, and encoder-decoder architectures. Experiments across image classification, language modeling, and vision-language tasks show that our approach substantially improves deferral performance.

Method: Proposed DA-GNN captures causal relationships in the data generating process using variational inference to handle dependency-aware noise that propagates through features, graph structure, and node labels.

Result: Extensive experiments show DA-GNN consistently outperforms existing baselines across various noise scenarios, including both the proposed dependency-aware noise and conventional noise models.

Conclusion: DA-GNN effectively handles realistic noise dependencies in graphs and new benchmark datasets enable more practical research on robust GNNs.

Abstract: In real-world applications, node features in graphs often contain noise from various sources, leading to significant performance degradation in GNNs. Although several methods have been developed to enhance robustness, they rely on the unrealistic assumption that noise in node features is independent of the graph structure and node labels, thereby limiting their applicability. To this end, we introduce a more realistic noise scenario, dependency-aware noise on graphs (DANG), where noise in node features create a chain of noise dependencies that propagates to the graph structure and node labels. We propose a novel robust GNN, DA-GNN, which captures the causal relationships among variables in the data generating process (DGP) of DANG using variational inference. In addition, we present new benchmark datasets that simulate DANG in real-world applications, enabling more practical research on robust GNNs. Extensive experiments demonstrate that DA-GNN consistently outperforms existing baselines across various noise scenarios, including both DANG and conventional noise models commonly considered in this field. Our code is available at https://github.com/yeonjun-in/torch-DA-GNN.

Method: Developed axiomatic framework with formal characterization of proper decision trees as K-permutations, then constructed generic dynamic programming recursion for exact solutions with constraints like tree depth and leaf size.

Result: Showed proper decision trees subsume various data structures (BSP trees, K-D trees, ML models) and proved they can be uniquely characterized, while demonstrating memoization impracticality due to space complexity.

Conclusion: Framework successfully analyzes both proper and non-proper decision trees, with proper trees enabling exact algorithmic solutions while revealing fundamental limitations in memoization approaches.

Abstract: We present an axiomatic framework for analyzing the algorithmic properties of decision trees. This framework supports the classification of decision tree problems through structural and ancestral constraints within a rigorous mathematical foundation. The central focus of this paper is a special class of decision tree problems-which we term proper decision trees-due to their versatility and effectiveness. In terms of versatility, this class subsumes several well-known data structures, including binary space partitioning trees, K-D trees, and machine learning decision tree models. Regarding effectiveness, we prove that only proper decision trees can be uniquely characterized as K-permutations, whereas typical non-proper decision trees correspond to binary-labeled decision trees with substantially greater complexity. Using this formal characterization, we develop a generic algorithmic approach for solving optimal decision tree problems over arbitrary splitting rules and objective functions for proper decision trees. We constructively derive a generic dynamic programming recursion for solving these problems exactly. However, we show that memoization is generally impractical in terms of space complexity, as both datasets and subtrees must be stored. This result contradicts claims in the literature that suggest a trade-off between memoizing datasets and subtrees. Our framework further accommodates constraints such as tree depth and leaf size, and can be accelerated using techniques such as thinning. Finally, we extend our analysis to several non-proper decision trees, including the commonly studied decision tree over binary feature data, the binary search tree, and the tree structure arising in the matrix chain multiplication problem. We demonstrate how these problems can be solved by appropriately modifying or discarding certain axioms.

Method: Uses a mixed-input Variational Autoencoder (VAE) to integrate categorical embeddings and numerical features into a unified latent manifold, enabling perturbations that preserve statistical consistency. Introduces In-Distribution Success Rate (IDSR) for joint evaluation.

Result: Achieves substantially lower outlier rates and more consistent performance across six datasets and three model architectures compared to traditional input-space attacks and other VAE-based methods. Shows superior practical utility and stability when reconstruction quality and sufficient training data are available.

Conclusion: The framework demonstrates the importance of maintaining on-manifold perturbations for generating realistic and robust adversarial examples in tabular domains, with effectiveness strongly dependent on reconstruction quality and training data availability.

Abstract: Adversarial attacks on tabular data present unique challenges due to the heterogeneous nature of mixed categorical and numerical features. Unlike images where pixel perturbations maintain visual similarity, tabular data lacks intuitive similarity metrics, making it difficult to define imperceptible modifications. Additionally, traditional gradient-based methods prioritise $\ell_p$-norm constraints, often producing adversarial examples that deviate from the original data distributions. To address this, we propose a latent-space perturbation framework using a mixed-input Variational Autoencoder (VAE) to generate statistically consistent adversarial examples. The proposed VAE integrates categorical embeddings and numerical features into a unified latent manifold, enabling perturbations that preserve statistical consistency. We introduce In-Distribution Success Rate (IDSR) to jointly evaluate attack effectiveness and distributional alignment. Evaluation across six publicly available datasets and three model architectures demonstrates that our method achieves substantially lower outlier rates and more consistent performance compared to traditional input-space attacks and other VAE-based methods adapted from image domain approaches, achieving substantially lower outlier rates and higher IDSR across six datasets and three model architectures. Our comprehensive analyses of hyperparameter sensitivity, sparsity control, and generative architecture demonstrate that the effectiveness of VAE-based attacks depends strongly on reconstruction quality and the availability of sufficient training data. When these conditions are met, the proposed framework achieves superior practical utility and stability compared with input-space methods. This work underscores the importance of maintaining on-manifold perturbations for generating realistic and robust adversarial examples in tabular domains.

Method: Uses teacher-student model architecture enhanced by knowledge distillation on bright-field microscopy images. Trained on 80,000 preprocessed images of lymphocyte subsets (T4, T8, B cells). Student model is highly compressed for FPGA deployment.

Result: Teacher model achieved 98% accuracy for T4 vs B cells and 93% zero-shot accuracy for T8 vs B cells. Student model with only 0.02% of teacher’s parameters achieved ultra-low inference latency of 14.5μs and total detection-to-sorting time of 24.7μs, providing 12x and 40x improvements over previous state-of-the-art.

Conclusion: The framework provides a scalable, cost-effective solution for lymphocyte classification and sets a new SOTA for real-time cell sorting using in situ deep learning on off-the-shelf computing hardware.

Abstract: Precise cell classification is essential in biomedical diagnostics and therapeutic monitoring, particularly for identifying diverse cell types involved in various diseases. Traditional cell classification methods such as flow cytometry depend on molecular labeling which is often costly, time-intensive, and can alter cell integrity. To overcome these limitations, we present a label-free machine learning framework for cell classification, designed for real-time sorting applications using bright-field microscopy images. This approach leverages a teacher-student model architecture enhanced by knowledge distillation, achieving high efficiency and scalability across different cell types. Demonstrated through a use case of classifying lymphocyte subsets, our framework accurately classifies T4, T8, and B cell types with a dataset of 80,000 preprocessed images, accessible via an open-source Python package for easy adaptation. Our teacher model attained 98% accuracy in differentiating T4 cells from B cells and 93% accuracy in zero-shot classification between T8 and B cells. Remarkably, our student model operates with only 0.02% of the teacher model’s parameters, enabling field-programmable gate array (FPGA) deployment. Our FPGA-accelerated student model achieves an ultra-low inference latency of just 14.5~$\mu$s and a complete cell detection-to-sorting trigger time of 24.7~$\mu$s, delivering 12x and 40x improvements over the previous state-of-the-art real-time cell analysis algorithm in inference and total latency, respectively, while preserving accuracy comparable to the teacher model. This framework provides a scalable, cost-effective solution for lymphocyte classification, as well as a new SOTA real-time cell sorting implementation for rapid identification of subsets using in situ deep learning on off-the-shelf computing hardware.

Method: Analyzed performance of Minibatch SGD, Local SGD, and Local SGD with momentum under standard assumptions and smooth non-convex client objectives for Markovian data streams.

Result: FL can support collaborative learning with Markovian data streams - sample complexity is proportional to inverse number of clients with communication complexity comparable to i.i.d. scenario, though sample complexity remains higher than i.i.d. sampling.

Conclusion: Federated learning is feasible with Markovian data streams under standard assumptions, achieving communication efficiency similar to i.i.d. cases but with higher sample complexity requirements.

Abstract: Federated learning (FL) is now recognized as a key framework for communication-efficient collaborative learning. Most theoretical and empirical studies, however, rely on the assumption that clients have access to pre-collected data sets, with limited investigation into scenarios where clients continuously collect data. In many real-world applications, particularly when data is generated by physical or biological processes, client data streams are often modeled by non-stationary Markov processes. Unlike standard i.i.d. sampling, the performance of FL with Markovian data streams remains poorly understood due to the statistical dependencies between client samples over time. In this paper, we investigate whether FL can still support collaborative learning with Markovian data streams. Specifically, we analyze the performance of Minibatch SGD, Local SGD, and a variant of Local SGD with momentum. We answer affirmatively under standard assumptions and smooth non-convex client objectives: the sample complexity is proportional to the inverse of the number of clients with a communication complexity comparable to the i.i.d. scenario. However, the sample complexity for Markovian data streams remains higher than for i.i.d. sampling.

Method: Uses Bayesian Optimization with Gaussian process regression as surrogate models to optimize process parameters, considering uncertainties in sorting accuracy calculations and minimizing required experiments.

Result: The method was evaluated with three example process parameters, successfully handling multiple optimization targets for both material output streams while accounting for uncertainties.

Conclusion: The proposed Bayesian Optimization approach effectively optimizes sensor-based sorting system parameters with minimal experiments, handling uncertainties and multiple optimization objectives simultaneously.

Abstract: Sensor-based sorting systems enable the physical separation of a material stream into two fractions. The sorting decision is based on the image data evaluation of the sensors used and is carried out using actuators. Various process parameters must be set depending on the properties of the material stream, the dimensioning of the system, and the required sorting accuracy. However, continuous verification and re-adjustment are necessary due to changing requirements and material stream compositions. In this paper, we introduce an approach for optimizing, recurrently monitoring and adjusting the process parameters of a sensor-based sorting system. Based on Bayesian Optimization, Gaussian process regression models are used as surrogate models to achieve specific requirements for system behavior with the uncertainties contained therein. This method minimizes the number of necessary experiments while simultaneously considering two possible optimization targets based on the requirements for both material output streams. In addition, uncertainties are considered during determining sorting accuracies in the model calculation. We evaluated the method with three example process parameters.

Method: Introduces reciprocal space formulation for RBMs, analyzes local curvature via singular values, studies symmetry breaking during training, and derives free energy in mean-field approximation. Experiments with RBMs of varying hidden layer sizes on MNIST dataset.

Result: At initialization, RBMs operate at saddle points with rotational symmetry following Marcenko-Pastur law. Training breaks rotational symmetry through hierarchical learning. In infinite size limit, reciprocal variables become Gaussian distributed, and some diffusion modes don’t converge to Boltzmann distribution.

Conclusion: The findings bridge gaps between disparate generative frameworks and illuminate learning processes in generative models, connecting RBMs to diffusion processes and coupled Bosons through symmetry breaking mechanisms.

Abstract: Deep generative models have become ubiquitous due to their ability to learn and sample from complex distributions. Despite the proliferation of various frameworks, the relationships among these models remain largely unexplored, a gap that hinders the development of a unified theory of AI learning. We address two central challenges: clarifying the connections between different deep generative models and deepening our understanding of their learning mechanisms. We focus on Restricted Boltzmann Machines (RBMs), known for their universal approximation capabilities for discrete distributions. By introducing a reciprocal space formulation, we reveal a connection between RBMs, diffusion processes, and coupled Bosons. We show that at initialization, the RBM operates at a saddle point, where the local curvature is determined by the singular values, whose distribution follows the Marcenko-Pastur law and exhibits rotational symmetry. During training, this rotational symmetry is broken due to hierarchical learning, where different degrees of freedom progressively capture features at multiple levels of abstraction. This leads to a symmetry breaking in the energy landscape, reminiscent of Landau theory. This symmetry breaking in the energy landscape is characterized by the singular values and the weight matrix eigenvector matrix. We derive the corresponding free energy in a mean-field approximation. We show that in the limit of infinite size RBM, the reciprocal variables are Gaussian distributed. Our findings indicate that in this regime, there will be some modes for which the diffusion process will not converge to the Boltzmann distribution. To illustrate our results, we trained replicas of RBMs with different hidden layer sizes using the MNIST dataset. Our findings bridge the gap between disparate generative frameworks and also shed light on the processes underpinning learning in generative models.

Method: Proposes Sign-In, which employs dynamic reparameterization that provably induces sign flips. This approach is orthogonal to dense-to-sparse training methods and complements their sign-flipping capabilities.

Result: Experiments and theory suggest performance improvements for PaI (training sparse networks from scratch), though the main challenge of closing the gap with dense-to-sparse training remains open.

Conclusion: Sign-In provides a promising orthogonal method to improve PaI performance through sign flips, but further work is needed to fully close the performance gap between PaI and dense-to-sparse training approaches.

Abstract: The performance gap between training sparse neural networks from scratch (PaI) and dense-to-sparse training presents a major roadblock for efficient deep learning. According to the Lottery Ticket Hypothesis, PaI hinges on finding a problem specific parameter initialization. As we show, to this end, determining correct parameter signs is sufficient. Yet, they remain elusive to PaI. To address this issue, we propose Sign-In, which employs a dynamic reparameterization that provably induces sign flips. Such sign flips are complementary to the ones that dense-to-sparse training can accomplish, rendering Sign-In as an orthogonal method. While our experiments and theory suggest performance improvements of PaI, they also carve out the main open challenge to close the gap between PaI and dense-to-sparse training.

Method: Proposes DMSC framework with three core components: EMPD for dynamic hierarchical patch decomposition with exponential granularities, TIB for joint modeling of intra-patch, inter-patch, and cross-variable dependencies, and ASR-MoE for adaptive fusion using specialized experts with temporal-aware weighting. These are integrated in a multi-layer progressive cascade architecture.

Result: Comprehensive experiments on thirteen real-world benchmarks demonstrate that DMSC consistently maintains state-of-the-art performance and superior computational efficiency for time series forecasting tasks.

Conclusion: DMSC effectively addresses key limitations in time series forecasting through its dynamic multi-scale coordination approach, achieving superior performance and efficiency across diverse benchmarks.

