Method: Uses comprehensive preprocessing pipeline with facial landmark detection, lip region isolation/stabilization, and multimodal quality assessment. Implements curriculum learning guided by quality scores, and adopts SigLIP encoder with Unified 3D-Resampler for high-frame-rate lip motion encoding.

Result: Achieves state-of-the-art performance in both literal accuracy and semantic fidelity on the HI-Dialogue dataset.

Conclusion: Establishes foundation for applying Omni-Models to assistive communication technology with end-to-end modeling framework and essential processing tools for future research.

Abstract: To provide a unified and flexible solution for daily communication among hearing-impaired individuals, we introduce the Omni-Model paradigm into assistive technology and present HI-TransPA, an instruction-driven audio-visual personal assistant. The model fuses indistinct speech with high-frame-rate lip dynamics, enabling both translation and dialogue within a single multimodal framework. To tackle the challenges of noisy and heterogeneous raw data and the limited adaptability of existing Omni-Models to hearing-impaired speech, we construct a comprehensive preprocessing and curation pipeline that detects facial landmarks, isolates and stabilizes the lip region, and quantitatively assesses multimodal sample quality. These quality scores guide a curriculum learning strategy that first trains on clean, high-confidence samples and progressively incorporates harder cases to strengthen model robustness. We further adopt a SigLIP encoder combined with a Unified 3D-Resampler to efficiently encode high-frame-rate lip motion. Experiments on our purpose-built HI-Dialogue dataset show that HI-TransPA achieves state-of-the-art performance in both literal accuracy and semantic fidelity. This work establishes a foundation for applying Omni-Models to assistive communication technology, providing an end-to-end modeling framework and essential processing tools for future research.

Method: Uses 7B parameter self-supervised pre-training for robust speech representations, encoder-decoder architecture with LLM-inspired decoder for zero-shot generalization, trained on massive diverse corpus combining public resources and community-sourced recordings.

Result: Achieves largest ASR coverage to date (1,600+ languages), substantial gains over prior systems especially in low-resource conditions, strong generalization, and releases models from 300M to 7B parameters.

Conclusion: Open-sourcing models and tools lowers barriers for researchers and communities, enabling new participation forms while addressing ethical considerations through compensated local partnerships.

Abstract: Automatic speech recognition (ASR) has advanced in high-resource languages, but most of the world’s 7,000+ languages remain unsupported, leaving thousands of long-tail languages behind. Expanding ASR coverage has been costly and limited by architectures that restrict language support, making extension inaccessible to most–all while entangled with ethical concerns when pursued without community collaboration. To transcend these limitations, we introduce Omnilingual ASR, the first large-scale ASR system designed for extensibility. Omnilingual ASR enables communities to introduce unserved languages with only a handful of data samples. It scales self-supervised pre-training to 7B parameters to learn robust speech representations and introduces an encoder-decoder architecture designed for zero-shot generalization, leveraging a LLM-inspired decoder. This capability is grounded in a massive and diverse training corpus; by combining breadth of coverage with linguistic variety, the model learns representations robust enough to adapt to unseen languages. Incorporating public resources with community-sourced recordings gathered through compensated local partnerships, Omnilingual ASR expands coverage to over 1,600 languages, the largest such effort to date–including over 500 never before served by ASR. Automatic evaluations show substantial gains over prior systems, especially in low-resource conditions, and strong generalization. We release Omnilingual ASR as a family of models, from 300M variants for low-power devices to 7B for maximum accuracy. We reflect on the ethical considerations shaping this design and conclude by discussing its societal impact. In particular, we highlight how open-sourcing models and tools can lower barriers for researchers and communities, inviting new forms of participation. Open-source artifacts are available at https://github.com/facebookresearch/omnilingual-asr.

Method: Conducted controlled experiments on classification and generation tasks using multiple model families (0.5B to 27B parameters) and GPT-5, comparing performance variance from different example orderings versus different example sets.

Result: Found that performance variance due to different example orderings is comparable to variance from using entirely different example sets. Strong orderings can be identified using development sets, achieving near-oracle performance.