Abstract: Time Series Forecasting (TSF) faces persistent challenges in modeling intricate temporal dependencies across different scales. Despite recent advances leveraging different decomposition operations and novel architectures based on CNN, MLP or Transformer, existing methods still struggle with static decomposition strategies, fragmented dependency modeling, and inflexible fusion mechanisms, limiting their ability to model intricate temporal dependencies. To explicitly solve the mentioned three problems respectively, we propose a novel Dynamic Multi-Scale Coordination Framework (DMSC) with Multi-Scale Patch Decomposition block (EMPD), Triad Interaction Block (TIB) and Adaptive Scale Routing MoE block (ASR-MoE). Specifically, EMPD is designed as a built-in component to dynamically segment sequences into hierarchical patches with exponentially scaled granularities, eliminating predefined scale constraints through input-adaptive patch adjustment. TIB then jointly models intra-patch, inter-patch, and cross-variable dependencies within each layer’s decomposed representations. EMPD and TIB are jointly integrated into layers forming a multi-layer progressive cascade architecture, where coarse-grained representations from earlier layers adaptively guide fine-grained feature extraction in subsequent layers via gated pathways. And ASR-MoE dynamically fuses multi-scale predictions by leveraging specialized global and local experts with temporal-aware weighting. Comprehensive experiments on thirteen real-world benchmarks demonstrate that DMSC consistently maintains state-of-the-art (SOTA) performance and superior computational efficiency for TSF tasks. Code is available at https://github.com/1327679995/DMSC.

Method: CTSketch decomposes symbolic programs into sub-programs and summarizes each with sketched tensors, approximating output distributions through simple tensor operations over input distributions and sketches.

Result: CTSketch achieves high accuracy on neurosymbolic benchmarks, including scalability tests with 1000 inputs, where neural predictors learn effectively despite only having supervision on the final output.

Conclusion: CTSketch significantly advances neurosymbolic learning scalability, making previously unattainable large-scale tasks feasible through program decomposition and tensor sketching techniques.

Abstract: Many computational tasks benefit from being formulated as the composition of neural networks followed by a discrete symbolic program. The goal of neurosymbolic learning is to train the neural networks using end-to-end input-output labels of the composite. We introduce CTSketch, a novel, scalable neurosymbolic learning algorithm. CTSketch uses two techniques to improve the scalability of neurosymbolic inference: decompose the symbolic program into sub-programs and summarize each sub-program with a sketched tensor. This strategy allows us to approximate the output distribution of the program with simple tensor operations over the input distributions and the sketches. We provide theoretical insight into the maximum approximation error. Furthermore, we evaluate CTSketch on benchmarks from the neurosymbolic learning literature, including some designed for evaluating scalability. Our results show that CTSketch pushes neurosymbolic learning to new scales that were previously unattainable, with neural predictors obtaining high accuracy on tasks with one thousand inputs, despite supervision only on the final output.

Method: Each agent solves a constrained Markov decision process (CMDP) with coordination through shared constraint vectors updated infrequently, using safe reinforcement learning to meet both local and global goals.

Result: The approach shows improved performance over centralized and independent baselines, especially in large-scale settings, with theoretical guarantees established under mild assumptions.

Conclusion: The proposed decentralized framework with implicit coordination through shared constraints enables effective resource management in edge computing with minimal communication requirements.

Abstract: In edge computing systems, autonomous agents must make fast local decisions while competing for shared resources. Existing MARL methods often resume to centralized critics or frequent communication, which fail under limited observability and communication constraints. We propose a decentralized framework in which each agent solves a constrained Markov decision process (CMDP), coordinating implicitly through a shared constraint vector. For the specific case of offloading, e.g., constraints prevent overloading shared server resources. Coordination constraints are updated infrequently and act as a lightweight coordination mechanism. They enable agents to align with global resource usage objectives but require little direct communication. Using safe reinforcement learning, agents learn policies that meet both local and global goals. We establish theoretical guarantees under mild assumptions and validate our approach experimentally, showing improved performance over centralized and independent baselines, especially in large-scale settings.

Method: Uses Principal Component Analysis (PCA) for feature reduction with reconstruction error for outlier detection, employing Incremental PCA to adapt to evolving data distributions and pre-scaling to normalize feature magnitudes while preserving relative variance.

Result: Successfully detected space plasma events including distinct space environments, dayside/nightside transient phenomena, and transition layers using NASA’s MMS mission data, and identified a dayside transient using THEMIS data with onboard-available measurements.

Conclusion: The adaptive outlier detection algorithm based on Incremental PCA is effective for real-time event detection in space missions, capable of handling evolving data distributions without predefined models.

Abstract: Analyzing multi-featured time series data is critical for space missions making efficient event detection, potentially onboard, essential for automatic analysis. However, limited onboard computational resources and data downlink constraints necessitate robust methods for identifying regions of interest in real time. This work presents an adaptive outlier detection algorithm based on the reconstruction error of Principal Component Analysis (PCA) for feature reduction, designed explicitly for space mission applications. The algorithm adapts dynamically to evolving data distributions by using Incremental PCA, enabling deployment without a predefined model for all possible conditions. A pre-scaling process normalizes each feature’s magnitude while preserving relative variance within feature types. We demonstrate the algorithm’s effectiveness in detecting space plasma events, such as distinct space environments, dayside and nightside transients phenomena, and transition layers through NASA’s MMS mission observations. Additionally, we apply the method to NASA’s THEMIS data, successfully identifying a dayside transient using onboard-available measurements.

Method: Integrates a model to predict underlying state with online conformal prediction to model uncertainties in non-stationary time series.

Result: Proves CPTC’s validity and improved adaptivity under minimum assumptions, and demonstrates effectiveness on 6 synthetic and real-world datasets with improved validity and adaptivity over state-of-the-art baselines.

Conclusion: CPTC provides a novel solution for uncertainty quantification in time series with change points, offering improved performance compared to existing methods.

Abstract: Conformal prediction has been explored as a general and efficient way to provide uncertainty quantification for time series. However, current methods struggle to handle time series data with change points - sudden shifts in the underlying data-generating process. In this paper, we propose a novel Conformal Prediction for Time-series with Change points (CPTC) algorithm, addressing this gap by integrating a model to predict the underlying state with online conformal prediction to model uncertainties in non-stationary time series. We prove CPTC’s validity and improved adaptivity in the time series setting under minimum assumptions, and demonstrate CPTC’s practical effectiveness on 6 synthetic and real-world datasets, showing improved validity and adaptivity compared to state-of-the-art baselines.

Method: Modifies DeepONet’s branch network to process input functions as unordered sets of location-value pairs using Deep Sets principles, ensuring permutation invariance while keeping the same parameter count.

Result: Achieves parity with DeepONet on fixed layouts while maintaining accuracy under variable sensor configurations, sensor drop-off, and unstructured point cloud inputs. Successfully applied to heat conduction with point sources, advection-diffusion modeling, and optimal transport problems.

Conclusion: SetONet significantly broadens operator learning applicability to problems with variable, incomplete, or unstructured input data through a lightweight design that handles point sets without rasterization or multi-stage pipelines.

Abstract: Neural operators, particularly the Deep Operator Network (DeepONet), have shown promise in learning mappings between function spaces for solving differential equations. However, standard DeepONet requires input functions to be sampled at fixed locations, limiting its applicability when sensor configurations vary or inputs exist on irregular grids. We introduce the Set Operator Network (SetONet), which modifies DeepONet’s branch network to process input functions as unordered sets of location-value pairs. By incorporating Deep Sets principles, SetONet ensures permutation invariance while maintaining the same parameter count as the baseline. On classical operator-learning benchmarks, SetONet achieves parity with DeepONet on fixed layouts while sustaining accuracy under variable sensor configurations or sensor drop-off - conditions for which standard DeepONet is not applicable. More significantly, SetONet natively handles problems where inputs are naturally represented as unstructured point clouds (such as point sources or density samples) rather than values on fixed grids, a capability standard DeepONet lacks. On heat conduction with point sources, advection-diffusion modeling chemical plumes, and optimal transport between density samples, SetONet learns operators end-to-end without rasterization or multi-stage pipelines. These problems feature inputs that are naturally discrete point sets (point sources or density samples) rather than functions on fixed grids. SetONet is a DeepONet-class architecture that addresses such problems with a lightweight design, significantly broadening the applicability of operator learning to problems with variable, incomplete, or unstructured input data.

Method: Extends sparse autoencoders (SAEs) to lifted spaces and infinite-dimensional function spaces, comparing inference and training dynamics of SAEs, lifted-SAE, and SAE neural operators.

Result: Lifting and operator modules introduce beneficial inductive biases that enable faster recovery, improved recovery of smooth concepts, and robust inference across varying resolutions - a property unique to neural operators.

Conclusion: The framework successfully enables mechanistic interpretability of large neural operators and demonstrates that lifting and operator modules provide significant advantages in representation recovery and robustness across resolutions.

Abstract: We frame the problem of unifying representations in neural models as one of sparse model recovery and introduce a framework that extends sparse autoencoders (SAEs) to lifted spaces and infinite-dimensional function spaces, enabling mechanistic interpretability of large neural operators (NO). While the Platonic Representation Hypothesis suggests that neural networks converge to similar representations across architectures, the representational properties of neural operators remain underexplored despite their growing importance in scientific computing. We compare the inference and training dynamics of SAEs, lifted-SAE, and SAE neural operators. We highlight how lifting and operator modules introduce beneficial inductive biases, enabling faster recovery, improved recovery of smooth concepts, and robust inference across varying resolutions, a property unique to neural operators.

Method: Extend GRPO and PPO to multi-turn variants with turn-level rewards, including verifiable and LLM-as-judge rewards for fine-grained credit assignment in reasoning-augmented search agents.

Result: Experiments show significant outperformance over baseline methods with trajectory-level rewards, achieving greater stability, faster convergence, higher accuracy, highest answer correctness, and 100% format correctness.

Conclusion: Well-designed turn-level rewards enable RL algorithms to achieve superior performance in multi-turn reasoning tasks, providing a systematic approach for reward design in multi-turn agent applications.

Abstract: This paper investigates Reinforcement Learning (RL) approaches to enhance the reasoning capabilities of Large Language Model (LLM) agents in long-horizon, multi-turn scenarios. Although RL algorithms such as Group Relative Policy Optimization (GRPO) and Proximal Policy Optimization (PPO) have been widely applied to train multi-turn LLM agents, they typically rely only on sparse outcome rewards and lack dense intermediate signals across multiple decision steps, limiting their performance on complex reasoning tasks. To bridge this gap, we present the first systematic study of \textit{turn-level reward design} for multi-turn RL algorithms and agent applications. By integrating turn-level rewards, we extend GRPO and PPO to their respective multi-turn variants, enabling fine-grained credit assignment. We conduct case studies on multi-turn reasoning-augmented search agents, where we carefully design two types of turn-level rewards: verifiable and LLM-as-judge. Our experiments on multi-turn search tasks demonstrate that incorporating well-designed turn-level rewards enables RL algorithms to significantly outperform baseline methods with trajectory-level rewards. Both training and validation reward curves illustrate that our method achieves \textit{greater stability}, \textit{faster convergence}, and \textit{higher accuracy}. Numerical results across diverse question-answering datasets further show that our approach consistently delivers highest answer correctness and 100% format correctness.

Method: Uses two state representation strategies (Feature-Centric and Object-Centric) with a Transformer-based Advantage Actor-Critic (A2C) algorithm for training.

Result: Evaluated models across multiple rule-based and trial-list-based experimental setups, analyzing transfer effects and representation impact on learning efficiency.

Conclusion: The study provides insights into how different state representations affect learning efficiency and transfer capabilities in hidden rule inference tasks.

Abstract: We investigate reinforcement learning in the Game Of Hidden Rules (GOHR) environment, a complex puzzle in which an agent must infer and execute hidden rules to clear a 6$\times$6 board by placing game pieces into buckets. We explore two state representation strategies, namely Feature-Centric (FC) and Object-Centric (OC), and employ a Transformer-based Advantage Actor-Critic (A2C) algorithm for training. The agent has access only to partial observations and must simultaneously infer the governing rule and learn the optimal policy through experience. We evaluate our models across multiple rule-based and trial-list-based experimental setups, analyzing transfer effects and the impact of representation on learning efficiency.

Method: Three techniques: 1) Woodbury formula to reduce computational complexity, 2) Subsampled Projected-Increment Natural Gradient Descent algorithm adaptation, 3) Randomized algorithms for large batch sizes.

Result: Methods outperform previous approaches, achieving same L² error as original ENGD up to 75× faster. Randomization helps early training for low-dimensional problems.

Conclusion: The proposed suite of techniques significantly improves ENGD efficiency for PINNs while maintaining accuracy, with identified limitations for certain scenarios.

Abstract: Natural gradient methods significantly accelerate the training of Physics-Informed Neural Networks (PINNs), but are often prohibitively costly. We introduce a suite of techniques to improve the accuracy and efficiency of energy natural gradient descent (ENGD) for PINNs. First, we leverage the Woodbury formula to dramatically reduce the computational complexity of ENGD. Second, we adapt the Subsampled Projected-Increment Natural Gradient Descent algorithm from the variational Monte Carlo literature to accelerate the convergence. Third, we explore the use of randomized algorithms to further reduce the computational cost in the case of large batch sizes. We find that randomization accelerates progress in the early stages of training for low-dimensional problems, and we identify key barriers to attaining acceleration in other scenarios. Our numerical experiments demonstrate that our methods outperform previous approaches, achieving the same $L^2$ error as the original ENGD up to $75\times$ faster.

Method: Identifies and reuses intrinsic computational redundancies in the diffusion process through feature-level cross-step reuse and inter-layer scheduling, without modifying model parameters. Evolves from static reuse to dynamic prediction approaches.

Result: Enables efficient inference by reducing computation while maintaining quality. Enhances caching flexibility across diverse tasks and enables integration with other acceleration techniques like sampling optimization and model distillation.

Conclusion: Diffusion Caching represents a promising paradigm that will become a key enabler for real-time and efficient generative AI, providing a unified inference framework for future multimodal and interactive applications.

Abstract: Diffusion Models have become a cornerstone of modern generative AI for their exceptional generation quality and controllability. However, their inherent \textit{multi-step iterations} and \textit{complex backbone networks} lead to prohibitive computational overhead and generation latency, forming a major bottleneck for real-time applications. Although existing acceleration techniques have made progress, they still face challenges such as limited applicability, high training costs, or quality degradation. Against this backdrop, \textbf{Diffusion Caching} offers a promising training-free, architecture-agnostic, and efficient inference paradigm. Its core mechanism identifies and reuses intrinsic computational redundancies in the diffusion process. By enabling feature-level cross-step reuse and inter-layer scheduling, it reduces computation without modifying model parameters. This paper systematically reviews the theoretical foundations and evolution of Diffusion Caching and proposes a unified framework for its classification and analysis. Through comparative analysis of representative methods, we show that Diffusion Caching evolves from \textit{static reuse} to \textit{dynamic prediction}. This trend enhances caching flexibility across diverse tasks and enables integration with other acceleration techniques such as sampling optimization and model distillation, paving the way for a unified, efficient inference framework for future multimodal and interactive applications. We argue that this paradigm will become a key enabler of real-time and efficient generative AI, injecting new vitality into both theory and practice of \textit{Efficient Generative Intelligence}.

Method: Parameterize Q-function using a velocity field and train it using flow-matching techniques from generative modeling. The velocity field is trained with TD-learning objective that bootstraps from values produced by a target velocity field computed through multiple steps of numerical integration.

Result: Floq improves performance by nearly 1.8x across challenging offline RL benchmarks and online fine-tuning tasks. It scales capacity far better than standard TD-learning architectures.

Conclusion: Iterative computation shows significant potential for value learning in RL, enabling more fine-grained control and better scaling of Q-function capacity through appropriate setting of integration steps.

Abstract: A hallmark of modern large-scale machine learning techniques is the use of training objectives that provide dense supervision to intermediate computations, such as teacher forcing the next token in language models or denoising step-by-step in diffusion models. This enables models to learn complex functions in a generalizable manner. Motivated by this observation, we investigate the benefits of iterative computation for temporal difference (TD) methods in reinforcement learning (RL). Typically they represent value functions in a monolithic fashion, without iterative compute. We introduce floq (flow-matching Q-functions), an approach that parameterizes the Q-function using a velocity field and trains it using techniques from flow-matching, typically used in generative modeling. This velocity field underneath the flow is trained using a TD-learning objective, which bootstraps from values produced by a target velocity field, computed by running multiple steps of numerical integration. Crucially, floq allows for more fine-grained control and scaling of the Q-function capacity than monolithic architectures, by appropriately setting the number of integration steps. Across a suite of challenging offline RL benchmarks and online fine-tuning tasks, floq improves performance by nearly 1.8x. floq scales capacity far better than standard TD-learning architectures, highlighting the potential of iterative computation for value learning.

Method: Formalized the KKT point embedding principle, proved it holds for neuron splitting in fully-connected networks and channel splitting in CNNs, and connected static embedding to gradient flow training dynamics with smooth losses.

Result: Proved that KKT points can be embedded between networks of different sizes, and showed that training trajectories initiated from mapped points remain mapped throughout training, preserving alignment with KKT directions dynamically.

Conclusion: The findings provide insights into network width effects, parameter redundancy, and structural connections between solutions in homogeneous networks of varying sizes, with implications for understanding optimization behavior across different network architectures.

Abstract: In this paper, we study the Karush-Kuhn-Tucker (KKT) points of the associated maximum-margin problem in homogeneous neural networks, including fully-connected and convolutional neural networks. In particular, We investigates the relationship between such KKT points across networks of different widths generated. We introduce and formalize the \textbf{KKT point embedding principle}, establishing that KKT points of a homogeneous network’s max-margin problem ($P_{\Phi}$) can be embedded into the KKT points of a larger network’s problem ($P_{\tilde{\Phi}}$) via specific linear isometric transformations. We rigorously prove this principle holds for neuron splitting in fully-connected networks and channel splitting in convolutional neural networks. Furthermore, we connect this static embedding to the dynamics of gradient flow training with smooth losses. We demonstrate that trajectories initiated from appropriately mapped points remain mapped throughout training and that the resulting $\omega$-limit sets of directions are correspondingly mapped, thereby preserving the alignment with KKT directions dynamically when directional convergence occurs. We conduct several experiments to justify that trajectories are preserved. Our findings offer insights into the effects of network width, parameter redundancy, and the structural connections between solutions found via optimization in homogeneous networks of varying sizes.

Method: Proposes using multiple alternative risk indicators including marginalizing over LLM-as-a-judge variants, structured tasks, out-of-distribution detection, and Elo rating systems for comprehensive evaluation.

Result: Shows that marginalizing over multiple LLM-as-a-judge variants reduces evaluation biases and that structured tasks provide more robust and controllable risk indicators.

Conclusion: The proposed evaluation framework provides more robust and objective assessment of uncertainty estimation methods for detecting LLM confabulations across various settings.

Abstract: Hallucinations are a common issue that undermine the reliability of large language models (LLMs). Recent studies have identified a specific subset of hallucinations, known as confabulations, which arise due to predictive uncertainty of LLMs. To detect confabulations, various methods for estimating predictive uncertainty in natural language generation (NLG) have been developed. These methods are typically evaluated by correlating uncertainty estimates with the correctness of generated text, with question-answering (QA) datasets serving as the standard benchmark. However, commonly used approximate correctness functions have substantial disagreement between each other and, consequently, in the ranking of the uncertainty estimation methods. This allows one to inflate the apparent performance of uncertainty estimation methods. We propose using several alternative risk indicators for risk correlation experiments that improve robustness of empirical assessment of UE algorithms for NLG. For QA tasks, we show that marginalizing over multiple LLM-as-a-judge variants leads to reducing the evaluation biases. Furthermore, we explore structured tasks as well as out of distribution and perturbation detection tasks which provide robust and controllable risk indicators. Finally, we propose to use an Elo rating of uncertainty estimation methods to give an objective summarization over extensive evaluation settings.

Method: Iteratively train two neural networks: one for optimal policy representation and another for value function approximation, using continuous-time dynamic programming and Hamilton-Jacobi-Bellman equation objectives.

Result: Empirical evaluations show the approach achieves accuracy and scalability in solving complex, high-dimensional stochastic control tasks.

Conclusion: The proposed method effectively solves challenging stochastic control problems with jumps, demonstrating practical applicability through empirical validation.

Abstract: In this paper, we introduce a model-based deep-learning approach to solve finite-horizon continuous-time stochastic control problems with jumps. We iteratively train two neural networks: one to represent the optimal policy and the other to approximate the value function. Leveraging a continuous-time version of the dynamic programming principle, we derive two different training objectives based on the Hamilton-Jacobi-Bellman equation, ensuring that the networks capture the underlying stochastic dynamics. Empirical evaluations on different problems illustrate the accuracy and scalability of our approach, demonstrating its effectiveness in solving complex, high-dimensional stochastic control tasks.

Method: Introduces Tropical Attention that lifts attention kernel into tropical projective space, making reasoning piecewise-linear and 1-Lipschitz. Multi-Head Tropical Attention (MHTA) universally approximates tropical circuits and realizes tropical transitive closure through composition.

Result: Empirical results show stronger out-of-distribution generalization, high robustness against noise, faster inference with fewer parameters compared to Softmax-based and recurrent attention baselines. Successfully extends neural algorithmic reasoning to NP-hard and NP-complete problems.

Conclusion: Tropical Attention paves the way for sharper and more expressive Large Reasoning Models capable of tackling complex combinatorial challenges in various domains including phylogenetics, cryptography, particle physics, and mathematical discovery.

Abstract: Can algebraic geometry enhance the sharpness, robustness, and interpretability of modern neural reasoning models by equipping them with a mathematically grounded inductive bias? To answer this, we introduce Tropical Attention, an attention mechanism grounded in tropical geometry that lifts the attention kernel into tropical projective space, where reasoning is piecewise-linear and 1-Lipschitz, thus preserving the polyhedral decision structure inherent to combinatorial reasoning. We prove that Multi-Head Tropical Attention (MHTA) stacks universally approximate tropical circuits and realize tropical transitive closure through composition, achieving polynomial resource bounds without invoking recurrent mechanisms. These guarantees explain why the induced polyhedral decision boundaries remain sharp and scale-invariant, rather than smoothed by Softmax. Empirically, we show that Tropical Attention delivers stronger out-of-distribution generalization in both length and value, with high robustness against perturbative noise, and substantially faster inference with fewer parameters compared to Softmax-based and recurrent attention baselines. For the first time, we extend neural algorithmic reasoning beyond PTIME problems to NP-hard and NP-complete problems, paving the way toward sharper and more expressive Large Reasoning Models (LRMs) capable of tackling complex combinatorial challenges in phylogenetics, cryptography, particle physics, and mathematical discovery.

Method: Replaces exponential kernel with sharpened angular (cosine) similarity, and approximates attention outputs via randomized projections and soft Locality-Sensitive Hashing (LSH).

Result: Matches accuracy of strong baselines across language modeling, masked language modeling, and text classification while reducing runtime and memory. Processes up to 12M tokens on GPU and 75M tokens on CPU.

Conclusion: RACE Attention provides a practical, theoretically grounded mechanism for extremely long context windows on current hardware, enabling processing well beyond state-of-the-art attention implementations.

Abstract: Softmax Attention has a quadratic time complexity, which becomes prohibitive to run at long contexts, even with highly optimized GPU kernels. For example, FlashAttention (an exact, GPU-optimized implementation of Softmax Attention) cannot complete a single forward-backward pass of a multi-head attention layer once the context exceeds ~4 million tokens on an NVIDIA GH200 (96 GB). We introduce RACE Attention, a kernel-inspired alternative to Softmax Attention that is linear in sequence length and embedding dimension. RACE Attention replaces the exponential kernel with a sharpened angular (cosine) similarity, and approximates attention outputs via randomized projections and soft Locality-Sensitive Hashing (LSH). Across language modeling, masked language modeling, and text classification, RACE Attention matches the accuracy of strong baselines while reducing runtime and memory. In a controlled scale test, it processes up to 12 million tokens during a single forward-backward pass on an NVIDIA GH200 GPU and 75 million tokens on an Intel Xeon Gold 5220R CPU, well beyond the practical limits of the current state-of-the-art attention implementations. RACE Attention thus offers a practical, theoretically grounded mechanism for outrageously long context windows on today’s hardware. We hope that it gets adopted in practice.

Method: Proposed two approaches: 1) estimator with provable asymptotic convergence rates when informative source domains are known, 2) data-driven transfer learning procedure to mitigate negative transfer when informative domains are unknown.

Result: The methods are supported by rigorous theoretical analysis and validated through extensive simulations and real-world applications, demonstrating effective transfer learning for distributional outputs.

Conclusion: The framework successfully extends transfer learning to Wasserstein space for probability distribution outputs, providing both theoretical guarantees and practical solutions for domain transfer scenarios.

Abstract: Transfer learning is a powerful paradigm for leveraging knowledge from source domains to enhance learning in a target domain. However, traditional transfer learning approaches often focus on scalar or multivariate data within Euclidean spaces, limiting their applicability to complex data structures such as probability distributions. To address this limitation, we introduce a novel transfer learning framework for regression models whose outputs are probability distributions residing in the Wasserstein space. When the informative subset of transferable source domains is known, we propose an estimator with provable asymptotic convergence rates, quantifying the impact of domain similarity on transfer efficiency. For cases where the informative subset is unknown, we develop a data-driven transfer learning procedure designed to mitigate negative transfer. The proposed methods are supported by rigorous theoretical analysis and are validated through extensive simulations and real-world applications. The code is available at https://github.com/h7nian/WaTL

Method: Built comprehensive modular algorithmic space with hundreds of components, then used dual-agent reinforcement learning system for structural and parametric design through cooperative training across 10k diverse instances.

Result: DesignX-generated optimizers surpass human-crafted ones by orders of magnitude on synthetic testbeds and real scenarios like Protein-docking, AutoML, and UAV path planning.

Conclusion: DesignX can discover non-trivial algorithm patterns beyond expert intuition, providing valuable design insights for the optimization community.

Abstract: Designing effective black-box optimizers is hampered by limited problem-specific knowledge and manual control that spans months for almost every detail. In this paper, we present \textit{DesignX}, the first automated algorithm design framework that generates an effective optimizer specific to a given black-box optimization problem within seconds. Rooted in the first principles, we identify two key sub-tasks: 1) algorithm structure generation and 2) hyperparameter control. To enable systematic construction, a comprehensive modular algorithmic space is first built, embracing hundreds of algorithm components collected from decades of research. We then introduce a dual-agent reinforcement learning system that collaborates on structural and parametric design through a novel cooperative training objective, enabling large-scale meta-training across 10k diverse instances. Remarkably, through days of autonomous learning, the DesignX-generated optimizers continuously surpass human-crafted optimizers by orders of magnitude, either on synthetic testbed or on realistic optimization scenarios such as Protein-docking, AutoML and UAV path planning. Further in-depth analysis reveals DesignX’s capability to discover non-trivial algorithm patterns beyond expert intuition, which, conversely, provides valuable design insights for the optimization community. We provide DesignX’s Python project at~ https://github.com/MetaEvo/DesignX.

Method: RT tokenizes cells with table/column metadata, uses masked token prediction for pretraining, and employs a novel Relational Attention mechanism over columns, rows, and primary-foreign key links.

Result: Pretrained on RelBench datasets, RT achieves 93% of fully supervised AUROC on binary classification tasks with zero-shot performance using a 22M parameter model, outperforming a 27B LLM (84%). Fine-tuning yields state-of-the-art results with high sample efficiency.

Conclusion: RT provides a practical path toward foundation models for relational data by effectively harnessing task-table context, relational attention patterns, and schema semantics for zero-shot transfer.