Conclusion: Example selection and ordering are equally important and intertwined in prompt design, requiring reexamination of ICL assumptions.

Abstract: In-context learning (ICL) enables large language models to perform new tasks by conditioning on a sequence of examples. Most prior work reasonably and intuitively assumes that which examples are chosen has a far greater effect on performance than how those examples are ordered, leading to a focus on example selection. We revisit this assumption and conduct a systematic comparison between the effect of selection and ordering. Through controlled experiments on both classification and generation tasks, using multiple open-source model families (0.5B to 27B parameters) and GPT-5, we find that the variance in performance due to different example orderings is comparable to that from using entirely different example sets. Furthermore, we show that strong orderings can be identified using only a development set, achieving performance close to an oracle that selects the best ordering based on test labels. Our findings highlight the equal and intertwined importance of example selection and ordering in prompt design, calling for a reexamination of the assumptions held in ICL.

Method: Investigated morphologically-aware tokenization for Latin by incorporating linguistic resources and morphological guidance into the tokenization process.

Result: Morphologically-guided tokenization improved overall performance on four downstream tasks, with most significant gains for out-of-domain texts, indicating better generalization ability.

Conclusion: Linguistic resources can effectively improve language modeling for morphologically complex languages, serving as a feasible alternative for low-resource languages lacking large-scale pretraining data.

Abstract: Tokenization is a critical component of language model pretraining, yet standard tokenization methods often prioritize information-theoretical goals like high compression and low fertility rather than linguistic goals like morphological alignment. In fact, they have been shown to be suboptimal for morphologically rich languages, where tokenization quality directly impacts downstream performance. In this work, we investigate morphologically-aware tokenization for Latin, a morphologically rich language that is medium-resource in terms of pretraining data, but high-resource in terms of curated lexical resources – a distinction that is often overlooked but critical in discussions of low-resource language modeling. We find that morphologically-guided tokenization improves overall performance on four downstream tasks. Performance gains are most pronounced for out of domain texts, highlighting our models’ improved generalization ability. Our findings demonstrate the utility of linguistic resources to improve language modeling for morphologically complex languages. For low-resource languages that lack large-scale pretraining data, the development and incorporation of linguistic resources can serve as a feasible alternative to improve LM performance.

Method: Used Natural Language Processing to analyze Presidential Directives corpus, comparing NLP results with human analyst identifications of relevant documents and themes.

Result: Both NLP and human analysts identified relevant documents, demonstrating NLP’s potential utility. However, discrepancies between NLP and human-labeled results were found.

Conclusion: NLP shows promise for large corpus analysis in social science, but validity concerns remain due to discrepancies with human analysis, and rapid AI/ML evolution means current tools may already be outdated.

Abstract: Our research investigates how Natural Language Processing (NLP) can be used to extract main topics from a larger corpus of written data, as applied to the case of identifying signaling themes in Presidential Directives (PDs) from the Reagan through Clinton administrations. Analysts and NLP both identified relevant documents, demonstrating the potential utility of NLPs in research involving large written corpuses. However, we also identified discrepancies between NLP and human-labeled results that indicate a need for more research to assess the validity of NLP in this use case. The research was conducted in 2023, and the rapidly evolving landscape of AIML means existing tools have improved and new tools have been developed; this research displays the inherent capabilities of a potentially dated AI tool in emerging social science applications.

Method: Benchmark sub-2B models on English->German Critical Error Detection using standardized prompts, logit-bias calibration, and majority voting across WMT21, WMT22, and SynCED-EnDe-2025 datasets.

Result: Gemma-3-1B achieves MCC=0.77 with F1-ERR=0.98 on SynCED-EnDe-2025 after fine-tuning, with 400ms latency on MacBook Pro M4 Pro. Qwen-3-1.7B attains highest MCC but with higher compute cost.

Conclusion: Compact instruction-tuned LLMs with lightweight calibration can deliver trustworthy on-device critical error detection for machine translation, enabling private low-cost error screening.