Abstract: Pretrained transformers readily adapt to new sequence modeling tasks via zero-shot prompting, but relational domains still lack architectures that transfer across datasets and tasks. The core challenge is the diversity of relational data, with varying heterogeneous schemas, graph structures and functional dependencies. In this paper, we present the Relational Transformer (RT) architecture, which can be pretrained on diverse relational databases and directly applied to unseen datasets and tasks without task- or dataset-specific fine-tuning, or retrieval of in-context examples. RT (i) tokenizes cells with table/column metadata, (ii) is pretrained via masked token prediction, and (iii) utilizes a novel Relational Attention mechanism over columns, rows, and primary-foreign key links. Pretrained on RelBench datasets spanning tasks such as churn and sales forecasting, RT attains strong zero-shot performance, averaging 93% of fully supervised AUROC on binary classification tasks with a single forward pass of a 22M parameter model, as opposed to 84% for a 27B LLM. Fine-tuning yields state-of-the-art results with high sample efficiency. Our experiments show that RT’s zero-shot transfer harnesses task-table context, relational attention patterns and schema semantics. Overall, RT provides a practical path toward foundation models for relational data.

Method: GAOT uses multiscale attentional graph neural operator encoders and decoders, geometry embeddings, and vision transformer processors to map domain information and inputs into PDE solutions. The implementation includes multiple innovations for computational efficiency and scalability.

Result: GAOT demonstrates significant gains in both accuracy and efficiency compared to several baselines across diverse PDE learning tasks. It achieves state-of-the-art performance on three large-scale 3D industrial CFD datasets.

Conclusion: The proposed GAOT framework successfully addresses the accuracy-efficiency trade-off in PDE operator learning, providing a robust and scalable solution for industrial simulations on arbitrary domains.

Abstract: The very challenging task of learning solution operators of PDEs on arbitrary domains accurately and efficiently is of vital importance to engineering and industrial simulations. Despite the existence of many operator learning algorithms to approximate such PDEs, we find that accurate models are not necessarily computationally efficient and vice versa. We address this issue by proposing a geometry aware operator transformer (GAOT) for learning PDEs on arbitrary domains. GAOT combines novel multiscale attentional graph neural operator encoders and decoders, together with geometry embeddings and (vision) transformer processors to accurately map information about the domain and the inputs into a robust approximation of the PDE solution. Multiple innovations in the implementation of GAOT also ensure computational efficiency and scalability. We demonstrate this significant gain in both accuracy and efficiency of GAOT over several baselines on a large number of learning tasks from a diverse set of PDEs, including achieving state of the art performance on three large scale three-dimensional industrial CFD datasets.

Method: 1) A novel membership inference game to efficiently audit approximate worst-case privacy risks of data records. 2) Enhanced DP-SGD algorithm with adaptive group-specific gradient clipping strategy inspired by differential privacy auditing canaries.

Result: Experimental results show the method provides more stringent measurement of group privacy risks and reliable assessment of disparity. The enhanced DP-SGD algorithm effectively reduces disparity in group privacy risks, improving fairness of privacy protection in DPML.

Conclusion: The proposed approaches successfully address the limitations in current group privacy risk assessment and enhance fairness in privacy protection for differentially private machine learning systems.

Abstract: While significant progress has been made in conventional fairness-aware machine learning (ML) and differentially private ML (DPML), the fairness of privacy protection across groups remains underexplored. Existing studies have proposed methods to assess group privacy risks, but these are based on the average-case privacy risks of data records. Such approaches may underestimate the group privacy risks, thereby potentially underestimating the disparity across group privacy risks. Moreover, the current method for assessing the worst-case privacy risks of data records is time-consuming, limiting their practical applicability. To address these limitations, we introduce a novel membership inference game that can efficiently audit the approximate worst-case privacy risks of data records. Experimental results demonstrate that our method provides a more stringent measurement of group privacy risks, yielding a reliable assessment of the disparity in group privacy risks. Furthermore, to promote privacy protection fairness in DPML, we enhance the standard DP-SGD algorithm with an adaptive group-specific gradient clipping strategy, inspired by the design of canaries in differential privacy auditing studies. Extensive experiments confirm that our algorithm effectively reduces the disparity in group privacy risks, thereby enhancing the fairness of privacy protection in DPML.

Method: Conducted a large-scale empirical study on hyperparameter sensitivity of common data attribution methods, performed theoretical analysis of regularization terms in influence function methods, and proposed a lightweight procedure for selecting regularization values without model retraining.

Result: Most data attribution methods are sensitive to certain key hyperparameters, but evaluating performance requires costly model retraining, creating a practical challenge. The proposed lightweight regularization selection procedure was validated as effective across standard benchmarks.

Conclusion: The study identifies a fundamental but overlooked challenge in practical data attribution applications and emphasizes the importance of careful hyperparameter selection discussions in future method development.

Abstract: Data attribution methods, which quantify the influence of individual training data points on a machine learning model, have gained increasing popularity in data-centric applications in modern AI. Despite a recent surge of new methods developed in this space, the impact of hyperparameter tuning in these methods remains under-explored. In this work, we present the first large-scale empirical study to understand the hyperparameter sensitivity of common data attribution methods. Our results show that most methods are indeed sensitive to certain key hyperparameters. However, unlike typical machine learning algorithms – whose hyperparameters can be tuned using computationally-cheap validation metrics – evaluating data attribution performance often requires retraining models on subsets of training data, making such metrics prohibitively costly for hyperparameter tuning. This poses a critical open challenge for the practical application of data attribution methods. To address this challenge, we advocate for better theoretical understandings of hyperparameter behavior to inform efficient tuning strategies. As a case study, we provide a theoretical analysis of the regularization term that is critical in many variants of influence function methods. Building on this analysis, we propose a lightweight procedure for selecting the regularization value without model retraining, and validate its effectiveness across a range of standard data attribution benchmarks. Overall, our study identifies a fundamental yet overlooked challenge in the practical application of data attribution, and highlights the importance of careful discussion on hyperparameter selection in future method development.

Method: Proposed three methods to blend KV-memory and FW-memory into a single system, differing in how/when input information is delivered to each system. Evaluated on language modeling, retrieval tasks, synthetic algorithmic tasks, and reinforcement learning in partially observable environments using 340M- and 1.3B-parameter models.

Result: Hybrid memory systems demonstrated improved performance by overcoming individual limitations of KV-memory and FW-memory components.

Conclusion: Well-designed hybrid memory architectures can overcome limitations of individual components, providing new insights into neural memory system design principles.

Abstract: We develop hybrid memory architectures for general-purpose sequence processing neural networks, that combine key-value memory using softmax attention (KV-memory) with fast weight memory through dynamic synaptic modulation (FW-memory) – the core principles of quadratic and linear transformers, respectively. These two memory systems have complementary but individually limited properties: KV-memory offers precise retrieval but is constrained by quadratic complexity in sequence length, while FW-memory supports arbitrarily long sequences and enables more expressive computation but sacrifices precise recall. We propose and compare three methods to blend these two systems into a single memory system, differing in how and when input information is delivered to each system, to leverage the strengths of both. We conduct experiments on general language modeling and retrieval tasks by training 340M- and 1.3B-parameter models from scratch, as well as on synthetic algorithmic tasks designed to precisely illustrate the benefits of certain hybrid methods over others. We also evaluate our hybrid memory systems on reinforcement learning in partially observable environments. Overall, we demonstrate how a well-designed hybrid can overcome the limitations of its individual components, offering new insights into the design principle of neural memory systems.

Method: Uses Kalman-based optimization that recursively updates compact gradient covariance products to estimate parameter covariance matrix, avoiding storage of full covariance matrix and large matrix inversions.

Result: Achieves accuracy on par or better than state-of-the-art first- and second-order optimizers across diverse tasks including image classification and language modeling.

Conclusion: KOALA++ maintains the efficiency of first-order methods while achieving competitive or superior performance compared to both first- and second-order optimization approaches.

Abstract: We propose KOALA++, a scalable Kalman-based optimization algorithm that explicitly models structured gradient uncertainty in neural network training. Unlike second-order methods, which rely on expensive second order gradient calculation, our method directly estimates the parameter covariance matrix by recursively updating compact gradient covariance products. This design improves upon the original KOALA framework that assumed diagonal covariance by implicitly capturing richer uncertainty structure without storing the full covariance matrix and avoiding large matrix inversions. Across diverse tasks, including image classification and language modeling, KOALA++ achieves accuracy on par or better than state-of-the-art first- and second-order optimizers while maintaining the efficiency of first-order methods.

Method: ROBIN integrates two innovations: (1) Rolling Diffusion-Batched Inference (ROBI) - a parallelized inference scheme that amortizes diffusion-based refinement costs across physical time steps, and (2) Hierarchical Graph Neural Network using algebraic multigrid coarsening for multiscale message passing across different mesh resolutions.

Result: ROBIN achieves state-of-the-art accuracy on challenging 2D and 3D solid mechanics benchmarks with geometric, material, and contact nonlinearities, while reducing inference time by up to an order of magnitude compared to standard diffusion simulators.

Conclusion: ROBIN successfully addresses limitations of existing learned simulators by combining rolling diffusion inference with hierarchical graph networks, enabling efficient and accurate simulation of complex solid mechanics phenomena.

Abstract: Graph-based learned simulators have emerged as a promising approach for simulating physical systems on unstructured meshes, offering speed and generalization across diverse geometries. However, they often struggle with capturing global phenomena, such as bending or long-range correlations usually occurring in solid mechanics, and suffer from error accumulation over long rollouts due to their reliance on local message passing and direct next-step prediction. We address these limitations by introducing the Rolling Diffusion-Batched Inference Network (ROBIN), a novel learned simulator that integrates two key innovations: (i) Rolling Diffusion-Batched Inference (ROBI), a parallelized inference scheme that amortizes the cost of diffusion-based refinement across physical time steps by overlapping denoising steps across a temporal window. (ii) A Hierarchical Graph Neural Network built on algebraic multigrid coarsening, enabling multiscale message passing across different mesh resolutions. This architecture, implemented via Algebraic-hierarchical Message Passing Networks, captures both fine-scale local dynamics and global structural effects critical for phenomena like beam bending or multi-body contact. We validate ROBIN on challenging 2D and 3D solid mechanics benchmarks involving geometric, material, and contact nonlinearities. ROBIN achieves state-of-the-art accuracy on all tasks, substantially outperforming existing next-step learned simulators while reducing inference time by up to an order of magnitude compared to standard diffusion simulators.

Method: Comprehensive empirical study using One-vs-Rest logistic regression on UCI Wine dataset (178 samples, 3 cultivars, 13 chemical features), comparing manual gradient descent against scikit-learn solvers and analyzing L1 regularization effects.

Result: Manual gradient descent achieved 92.59% accuracy, scikit-learn provided 98.15% accuracy with 24x speedup. L1 regularization reduced features by 54-69% with only 4.63% accuracy decrease. Optimal 5-feature subset achieved 62% complexity reduction with 92-94% accuracy.

Conclusion: The findings provide actionable guidelines for balancing comprehensive chemical analysis against targeted feature measurement in resource-constrained environments, enabling cost-effective deployment with significant time and cost savings.

Abstract: Multi-class wine classification presents fundamental trade-offs between model accuracy, feature dimensionality, and interpretability - critical factors for production deployment in analytical chemistry. This paper presents a comprehensive empirical study of One-vs-Rest logistic regression on the UCI Wine dataset (178 samples, 3 cultivars, 13 chemical features), comparing from-scratch gradient descent implementation against scikit-learn’s optimized solvers and quantifying L1 regularization effects on feature sparsity. Manual gradient descent achieves 92.59 percent mean test accuracy with smooth convergence, validating theoretical foundations, though scikit-learn provides 24x training speedup and 98.15 percent accuracy. Class-specific analysis reveals distinct chemical signatures with heterogeneous patterns where color intensity varies dramatically (0.31 to 16.50) across cultivars. L1 regularization produces 54-69 percent feature reduction with only 4.63 percent accuracy decrease, demonstrating favorable interpretability-performance trade-offs. We propose an optimal 5-feature subset achieving 62 percent complexity reduction with estimated 92-94 percent accuracy, enabling cost-effective deployment with 80 dollars savings per sample and 56 percent time reduction. Statistical validation confirms robust generalization with sub-2ms prediction latency suitable for real-time quality control. Our findings provide actionable guidelines for practitioners balancing comprehensive chemical analysis against targeted feature measurement in resource-constrained environments.

Method: Uses top-k masking with statistical top-k algorithm for explicit sparsity control, reallocates existing FFN parameters and attention key embeddings to form low-cost predictors for activated entries, and maintains standard training procedures.

Result: Achieves 92% sparsity (only 8% FFN neurons activated), each token attends to max 256 tokens, 2.5x FLOPs reduction, with decoding speedups of 1.79x on CPU and 1.40x on GPU while maintaining competitive performance on benchmarks.

Conclusion: Spark Transformer successfully achieves high activation sparsity in both FFN and attention mechanisms without compromising model quality, parameter count, or training procedures, providing significant computational efficiency gains.

Abstract: The discovery of the lazy neuron phenomenon in trained Transformers, where the vast majority of neurons in their feed-forward networks (FFN) are inactive for each token, has spurred tremendous interests in activation sparsity for enhancing large model efficiency. While notable progress has been made in translating such sparsity to wall-time benefits, modern Transformers have moved away from the ReLU activation function crucial to this phenomenon. Existing efforts on re-introducing activation sparsity often degrade model quality, increase parameter count, complicate or slow down training. Sparse attention, the application of sparse activation to the attention mechanism, often faces similar challenges. This paper introduces the Spark Transformer, a novel architecture that achieves a high level of activation sparsity in both FFN and the attention mechanism while maintaining model quality, parameter count, and standard training procedures. Our method realizes sparsity via top-k masking for explicit control over sparsity level. Crucially, we introduce statistical top-k, a hardware-accelerator-friendly, linear-time approximate algorithm that avoids costly sorting and mitigates significant training slowdown from standard top-$k$ operators. Furthermore, Spark Transformer reallocates existing FFN parameters and attention key embeddings to form a low-cost predictor for identifying activated entries. This design not only mitigates quality loss from enforced sparsity, but also enhances wall-time benefit. Pretrained with the Gemma-2 recipe, Spark Transformer demonstrates competitive performance on standard benchmarks while exhibiting significant sparsity: only 8% of FFN neurons are activated, and each token attends to a maximum of 256 tokens. This sparsity translates to a 2.5x reduction in FLOPs, leading to decoding wall-time speedups of up to 1.79x on CPU and 1.40x on GPU.

Method: MIRA incorporates three key components: 1) Continuous-Time Rotary Positional Encoding for variable time intervals, 2) frequency-specific mixture-of-experts layer for temporal specialization, and 3) Continuous Dynamics Extrapolation Block based on Neural ODE for modeling continuous latent state trajectories.

Result: Pretrained on 454+ billion time points, MIRA reduces forecasting errors by 10% in out-of-distribution and 7% in in-distribution scenarios compared to zero-shot and fine-tuned baselines.

Conclusion: MIRA establishes a comprehensive benchmark for medical time series modeling and demonstrates superior performance in handling medical time series forecasting challenges.

Abstract: A unified foundation model for medical time series – pretrained on open access and ethics board-approved medical corpora – offers the potential to reduce annotation burdens, minimize model customization, and enable robust transfer across clinical institutions, modalities, and tasks, particularly in data-scarce or privacy-constrained environments. However, existing generalist time series foundation models struggle to handle medical time series data due to their inherent challenges, including irregular intervals, heterogeneous sampling rates, and frequent missing values. To address these challenges, we introduce MIRA, a unified foundation model specifically designed for medical time series forecasting. MIRA incorporates a Continuous-Time Rotary Positional Encoding that enables fine-grained modeling of variable time intervals, a frequency-specific mixture-of-experts layer that routes computation across latent frequency regimes to further promote temporal specialization, and a Continuous Dynamics Extrapolation Block based on Neural ODE that models the continuous trajectory of latent states, enabling accurate forecasting at arbitrary target timestamps. Pretrained on a large-scale and diverse medical corpus comprising over 454 billion time points collect from publicly available datasets, MIRA achieves reductions in forecasting errors by an average of 10% and 7% in out-of-distribution and in-distribution scenarios, respectively, when compared to other zero-shot and fine-tuned baselines. We also introduce a comprehensive benchmark spanning multiple downstream clinical tasks, establishing a foundation for future research in medical time series modeling.

Method: BREEZE introduces behavioral regularization to transform policy optimization into stable in-sample learning, uses task-conditioned diffusion models for policy extraction to generate multimodal action distributions, and employs expressive attention-based architectures for representation modeling.

Result: Extensive experiments on ExORL and D4RL Kitchen benchmarks show BREEZE achieves best or near-best performance while demonstrating superior robustness compared to prior offline zero-shot RL methods.

Conclusion: BREEZE successfully addresses key limitations in zero-shot RL by enhancing learning stability, policy extraction capability, and representation quality through its integrated framework of behavioral regularization, diffusion models, and attention architectures.

Abstract: The recent development of zero-shot reinforcement learning (RL) has opened a new avenue for learning pre-trained generalist policies that can adapt to arbitrary new tasks in a zero-shot manner. While the popular Forward-Backward representations (FB) and related methods have shown promise in zero-shot RL, we empirically found that their modeling lacks expressivity and that extrapolation errors caused by out-of-distribution (OOD) actions during offline learning sometimes lead to biased representations, ultimately resulting in suboptimal performance. To address these issues, we propose Behavior-REgularizEd Zero-shot RL with Expressivity enhancement (BREEZE), an upgraded FB-based framework that simultaneously enhances learning stability, policy extraction capability, and representation learning quality. BREEZE introduces behavioral regularization in zero-shot RL policy learning, transforming policy optimization into a stable in-sample learning paradigm. Additionally, BREEZE extracts the policy using a task-conditioned diffusion model, enabling the generation of high-quality and multimodal action distributions in zero-shot RL settings. Moreover, BREEZE employs expressive attention-based architectures for representation modeling to capture the complex relationships between environmental dynamics. Extensive experiments on ExORL and D4RL Kitchen demonstrate that BREEZE achieves the best or near-the-best performance while exhibiting superior robustness compared to prior offline zero-shot RL methods. The official implementation is available at: https://github.com/Whiterrrrr/BREEZE.

Method: Multi-stage process: beam search for candidate completions, execute against test cases, incorporate execution signals into prompts during generation with consistent signals per line.

Result: Significantly improves code generation performance compared to standard approaches, achieving state-of-the-art results across various complexity levels.

Conclusion: Execution-guided generation with real-time feedback enables more executable code solutions and supports parallel exploration of diverse reasoning paths.

Abstract: We present a novel approach to neural code generation that incorporates real-time execution signals into the language model generation process. While large language models (LLMs) have demonstrated impressive code generation capabilities, they typically do not utilize execution feedback during inference, a critical signal that human programmers regularly leverage. Our method, Execution-Guided Classifier-Free Guidance (EG-CFG), dynamically incorporates execution signals as the model generates code, providing line-by-line feedback that guides the generation process toward executable solutions. EG-CFG employs a multi-stage process: first, we conduct beam search to sample candidate program completions for each line; second, we extract execution signals by executing these candidates against test cases; and finally, we incorporate these signals into the prompt during generation. By maintaining consistent signals across tokens within the same line and refreshing signals at line boundaries, our approach provides coherent guidance while preserving syntactic structure. Moreover, the method naturally supports native parallelism at the task level in which multiple agents operate in parallel, exploring diverse reasoning paths and collectively generating a broad set of candidate solutions. Our experiments across diverse coding tasks demonstrate that EG-CFG significantly improves code generation performance compared to standard approaches, achieving state-of-the-art results across various levels of complexity, from foundational problems to challenging competitive programming and data science tasks. Our code is available at: https://github.com/boazlavon/eg_cfg

Method: Two symmetry-aware architectures: parallel design (independent processing of standard and equivariant features with fusion) and cascaded design (sequential equivariant operations). Theoretical analysis shows reduced hypothesis space complexity and better gradient regularization.

Result: Models consistently improve adversarial robustness and generalization on CIFAR-10, CIFAR-100, and CIFAR-10C under FGSM and PGD attacks, without adversarial training. Tighter certified robustness bounds under CLEVER framework.

Conclusion: Symmetry-enforcing architectures offer efficient and principled alternatives to data augmentation-based defenses, demonstrating the potential of architectural approaches for adversarial robustness.

Abstract: Adversarial examples reveal critical vulnerabilities in deep neural networks by exploiting their sensitivity to imperceptible input perturbations. While adversarial training remains the predominant defense strategy, it often incurs significant computational cost and may compromise clean-data accuracy. In this work, we investigate an architectural approach to adversarial robustness by embedding group-equivariant convolutions-specifically, rotation- and scale-equivariant layers-into standard convolutional neural networks (CNNs). These layers encode symmetry priors that align model behavior with structured transformations in the input space, promoting smoother decision boundaries and greater resilience to adversarial attacks. We propose and evaluate two symmetry-aware architectures: a parallel design that processes standard and equivariant features independently before fusion, and a cascaded design that applies equivariant operations sequentially. Theoretically, we demonstrate that such models reduce hypothesis space complexity, regularize gradients, and yield tighter certified robustness bounds under the CLEVER (Cross Lipschitz Extreme Value for nEtwork Robustness) framework. Empirically, our models consistently improve adversarial robustness and generalization across CIFAR-10, CIFAR-100, and CIFAR-10C under both FGSM and PGD attacks, without requiring adversarial training. These findings underscore the potential of symmetry-enforcing architectures as efficient and principled alternatives to data augmentation-based defenses.

Method: Analyzed shallow Transformers during training plateaus, examining output patterns, internal representations, and attention map evolution. Also validated findings on large language models (Pythia, OLMo) during early pre-training.

Result: During plateaus, models develop partial solutions but exhibit strong repetition bias and representation collapse (parallel hidden states). Slow attention learning is the bottleneck, with hidden progress in attention preceding rapid convergence. Attention interventions alter plateau duration and severity of these phenomena.

Conclusion: Abrupt learning in Transformers stems from slow attention optimization, with plateaus characterized by repetition bias and representation collapse that resolve when attention maps finally converge.

Abstract: Training Transformers on algorithmic tasks frequently demonstrates an intriguing abrupt learning phenomenon: an extended performance plateau followed by a sudden, sharp improvement. This work investigates the underlying mechanisms for such dynamics, primarily in shallow Transformers. We reveal that during the plateau, the model often develops an interpretable partial solution while simultaneously exhibiting a strong repetition bias in their outputs. This output degeneracy is accompanied by internal representation collapse, where hidden states across different tokens become nearly parallel. We further identify the slow learning of optimal attention maps as a key bottleneck. Hidden progress in attention configuration during the plateau precedes the eventual rapid convergence, and directly intervening on attention significantly alters plateau duration and the severity of repetition bias and representational collapse. We validate that these identified phenomena-repetition bias and representation collapse-are not artifacts of toy setups but also manifest in the early pre-training stage of large language models like Pythia and OLMo.

Method: Reuses half of Value heads from the first layer in deeper layers while computing the other half normally, reducing Value projections and V cache by nearly 50%. Theoretically restores information lost to compression and accelerates implicit mesa-optimization.

Result: Consistent ~25% KV cache reduction across different model scales while improving perplexity compared to standard MHA Transformers and advanced variants. Can be combined with other methods like GQA and MLA for further savings.

Conclusion: SkipV1Former effectively balances representation quality and resource efficiency, offering practical KV cache reduction with performance improvements, and provides a cost-effective uptraining path for existing models.

Abstract: Transformer models have driven breakthroughs across various language tasks by their strong capability to learn rich contextual representations. Scaling them to improve representation, however, often demands substantial memory and compute costs, such as the Key-Value (KV) cache used during auto-regressive decoding. Skip connections offer a promising way to improve representation without bloating resource usage, yet most prior works either improve expressivity while leaving KV costs unchanged, or reduce memory at the cost of weaker representation. In this work, we propose SkipV1Former, a Transformer variant that uses skip connections from the first layer’s Value heads to strengthen model representation and reduce KV cache. Specifically, from the second block onward, each layer reuses half of its Value heads from the very first layer, while computing the other half as usual-cutting Value projections and V cache by nearly 50 %. Theoretically, we show that routing uncompressed first-layer Values into deeper layers restores information lost to compression and accelerates the model’s implicit mesa-optimization-a key pattern of Transformer in auto-regressive tasks. Empirically, across different model scales, SkipV1Former delivers consistent reductions of approximately 25 % in KV cache while improving perplexity relative to standard Multi-Head Attention (MHA) Transformers and some advanced variants. Moreover, we propose a recipe for uptraining existing MHA Transformer checkpoints to SkipV1Former with only 10-15% additional compute. Finally, SkipV1Former can seamlessly combine advanced methods like Group-Query Attention and Multi-Latent Attention to achieve further KV cache savings and performance improvement. When combined with YOCO, it cuts KV cache size by nearly 50 % while still improving performance.

Method: Adapts Kolmogorov-Arnold representation theorem to constrain priors to univariate relationships, enabling fast exact inference via inverse transform method and efficient importance sampling in low-dimensional latent spaces.

Result: T-KAM achieves fast inference, interpretability, stable training, and efficient multimodal sampling through importance sampling and population-based LMC, while being compatible with future hardware.

Conclusion: T-KAM elegantly balances trade-offs in generative modeling by offering interpretability, efficiency, and stability while leveraging structural biases for improved performance.

Abstract: Learning an energy-based model (EBM) in the latent space of a top-down generative model offers a versatile framework for generation across multiple data modalities. However, it remains unclear how its interpretability can be used to guide model design, improve generative quality, and reduce training time. Moreover, the reliance on Langevin Monte Carlo (LMC) sampling presents challenges in efficiency and sampling multimodal latent distributions. In this work, we propose a novel adaptation of the Kolmogorov-Arnold representation theorem for generative modeling and introduce the Thermodynamic Kolmogorov-Arnold Model (T-KAM) to take advantage of structural and inductive biases. By constraining the prior to univariate relationships, T-KAM enables fast and exact inference via the inverse transform method. With the low dimensionality of the latent space and suitable inductive biases encoded, we demonstrate that importance sampling (IS) becomes a viable, unbiased, and highly efficient posterior sampler. For situations where IS fails, we investigate a novel strategy using population-based LMC, which decomposes posterior sampling into a sequence of annealed distributions to improve multimodal sampling. T-KAM elegantly balances common trade-offs in generative modeling, offering fast inference, interpretability, and stable training, while being naturally suited to upcoming Zettascale Computing Corp. hardware.

Method: Systematically investigates required symmetries for graph foundation models, characterizes linear transformations equivariant to node/label permutations and invariant to feature permutations, proves universality of resulting networks on multisets, and applies these layers to local graph neighborhoods.

Result: Extensive experiments on 29 real-world node classification datasets show strong zero-shot empirical performance and consistent improvement as the number of training graphs increases.

Conclusion: The proposed symmetry-based recipe successfully builds graph foundation models that generalize well across different graphs and features, demonstrating both strong zero-shot performance and scalability with more training data.

Abstract: Graph machine learning architectures are typically tailored to specific tasks on specific datasets, which hinders their broader applicability. This has led to a new quest in graph machine learning: how to build graph foundation models capable of generalizing across arbitrary graphs and features? In this work, we present a recipe for designing graph foundation models for node-level tasks from first principles. The key ingredient underpinning our study is a systematic investigation of the symmetries that a graph foundation model must respect. In a nutshell, we argue that label permutation-equivariance alongside feature permutation-invariance are necessary in addition to the common node permutation-equivariance on each local neighborhood of the graph. To this end, we first characterize the space of linear transformations that are equivariant to permutations of nodes and labels, and invariant to permutations of features. We then prove that the resulting network is a universal approximator on multisets that respect the aforementioned symmetries. Our recipe uses such layers on the multiset of features induced by the local neighborhood of the graph to obtain a class of graph foundation models for node property prediction. We validate our approach through extensive experiments on 29 real-world node classification datasets, demonstrating both strong zero-shot empirical performance and consistent improvement as the number of training graphs increases.

Method: Uses Permutation Relative Policy Optimization (PRPO), a reinforcement learning method that encodes column-permutation invariance as structural prior, constructing multiple label-preserving permutations per sample and estimating advantages within and across permutations.

Result: TabR1 achieves performance comparable to strong baselines under full-supervision, approaches 32-shot baseline performance in zero-shot setting, and substantially outperforms much larger LLMs (53.17% improvement over DeepSeek-R1 685B).

Conclusion: PRPO effectively activates LLM reasoning abilities for tabular prediction, enhancing few-shot/zero-shot performance and interpretability while maintaining competitive performance with limited supervision.

Abstract: Tabular prediction has traditionally relied on gradient-boosted decision trees and specialized deep learning models, which excel within tasks but provide limited interpretability and weak transfer across tables. Reasoning large language models (LLMs) promise cross-task adaptability with trans- parent reasoning traces, yet their potential has not been fully realized for tabular data. This paper presents TabR1, the first reasoning LLM for tabular prediction with multi-step reasoning. At its core is Permutation Relative Policy Optimization (PRPO), a simple yet efficient reinforcement learning method that encodes column-permutation invariance as a structural prior. By construct- ing multiple label-preserving permutations per sample and estimating advantages both within and across permutations, PRPO transforms sparse rewards into dense learning signals and improves generalization. With limited supervision, PRPO activates the reasoning ability of LLMs for tabular prediction, enhancing few-shot and zero-shot performance as well as interpretability. Comprehensive experiments demonstrate that TabR1 achieves performance comparable to strong baselines under full-supervision fine-tuning. In the zero-shot setting, TabR1 approaches the performance of strong baselines under the 32-shot setting. Moreover, TabR1 (8B) substantially outperforms much larger LLMs across various tasks, achieving up to 53.17% improvement over DeepSeek-R1 (685B).

Method: Applies SVD to pre-trained weight matrices and introduces learnable routers to adjust singular values. Creates orthogonal rank-one experts from singular vectors, with optional orthogonal transforms for capacity enhancement.

Result: MoORE consistently outperforms existing multi-task adaptation methods across various datasets, demonstrating superior conflict- and oblivion-resistance.

Conclusion: MoORE provides an effective solution for multi-task adaptation that maintains original task performance while learning new tasks, with guaranteed orthogonality and preservation of original weight matrix properties.

Abstract: Adapting large-scale foundation models in multi-task scenarios often suffers from task conflict and oblivion. To mitigate such issues, we propose a novel ‘‘model MoE-ization’’ strategy that leads to a conflict- and oblivion-resistant multi-task adaptation method. Given a weight matrix of a pre-trained model, our method applies SVD to it and introduces a learnable router to adjust its singular values based on tasks and samples. Accordingly, the weight matrix becomes a Mixture of Orthogonal Rank-one Experts (MoORE), in which each expert corresponds to the outer product of a left singular vector and the corresponding right one. We can improve the model capacity by imposing a learnable orthogonal transform on the right singular vectors. Unlike low-rank adaptation (LoRA) and its MoE-driven variants, MoORE guarantees the experts’ orthogonality and maintains the column space of the original weight matrix. These two properties make the adapted model resistant to the conflicts among the new tasks and the oblivion of its original tasks, respectively. Experiments on various datasets demonstrate that MoORE outperforms existing multi-task adaptation methods consistently, showing its superiority in terms of conflict- and oblivion-resistance. The code of the experiments is available at https://github.com/DaShenZi721/MoORE.

Method: Propose scaling rule for Adam hyperparameters where the decay rate of second moment is adjusted to maintain fixed half-life in terms of tokens across different batch sizes, enabling stable training with very small batches including batch size 1.

Result: Small batch sizes train stably, are more robust to hyperparameter choices, achieve equal or better per-FLOP performance than larger batches, and enable stable training with vanilla SGD without momentum. Small batches with state-efficient optimizers can match full fine-tuning performance while maintaining LoRA-like memory footprint.

Conclusion: Recommend using small batch sizes with properly scaled Adam hyperparameters, advise against gradient accumulation unless training on multiple devices, and show small batches with state-efficient optimizers provide full fine-tuning benefits with low memory usage.

Abstract: Conventional wisdom dictates that small batch sizes make language model pretraining and fine-tuning unstable, motivating gradient accumulation, which trades off the number of optimizer steps for a proportional increase in batch size. While it is common to decrease the learning rate for smaller batch sizes, other hyperparameters are often held fixed. In this work, we revisit small batch sizes all the way down to batch size one, and we propose a rule for scaling Adam hyperparameters to small batch sizes. In particular, rather than holding the decay rate of the second moment fixed across batch sizes, we propose to hold its half-life fixed in terms of tokens. We find that small batch sizes (1) train stably, (2) are consistently more robust to hyperparameter choices, (3) achieve equal or better per-FLOP performance than larger batch sizes, and (4) notably enable stable language model training with vanilla SGD, even without momentum, despite storing no optimizer state. Building on these results, we provide practical recommendations for selecting a batch size and setting optimizer hyperparameters. We further recommend against gradient accumulation unless training on multiple devices with multiple model replicas. Finally, we show that a small batch size combined with an optimizer with a small state size can provide the performance benefits of full fine-tuning while maintaining a similar memory footprint to LoRA.

Method: xRFM combines feature learning kernel machines with a tree structure to adapt to local data structure and scale to unlimited training data. It uses Average Gradient Outer Product for native interpretability.

Result: xRFM achieves best performance across 100 regression datasets and is competitive with best methods across 200 classification datasets, outperforming GBDTs. It was tested against 31 other methods including TabPFNv2.

Conclusion: xRFM provides a powerful alternative to traditional GBDTs for tabular data, offering superior performance on regression tasks and competitive classification performance while maintaining interpretability through its native feature importance mechanism.

Abstract: Inference from tabular data, collections of continuous and categorical variables organized into matrices, is a foundation for modern technology and science. Yet, in contrast to the explosive changes in the rest of AI, the best practice for these predictive tasks has been relatively unchanged and is still primarily based on variations of Gradient Boosted Decision Trees (GBDTs). Very recently, there has been renewed interest in developing state-of-the-art methods for tabular data based on recent developments in neural networks and feature learning methods. In this work, we introduce xRFM, an algorithm that combines feature learning kernel machines with a tree structure to both adapt to the local structure of the data and scale to essentially unlimited amounts of training data. We show that compared to $31$ other methods, including recently introduced tabular foundation models (TabPFNv2) and GBDTs, xRFM achieves best performance across $100$ regression datasets and is competitive to the best methods across $200$ classification datasets outperforming GBDTs. Additionally, xRFM provides interpretability natively through the Average Gradient Outer Product.

Method: Learn a probabilistic bridge that transforms distributions from low-value to high-value inputs using synthetic functions constructed as mean posteriors of multiple Gaussian processes fitted on offline data.

Result: The proposed approach achieves significant improvement over recent methods and establishes new state-of-the-art performance on extensive benchmarks.

Conclusion: Framing offline optimization as distributional translation with synthetic functions effectively mitigates data limitations and advances the field.

Abstract: This paper studies the black-box optimization task which aims to find the maxima of a black-box function using a static set of its observed input-output pairs. This is often achieved via learning and optimizing a surrogate function with that offline data. Alternatively, it can also be framed as an inverse modeling task that maps a desired performance to potential input candidates that achieve it. Both approaches are constrained by the limited amount of offline data. To mitigate this limitation, we introduce a new perspective that casts offline optimization as a distributional translation task. This is formulated as learning a probabilistic bridge transforming an implicit distribution of low-value inputs (i.e., offline data) into another distribution of high-value inputs (i.e., solution candidates). Such probabilistic bridge can be learned using low- and high-value inputs sampled from synthetic functions that resemble the target function. These synthetic functions are constructed as the mean posterior of multiple Gaussian processes fitted with different parameterizations on the offline data, alleviating the data bottleneck. The proposed approach is evaluated on an extensive benchmark comprising most recent methods, demonstrating significant improvement and establishing a new state-of-the-art performance. Our code is publicly available at https://github.com/cuong-dm/ROOT.

Method: Proposes Arbitrary Entropy Policy Optimization (AEPO) which replaces entropy bonuses with REINFORCE policy gradient on temperature-adjusted distributions, using three key designs: policy gradient as regularization, distribution as regularization, and REINFORCE as regularization.

Result: AEPO stabilizes entropy at arbitrary target levels, effectively removing collapse in GRPO; reveals a non-monotonic relationship where performance first improves then declines with increasing entropy; and generalizes beyond entropy to provide a broader RFT paradigm.

Conclusion: AEPO successfully resolves entropy collapse in reinforcement fine-tuning, clarifies the connection between entropy, exploration, and reasoning performance, and provides a more general framework for policy optimization.

Abstract: Reinforcement fine-tuning (RFT) is essential for enhancing the reasoning capabilities of large language models (LLM), yet the widely adopted Group Relative Policy Optimization (GRPO) suffers from entropy collapse, where entropy monotonically decreases, exploration vanishes, and policies converge prematurely. Existing entropy-regularized methods only partially alleviate this issue while introducing bias and instability, leaving entropy control unresolved and the connection between entropy, exploration, and performance unclear. We propose Arbitrary Entropy Policy Optimization (AEPO), which eliminates entropy collapse by replacing entropy bonuses with REINFORCE policy gradient on temperature-adjusted distributions and stabilizing entropy through temperature regulation. AEPO integrates three key designs: policy gradient as regularization, distribution as regularization, and REINFORCE as regularization, enabling precise entropy control without distorting optimization. Experiments demonstrate three major contributions: AEPO (1) stabilizes entropy at arbitrary target levels, effectively removing collapse in GRPO; (2) reveals a non-monotonic relation where performance first improves then declines with increasing entropy, clarifying the link between entropy, exploration, and reasoning; and (3) generalizes beyond entropy, providing a broader RFT paradigm where superior target distributions can serve as REINFORCE regularizers.

Method: The authors propose two continuous distance functions designed to theoretically overcome the limitations of traditional crystal distance functions used in evaluating generative AI models.

Result: Experiments show that the proposed continuous distance functions reveal insights missed by traditional distance functions, providing more reliable evaluation of generative models for inorganic crystals.

Conclusion: The new continuous distance functions offer a more robust and insightful basis for evaluating and comparing generative AI models in materials science, particularly for inorganic crystal generation.

Abstract: To address pressing scientific challenges such as climate change, increasingly sophisticated generative artificial intelligence models are being developed that can efficiently sample the large chemical space of possible functional materials. These models can quickly sample new chemical compositions paired with crystal structures. They are typically evaluated using uniqueness and novelty metrics, which depend on a chosen crystal distance function. However, the most prevalent distance function has four limitations: it fails to quantify the degree of similarity between compounds, cannot distinguish compositional difference and structural difference, lacks Lipschitz continuity against shifts in atomic coordinates, and results in a uniqueness metric that is not invariant against the permutation of generated samples. In this work, we propose using two continuous distance functions to evaluate uniqueness and novelty, which theoretically overcome these limitations. Our experiments show that these distances reveal insights missed by traditional distance functions, providing a more reliable basis for evaluating and comparing generative models for inorganic crystals.

Method: Uses Gated Recurrent Units (GRUs) for temporal patterns, Directed Acyclic Graph (DAG) learning for causal structures, Hill equation-based saturation curves, multi-region modeling with shared/region-specific parameters, and robust statistical methods including Huber loss.

Result: The package enables data-driven hyperparameter learning, automatic transformation estimation, and comprehensive response curve analysis for understanding channel saturation effects.

Conclusion: DeepCausalMMM provides an advanced, automated approach to MMM that overcomes limitations of traditional methods by combining deep learning and causal inference techniques.

Abstract: Marketing Mix Modeling (MMM) is a statistical technique used to estimate the impact of marketing activities on business outcomes such as sales, revenue, or customer visits. Traditional MMM approaches often rely on linear regression or Bayesian hierarchical models that assume independence between marketing channels and struggle to capture complex temporal dynamics and non-linear saturation effects [@Chan2017; @Hanssens2005; @Ng2021Bayesian]. DeepCausalMMM is a Python package that addresses these limitations by combining deep learning, causal inference, and advanced marketing science. The package uses Gated Recurrent Units (GRUs) to automatically learn temporal patterns such as adstock (carryover effects) and lag, while simultaneously learning statistical dependencies and potential causal structures between marketing channels through Directed Acyclic Graph (DAG) learning [@Zheng2018NOTEARS; @Gong2024CausalMMM]. Additionally, it implements Hill equation-based saturation curves to model diminishing returns and optimize budget allocation. Key features include: (1) a data-driven design where hyperparameters and transformations (e.g., adstock decay, saturation curves) are learned or estimated from data with sensible defaults, rather than requiring fixed heuristics or manual specification, (2) multi-region modeling with both shared and region-specific parameters, (3) robust statistical methods including Huber loss and advanced regularization, (4) comprehensive response curve analysis for understanding channel saturation.

Method: The authors analyze various generalization measures (path norm, spectral norm, Frobenius norms, flatness proxies, PAC-Bayes bounds) by testing their sensitivity to minor hyperparameter changes like learning rate adjustments and SGD variants.

Result: Found that many generalization measures are fragile - small training changes can reverse learning curve slopes. PAC-Bayes origin measure is more robust to hyperparameter changes but fails to capture data complexity differences. Function-based marginal-likelihood PAC-Bayes captures data complexity but isn’t post-mortem.

Conclusion: Developers of new generalization measures should explicitly audit them for fragility, as many current measures provide unreliable qualitative trends due to their sensitivity to minor training modifications.

Abstract: A wide variety of generalization measures have been applied to deep neural networks (DNNs). Although obtaining tight bounds remains challenging, such measures are often assumed to reproduce qualitative generalization trends. In this position paper, we argue that many post-mortem generalization measures – those computed on trained networks – are \textbf{fragile}: small training modifications that barely affect the underlying DNN can substantially change a measure’s value, trend, or scaling behavior. For example, minor hyperparameter changes, such as learning rate adjustments or switching between SGD variants can reverse the slope of a learning curve in widely used generalization measures like the path norm. We also identify subtler forms of fragility. For instance, the PAC-Bayes origin measure is regarded as one of the most reliable, and is indeed less sensitive to hyperparameter tweaks than many other measures. However, it completely fails to capture differences in data complexity across learning curves. This data fragility contrasts with the function-based marginal-likelihood PAC-Bayes bound, which does capture differences in data-complexity, including scaling behavior, in learning curves, but which is not a post-mortem measure. Beyond demonstrating that many bounds – such as path, spectral and Frobenius norms, flatness proxies, and deterministic PAC-Bayes surrogates – are fragile, this position paper also argues that developers of new measures should explicitly audit them for fragility.

Method: Cross-architecture distillation via Attention Bridge (CAB) with token-level supervision through lightweight bridge and flexible layer-wise alignment strategies to handle architectural discrepancies.

Result: Extensive experiments show CAB consistently improves SSM performance across vision and language domains, even with limited data, outperforming standard and cross-architecture distillation methods.

Conclusion: Attention-based knowledge can be efficiently transferred to recurrent models, enabling rapid utilization of Transformer expertise for building stronger SSM communities.

Abstract: State-space models (SSMs) have emerged as efficient alternatives to Transformers for sequence modeling, offering superior scalability through recurrent structures. However, their training remains costly and the ecosystem around them is far less mature than that of Transformers. Moreover, the structural heterogeneity between SSMs and Transformers makes it challenging to efficiently distill knowledge from pretrained attention models. In this work, we propose Cross-architecture distillation via Attention Bridge (CAB), a novel data-efficient distillation framework that efficiently transfers attention knowledge from Transformer teachers to state-space student models. Unlike conventional knowledge distillation that transfers knowledge only at the output level, CAB enables token-level supervision via a lightweight bridge and flexible layer-wise alignment, improving both efficiency and transferability. We further introduce flexible layer-wise alignment strategies to accommodate architectural discrepancies between teacher and student. Extensive experiments across vision and language domains demonstrate that our method consistently improves the performance of state-space models, even under limited training data, outperforming both standard and cross-architecture distillation methods. Our findings suggest that attention-based knowledge can be efficiently transferred to recurrent models, enabling rapid utilization of Transformer expertise for building a stronger SSM community.

Method: Extend single-model consistency to combine responses from multiple LLMs with diverse training data, training schemes, and model architectures. Evaluate across 15 LLMs and explore optimal teaming conditions.

Result: Substantial improvements in hallucination detection and mitigation capabilities beyond single-model consistency, often with reduced inference costs.

Conclusion: Consortium consistency approach effectively improves LLM reliability while addressing cost concerns of single-model methods.

Abstract: Recent work has demonstrated state-of-the-art results in large language model (LLM) hallucination detection and mitigation through consistency-based approaches which involve aggregating multiple responses sampled from a single LLM for a given prompt. These approaches help offset limitations stemming from the imperfect data on which LLMs are trained, which includes biases and under-representation of information required at deployment time among other limitations which can lead to hallucinations. We show that extending these single-model consistency methods to combine responses from multiple LLMs with different training data, training schemes and model architectures can result in substantial further improvements in hallucination detection and mitigation capabilities beyond their single-model consistency counterparts. We evaluate this “consortium consistency” approach across many model teams from a pool of 15 LLMs and explore under what conditions it is beneficial to team together different LLMs in this manner. Further, we show that these performance improvements often come with reduced inference costs, offsetting a significant drawback with single-model consistency methods.

Method: HSD stacks draft models in a hierarchy where each model proposes tokens and the next larger model verifies them in a single forward pass, with the target model performing final verification. The algorithm includes polynomial-time selection of latency-optimal hierarchies.

Result: Empirical results show HSD provides up to 1.2x speed-up over the best single-draft baseline, demonstrating practical latency reduction beyond previous techniques.

Conclusion: Hierarchical Speculative Decoding effectively reduces transformer inference latency by leveraging multiple draft models in a hierarchical verification structure, with provable optimal hierarchy selection.

Abstract: Transformer language models generate text autoregressively, making inference latency proportional to the number of tokens generated. Speculative decoding reduces this latency without sacrificing output quality, by leveraging a small draft model to propose tokens that the larger target model verifies in parallel. In practice, however, there may exist a set of potential draft models- ranging from faster but less inaccurate, to slower yet more reliable. We introduce Hierarchical Speculative Decoding (HSD), an algorithm that stacks these draft models into a hierarchy, where each model proposes tokens, and the next larger model verifies them in a single forward pass, until finally the target model verifies tokens. We derive an expression for the expected latency of any such hierarchy and show that selecting the latency-optimal hierarchy can be done in polynomial time. Empirically, HSD gives up to 1.2x speed-up over the best single-draft baseline, demonstrating the practicality of our algorithm in reducing generation latency beyond previous techniques.

Method: CONFEX combines Conformal Prediction (CP) with Mixed-Integer Linear Programming (MILP). It uses a novel localized CP procedure that leverages an offline tree-based partitioning of the input space, enabling efficient MILP encoding for generating counterfactuals.

Result: CONFEX generates counterfactual explanations with rigorous guarantees on both predictive uncertainty and optimality. Evaluations show it produces robust and plausible explanations compared to state-of-the-art methods across diverse benchmarks.

Conclusion: The proposed CONFEX method successfully addresses the limitations of existing approaches by providing uncertainty-aware counterfactual explanations with formal guarantees, making explanations more reliable and actionable for users.

Abstract: Counterfactual explanations (CFXs) provide human-understandable justifications for model predictions, enabling actionable recourse and enhancing interpretability. To be reliable, CFXs must avoid regions of high predictive uncertainty, where explanations may be misleading or inapplicable. However, existing methods often neglect uncertainty or lack principled mechanisms for incorporating it with formal guarantees. We propose CONFEX, a novel method for generating uncertainty-aware counterfactual explanations using Conformal Prediction (CP) and Mixed-Integer Linear Programming (MILP). CONFEX explanations are designed to provide local coverage guarantees, addressing the issue that CFX generation violates exchangeability. To do so, we develop a novel localised CP procedure that enjoys an efficient MILP encoding by leveraging an offline tree-based partitioning of the input space. This way, CONFEX generates CFXs with rigorous guarantees on both predictive uncertainty and optimality. We evaluate CONFEX against state-of-the-art methods across diverse benchmarks and metrics, demonstrating that our uncertainty-aware approach yields robust and plausible explanations.

Method: C2C introduces two key innovations: (1) Alignment Factor (AF) metric to quantify agent task alignment, and (2) Sequential Action Framework that integrates stepwise execution with intelligent, cost-aware communication decisions.

Result: C2C reduces task completion time by about 40% with acceptable communication costs, successfully completes all tasks in standard configurations, and maintains effectiveness at scale across team sizes from 5 to 17 agents.

Conclusion: C2C establishes both a theoretical foundation for measuring communication effectiveness in multi-agent systems and a practical framework for complex collaborative tasks.

Abstract: Teamwork in workspace for complex tasks requires diverse communication strategies, but current multi-agent LLM systems lack systematic frameworks for task oriented communication. We introduce Communication to Completion (C2C), a scalable framework that addresses this gap through two key innovations: (1) the Alignment Factor (AF), a novel metric quantifying agent task alignment that directly impacts work efficiency, and (2) a Sequential Action Framework that integrates stepwise execution with intelligent communication decisions. C2C enables agents to make cost aware communication choices, dynamically improving task understanding through targeted interactions. We evaluated C2C on realistic coding workflows across three complexity tiers and team sizes from 5 to 17 agents, comparing against no communication and fixed steps baselines. The results show that C2C reduces the task completion time by about 40% with acceptable communication costs. The framework completes all tasks successfully in standard configurations and maintains effectiveness at scale. C2C establishes both a theoretical foundation for measuring communication effectiveness in multi-agent systems and a practical framework for complex collaborative tasks.

Method: Proposed Continued Fraction Q-Learning (QCoFr) framework that captures arbitrary-order agent interactions with linear complexity O(n), and uses variational information bottleneck to extract latent information for credit estimation and filtering noisy interactions.

Result: Extensive experiments show QCoFr consistently achieves better performance and provides interpretability that aligns with theoretical analysis.

Conclusion: QCoFr successfully addresses the limitations of previous MARL methods by providing an efficient, flexible framework for modeling agent interactions while maintaining interpretability and improving cooperation.

Abstract: The ability to model interactions among agents is crucial for effective coordination and understanding their cooperation mechanisms in multi-agent reinforcement learning (MARL). However, previous efforts to model high-order interactions have been primarily hindered by the combinatorial explosion or the opaque nature of their black-box network structures. In this paper, we propose a novel value decomposition framework, called Continued Fraction Q-Learning (QCoFr), which can flexibly capture arbitrary-order agent interactions with only linear complexity $\mathcal{O}\left({n}\right)$ in the number of agents, thus avoiding the combinatorial explosion when modeling rich cooperation. Furthermore, we introduce the variational information bottleneck to extract latent information for estimating credits. This latent information helps agents filter out noisy interactions, thereby significantly enhancing both cooperation and interpretability. Extensive experiments demonstrate that QCoFr not only consistently achieves better performance but also provides interpretability that aligns with our theoretical analysis.

Method: Holon formation is studied using a multiagent system model, analyzing how resource constraints (energy, messages, time) lead to the emergence of holonic structures from fully intercommunicating peers.

Result: Holonic structures demonstrate adaptability to environmental changes, autonomy, and cooperation capabilities, making them useful when resource shortages prevent communications or system components fail.

Conclusion: Holonic fusion represents a significant advancement over traditional hierarchical approaches, providing flexible, adaptive structures that optimize information fusion under resource constraints and uncertainty.

Abstract: There has recently been a major advance with respect to how information fusion is performed. Information fusion has gone from being conceived as a purely hierarchical procedure, as is the case of traditional military applications, to now being regarded collaboratively, as holonic fusion, which is better suited for civil applications and edge organizations. The above paradigm shift is being boosted as information fusion gains ground in different non-military areas, and human-computer and machine-machine communications, where holarchies, which are more flexible structures than ordinary, static hierarchies, become more widespread. This paper focuses on showing how holonic structures tend to be generated when there are constraints on resources (energy, available messages, time, etc.) for interactions based on a set of fully intercommunicating elements (peers) whose components fuse information as a means of optimizing the impact of vagueness and uncertainty present message exchanges. Holon formation is studied generically based on a multiagent system model, and an example of its possible operation is shown. Holonic structures have a series of advantages, such as adaptability, to sudden changes in the environment or its composition, are somewhat autonomous and are capable of cooperating in order to achieve a common goal. This can be useful when the shortage of resources prevents communications or when the system components start to fail.

Method: Used LLM agents in Diner’s Dilemma simulations with natural language reasoning and pairwise imitation for strategy adoption, comparing to Boyd and Richerson’s model.

Result: Agents followed Boyd and Richerson’s strategies, explicit punishment drove norm emergence and reinforced cooperation across different agent configurations.

Conclusion: LLM-based Multi-Agent Systems can replicate traditional mathematical models of cooperation evolution while providing more realistic testbeds through natural language reasoning.

Abstract: The evolution of cooperation has been extensively studied using abstract mathematical models and simulations. Recent advances in Large Language Models (LLMs) and the rise of LLM agents have demonstrated their ability to perform social reasoning, thus providing an opportunity to test the emergence of norms in more realistic agent-based simulations with human-like reasoning using natural language. In this research, we investigate whether the cooperation dynamics presented in Boyd and Richerson’s model persist in a more realistic simulation of the Diner’s Dilemma using LLM agents compared to the abstract mathematical nature in the work of Boyd and Richerson. Our findings indicate that agents follow the strategies defined in the Boyd and Richerson model, and explicit punishment mechanisms drive norm emergence, reinforcing cooperative behaviour even when the agent strategy configuration varies. Our results suggest that LLM-based Multi-Agent System simulations, in fact, can replicate the evolution of cooperation predicted by the traditional mathematical models. Moreover, our simulations extend beyond the mathematical models by integrating natural language-driven reasoning and a pairwise imitation method for strategy adoption, making them a more realistic testbed for cooperative behaviour in MASs.

Method: The paper introduces a developmental continuum framework with six levels to map different agentic architectures and their applications in social science research.

Result: The framework clarifies technical boundaries between agent architectures, showing how lower-tier systems streamline conventional tasks while higher-tier systems enable novel forms of social inquiry including group dynamics and norm formation studies.

Conclusion: While LLM-based agents have transformative potential for social sciences, realizing this requires careful deployment, robust validation protocols, interdisciplinary collaboration, and balancing technical innovation with ethical responsibility.

Abstract: As large language models (LLMs) transition from static tools to fully agentic systems, their potential for transforming social science research has become increasingly evident. This paper introduces a structured framework for understanding the diverse applications of LLM-based agents, ranging from simple data processors to complex, multi-agent systems capable of simulating emergent social dynamics. By mapping this developmental continuum across six levels, the paper clarifies the technical and methodological boundaries between different agentic architectures, providing a comprehensive overview of current capabilities and future potential. It highlights how lower-tier systems streamline conventional tasks like text classification and data annotation, while higher-tier systems enable novel forms of inquiry, including the study of group dynamics, norm formation, and large-scale social processes. However, these advancements also introduce significant challenges, including issues of reproducibility, ethical oversight, and the risk of emergent biases. The paper critically examines these concerns, emphasizing the need for robust validation protocols, interdisciplinary collaboration, and standardized evaluation metrics. It argues that while LLM-based agents hold transformative potential for the social sciences, realizing this promise will require careful, context-sensitive deployment and ongoing methodological refinement. The paper concludes with a call for future research that balances technical innovation with ethical responsibility, encouraging the development of agentic systems that not only replicate but also extend the frontiers of social science, offering new insights into the complexities of human behavior.

Method: Uses multiple AUVs equipped with optical and acoustic multimodal communication as relay nodes, controlled by multi-agent reinforcement learning for cooperative movement. USVs serve as surface relay nodes to coordinate and forward information from AUVs.

Result: Through simulation verification, the proposed scheme effectively achieves reliable and high-speed communication for divers.

Conclusion: The proposed maritime unmanned system approach successfully addresses underwater diver communication challenges by providing adaptive, reliable, and high-speed communication services.

Abstract: As underwater human activities are increasing, the demand for underwater communication service presents a significant challenge. Existing underwater diver communication methods face hurdles due to inherent disadvantages and complex underwater environments. To address this issue, we propose a scheme that utilizes maritime unmanned systems to assist divers with reliable and high-speed communication. Multiple AUVs are equipped with optical and acoustic multimodal communication devices as relay nodes, providing adaptive communication services based on changes in the diver’s activity area. By using a multi-agent reinforcement learning (MARL) approach to control the cooperative movement of AUVs, high-speed and reliable data transmission between divers can be achieved. At the same time, utilizing the advantages of on-demand deployment and wide coverage of unmanned surface vehicles (USVs) as surface relay nodes to coordinate and forward information from AUVs, and controlling AUVs to adaptively select relay USV nodes for data transmission, high-quality communication between divers and surface platform can be achieved. Through simulation verification, the proposed scheme can effectively achieve reliable and high-speed communication for divers.

Method: Conducted comprehensive experiments across 4 real-world environments, 13 uncertainty types, and 15 hyperparameters using over 82,620 experiments to evaluate cooperation, robustness, and resilience in MARL.

Result: Key findings: (1) Cooperation-robustness link weakens with intense perturbations; (2) No generalization across uncertainty modalities; (3) Hyperparameter tuning is critical - standard practices can hurt robustness while specific techniques consistently help.

Conclusion: Hyperparameter optimization alone can substantially improve cooperation, robustness and resilience across all MARL backbones, with findings generalizing to robust MARL methods, highlighting the importance of systematic tuning for trustworthy MARL systems.

Abstract: In cooperative Multi-Agent Reinforcement Learning (MARL), it is a common practice to tune hyperparameters in ideal simulated environments to maximize cooperative performance. However, policies tuned for cooperation often fail to maintain robustness and resilience under real-world uncertainties. Building trustworthy MARL systems requires a deep understanding of robustness, which ensures stability under uncertainties, and resilience, the ability to recover from disruptions–a concept extensively studied in control systems but largely overlooked in MARL. In this paper, we present a large-scale empirical study comprising over 82,620 experiments to evaluate cooperation, robustness, and resilience in MARL across 4 real-world environments, 13 uncertainty types, and 15 hyperparameters. Our key findings are: (1) Under mild uncertainty, optimizing cooperation improves robustness and resilience, but this link weakens as perturbations intensify. Robustness and resilience also varies by algorithm and uncertainty type. (2) Robustness and resilience do not generalize across uncertainty modalities or agent scopes: policies robust to action noise for all agents may fail under observation noise on a single agent. (3) Hyperparameter tuning is critical for trustworthy MARL: surprisingly, standard practices like parameter sharing, GAE, and PopArt can hurt robustness, while early stopping, high critic learning rates, and Leaky ReLU consistently help. By optimizing hyperparameters only, we observe substantial improvement in cooperation, robustness and resilience across all MARL backbones, with the phenomenon also generalizing to robust MARL methods across these backbones. Code and results available at https://github.com/BUAA-TrustworthyMARL/adv_marl_benchmark .

Method: Using local guidance with spatiotemporal cues in the vicinity of each agent, applied to the LaCAM configuration-based solver.

Result: Significantly improved solution quality without exceeding moderate time budget, establishing new performance frontiers for MAPF.

Conclusion: Local guidance with informative spatiotemporal cues can effectively enhance multi-agent pathfinding performance while maintaining computational efficiency.

Abstract: Guidance is an emerging concept that improves the empirical performance of real-time, sub-optimal multi-agent pathfinding (MAPF) methods. It offers additional information to MAPF algorithms to mitigate congestion on a global scale by considering the collective behavior of all agents across the entire workspace. This global perspective helps reduce agents’ waiting times, thereby improving overall coordination efficiency. In contrast, this study explores an alternative approach: providing local guidance in the vicinity of each agent. While such localized methods involve recomputation as agents move and may appear computationally demanding, we empirically demonstrate that supplying informative spatiotemporal cues to the planner can significantly improve solution quality without exceeding a moderate time budget. When applied to LaCAM, a leading configuration-based solver, this form of guidance establishes a new performance frontier for MAPF.

Method: Proposed a DNN architecture integrating feature-wise linear modulation (FiLM) that allows user-defined patterns as conditioning inputs, with progressive training strategies to enhance pattern approximation.

Result: UNDF generalizes to unseen user-defined patterns with higher directivities, scaling variations, and different steering directions, and outperforms conventional methods.

Conclusion: The FiLM-based UNDF architecture successfully enables flexible spatial filtering with user-defined directivity patterns during inference, demonstrating superior performance over traditional approaches.

Abstract: Spatial filtering with a desired directivity pattern is advantageous for many audio applications. In this work, we propose neural directional filtering with user-defined directivity patterns (UNDF), which enables spatial filtering based on directivity patterns that users can define during inference. To achieve this, we propose a DNN architecture that integrates feature-wise linear modulation (FiLM), allowing user-defined patterns to serve as conditioning inputs. Through analysis, we demonstrate that the FiLM-based architecture enables the UNDF to generalize to unseen user-defined patterns during interference with higher directivities, scaling variations, and different steering directions. Furthermore, we progressively refine training strategies to enhance pattern approximation and enable UNDF to approximate irregular shapes. Lastly, experimental comparisons show that UNDF outperforms conventional methods.

Method: SFM constructs intermediate states along flow matching paths from coarse representations, uses orthogonal projection to determine temporal positions of these states, applies single-segment piecewise flow construction, and integrates with TTS models via lightweight SFM head.

Result: SFM yields consistent gains in speech naturalness across objective and subjective evaluations, and significantly accelerates inference when using adaptive-step ODE solvers.

Conclusion: Shallow Flow Matching provides an effective coarse-to-fine generation paradigm for TTS that improves quality while reducing computational cost during inference.

Abstract: We propose Shallow Flow Matching (SFM), a novel mechanism that enhances flow matching (FM)-based text-to-speech (TTS) models within a coarse-to-fine generation paradigm. Unlike conventional FM modules, which use the coarse representations from the weak generator as conditions, SFM constructs intermediate states along the FM paths from these representations. During training, we introduce an orthogonal projection method to adaptively determine the temporal position of these states, and apply a principled construction strategy based on a single-segment piecewise flow. The SFM inference starts from the intermediate state rather than pure noise, thereby focusing computation on the latter stages of the FM paths. We integrate SFM into multiple TTS models with a lightweight SFM head. Experiments demonstrate that SFM yields consistent gains in speech naturalness across both objective and subjective evaluations, and significantly accelerates inference when using adaptive-step ODE solvers. Demo and codes are available at https://ydqmkkx.github.io/SFMDemo/.

Method: Spatially adjusting Quantization Parameters of block-based video codecs adaptively using eye tracking data, transmitting foveal region with high fidelity while reducing quality in peripheral regions.

Result: Bandwidth can be reduced to one third without sacrificing immersion, with both qualitative and quantitative analysis showing maintained transmission fidelity.

Conclusion: The proposed foveated compression method effectively reduces bandwidth requirements for immersive televisualization when eye tracking data is available, enabling high-quality transmission under constrained conditions.

Abstract: Immersive televisualization is important both for telepresence and teleoperation, but resolution and fidelity are often limited by communication bandwidth constraints. We propose a lightweight method for foveated compression of immersive televisualization video streams that can be easily integrated with common video codecs, reducing the required bandwidth if eye tracking data is available. Specifically, we show how to spatially adjust the Quantization Parameter of modern block-based video codecs in a adaptive way based on eye tracking information. The foveal region is transmitted with high fidelity while quality is reduced in the peripheral region, saving bandwidth. We integrate our method with the NimbRo avatar system, which won the ANA Avatar XPRIZE competition. Our experiments show that bandwidth can be reduced to a third without sacrificing immersion. We analyze transmission fidelity with qualitative examples and report quantitative results.

Method: Multi-resolution analysis using discrete wavelet transform to quantify fine TC structures from satellite imagery, enabling identification of physically meaningful structural features.

Result: The approach identifies structural features that strongly correlate with rapid intensity change in tropical cyclones.

Conclusion: The proposed method provides an interpretable framework for TC structure analysis that can be used with deep learning for improved short-term intensity forecasting guidance.

Abstract: Accurate tropical cyclone (TC) short-term intensity forecasting with a 24-hour lead time is essential for disaster mitigation in the Atlantic TC basin. Since most TCs evolve far from land-based observing networks, satellite imagery is critical to monitoring these storms; however, these complex and high-resolution spatial structures can be challenging to qualitatively interpret in real time by forecasters. Here we propose a concise, interpretable, and descriptive approach to quantify fine TC structures with a multi-resolution analysis (MRA) by the discrete wavelet transform, enabling data analysts to identify physically meaningful structural features that strongly correlate with rapid intensity change. Furthermore, deep-learning techniques can build on this MRA for short-term intensity guidance.

Method: Analyzed both dilation and eyelid occlusion using a large dataset of 555 distinct irises, examining correlation between probe image visible iris area and Hamming distance of iris code pairs.

Result: Demonstrated strong correlation between visible iris area and Hamming distance, showing that visible iris area is a robust indicator of probe image quality.

Conclusion: Visible iris area could be efficiently incorporated into the iris acquisition process to improve confidence in match predictions.

Abstract: With the increasing adoption of iris recognition systems and the expansion of large-scale enrollment databases, there is a growing need to efficiently assess iris image quality at the time of acquisition, particularly to model user non-compliance in real time. Image quality may degrade due to eyelid occlusion or pupil dilation. Although previous studies have shown that occlusion and changes in the pupil-to-iris ratio negatively impact recognition performance, these investigations were typically limited by small sample sizes and did not examine the combined effects of eyelid and pupil variations. In this study, we analyze both dilation and eyelid occlusion using a large dataset of 555 distinct irises and demonstrate a strong correlation between probe image visible iris area and the Hamming distance of iris code pairs. These results suggest that visible iris area is a robust indicator of probe image quality and could be efficiently incorporated into the iris acquisition process to improve confidence in match predictions.

Method: The system generates 3D assets from single images with accurate geometry, well-aligned textures, and realistic physically-based materials that can be directly integrated into physics engines with minimal configuration.

Result: Seed3D 1.0 produces assets that enable deployment in robotic manipulation and simulation training, and can scale to complete scene generation by assembling objects into coherent environments.

Conclusion: By enabling scalable simulation-ready content creation, Seed3D 1.0 provides a foundation for advancing physics-based world simulators and embodied AI agent training.

Abstract: Developing embodied AI agents requires scalable training environments that balance content diversity with physics accuracy. World simulators provide such environments but face distinct limitations: video-based methods generate diverse content but lack real-time physics feedback for interactive learning, while physics-based engines provide accurate dynamics but face scalability limitations from costly manual asset creation. We present Seed3D 1.0, a foundation model that generates simulation-ready 3D assets from single images, addressing the scalability challenge while maintaining physics rigor. Unlike existing 3D generation models, our system produces assets with accurate geometry, well-aligned textures, and realistic physically-based materials. These assets can be directly integrated into physics engines with minimal configuration, enabling deployment in robotic manipulation and simulation training. Beyond individual objects, the system scales to complete scene generation through assembling objects into coherent environments. By enabling scalable simulation-ready content creation, Seed3D 1.0 provides a foundation for advancing physics-based world simulators. Seed3D 1.0 is now available on https://console.volcengine.com/ark/region:ark+cn-beijing/experience/vision?modelId=doubao-seed3d-1-0-250928&tab=Gen3D

Method: Uses a modified Dark Channel Prior for initial dehazing, followed by a U-shaped green learning pipeline with unsupervised representation learning, feature engineering (RFT and LNT), and transparent supervised learning.

Result: Significantly reduces parameter count while achieving performance comparable to state-of-the-art deep learning models, with mathematical interpretability.

Conclusion: GUSL-Dehaze provides an effective, lightweight alternative to deep learning-based dehazing methods that is suitable for resource-constrained devices while maintaining transparency and interpretability.

Abstract: Image dehazing is a restoration task that aims to recover a clear image from a single hazy input. Traditional approaches rely on statistical priors and the physics-based atmospheric scattering model to reconstruct the haze-free image. While recent state-of-the-art methods are predominantly based on deep learning architectures, these models often involve high computational costs and large parameter sizes, making them unsuitable for resource-constrained devices. In this work, we propose GUSL-Dehaze, a Green U-Shaped Learning approach to image dehazing. Our method integrates a physics-based model with a green learning (GL) framework, offering a lightweight, transparent alternative to conventional deep learning techniques. Unlike neural network-based solutions, GUSL-Dehaze completely avoids deep learning. Instead, we begin with an initial dehazing step using a modified Dark Channel Prior (DCP), which is followed by a green learning pipeline implemented through a U-shaped architecture. This architecture employs unsupervised representation learning for effective feature extraction, together with feature-engineering techniques such as the Relevant Feature Test (RFT) and the Least-Squares Normal Transform (LNT) to maintain a compact model size. Finally, the dehazed image is obtained via a transparent supervised learning strategy. GUSL-Dehaze significantly reduces parameter count while ensuring mathematical interpretability and achieving performance on par with state-of-the-art deep learning models.

Method: Integrates ghost modules, efficient channel attention, and squeeze-and-excitation blocks for robust feature representation, with quantization-aware head and spike-compatible transformations for seamless SNN conversion.

Result: Achieves 91.6% Top-1 accuracy and 82.4% macro F1 on HAM10000, 90.8%/81.7% on clinical dataset; deployed on BrainChip Akida with 1.5 ms latency and 1.7 mJ energy per image, reducing latency and energy by over 94.6%/98.6% vs GPU-based CNNs.

Conclusion: QANA effectively enables accurate, real-time, and privacy-sensitive medical analysis in edge environments, significantly outperforming state-of-the-art CNN-to-SNN models.

Abstract: Accurate and efficient skin lesion classification on edge devices is critical for accessible dermatological care but remains challenging due to computational, energy, and privacy constraints. We introduce QANA, a novel quantization-aware neuromorphic architecture for incremental skin lesion classification on resource-limited hardware. QANA effectively integrates ghost modules, efficient channel attention, and squeeze-and-excitation blocks for robust feature representation with low-latency and energy-efficient inference. Its quantization-aware head and spike-compatible transformations enable seamless conversion to spiking neural networks (SNNs) and deployment on neuromorphic platforms. Evaluation on the large-scale HAM10000 benchmark and a real-world clinical dataset shows that QANA achieves 91.6% Top-1 accuracy and 82.4% macro F1 on HAM10000, and 90.8%/81.7% on the clinical dataset, significantly outperforming state-of-the-art CNN-to-SNN models under fair comparison. Deployed on BrainChip Akida hardware, QANA achieves 1.5 ms inference latency and 1.7,mJ energy per image, reducing inference latency and energy use by over 94.6%/98.6% compared to GPU-based CNNs surpassing state-of-the-art CNN-to-SNN conversion baselines. These results demonstrate the effectiveness of QANA for accurate, real-time, and privacy-sensitive medical analysis in edge environments.