Abstract: Large Language Models (LLMs) excel at evaluating machine translation (MT), but their scale and cost hinder deployment on edge devices and in privacy-sensitive workflows. We ask: how small can you get while still detecting meaning-altering translation errors? Focusing on English->German Critical Error Detection (CED), we benchmark sub-2B models (LFM2-350M, Qwen-3-0.6B/1.7B, Llama-3.2-1B-Instruct, Gemma-3-1B) across WMT21, WMT22, and SynCED-EnDe-2025. Our framework standardizes prompts, applies lightweight logit-bias calibration and majority voting, and reports both semantic quality (MCC, F1-ERR/F1-NOT) and compute metrics (VRAM, latency, throughput). Results reveal a clear sweet spot around one billion parameters: Gemma-3-1B provides the best quality-efficiency trade-off, reaching MCC=0.77 with F1-ERR=0.98 on SynCED-EnDe-2025 after merged-weights fine-tuning, while maintaining 400 ms single-sample latency on a MacBook Pro M4 Pro (24 GB). At larger scale, Qwen-3-1.7B attains the highest absolute MCC (+0.11 over Gemma) but with higher compute cost. In contrast, ultra-small models (0.6B) remain usable with few-shot calibration yet under-detect entity and number errors. Overall, compact, instruction-tuned LLMs augmented with lightweight calibration and small-sample supervision can deliver trustworthy, on-device CED for MT, enabling private, low-cost error screening in real-world translation pipelines. All datasets, prompts, and scripts are publicly available at our GitHub repository.

Method: The benchmark segments continuous full-duplex dialogues into discrete turns to enable turn-by-turn evaluation across dialogue quality, conversational dynamics, instruction following, and safety.

Result: Experimental results show current FD-SLMs struggle to maintain consistent performance across multiple rounds and evaluation dimensions.

Conclusion: The benchmark effectively addresses evaluation gaps for FD-SLMs and demonstrates the need for comprehensive multi-round assessment.

Abstract: Full-Duplex Speech Language Models (FD-SLMs) enable real-time, overlapping conversational interactions, offering a more dynamic user experience compared to traditional half-duplex models. However, existing benchmarks primarily focus on evaluating single-round interactions and conversational features, neglecting the complexities of multi-round communication and critical capabilities such as instruction following and safety. Evaluating FD-SLMs in multi-round settings poses significant challenges, including blurred turn boundaries in communication and context inconsistency during model inference. To address these gaps, we introduce MTR-DuplexBench, a novel benchmark that segments continuous full-duplex dialogues into discrete turns, enabling comprehensive, turn-by-turn evaluation of FD-SLMs across dialogue quality, conversational dynamics, instruction following, and safety. Experimental results reveal that current FD-SLMs face difficulties in maintaining consistent performance across multiple rounds and evaluation dimensions, highlighting the necessity and effectiveness of our proposed benchmark. The benchmark and code will be available in the future.

Method: Conducted analysis of predicate-argument structures using annotation projection experiments where each language served as source to project annotations to parallel sentences, supported by qualitative and quantitative analysis.

Result: Language transfer is asymmetric - performance differs significantly depending on which language is used as source, with clear structural divergences identified between Chinese and English predicate-argument structures.

Conclusion: Source language selection is crucial in cross-lingual transfer learning and requires investigation before making scientific claims about cross-lingual NLP systems.

Abstract: Cross-lingual Natural Language Processing (NLP) has gained significant traction in recent years, offering practical solutions in low-resource settings by transferring linguistic knowledge from resource-rich to low-resource languages. This field leverages techniques like annotation projection and model transfer for language adaptation, supported by multilingual pre-trained language models. However, linguistic divergences hinder language transfer, especially among typologically distant languages. In this paper, we present an analysis of predicate-argument structures in parallel Chinese and English sentences. We explore the alignment and misalignment of predicate annotations, inspecting similarities and differences and proposing a categorization of structural divergences. The analysis and the categorization are supported by a qualitative and quantitative analysis of the results of an annotation projection experiment, in which, in turn, one of the two languages has been used as source language to project annotations into the corresponding parallel sentences. The results of this analysis show clearly that language transfer is asymmetric. An aspect that requires attention when it comes to selecting the source language in transfer learning applications and that needs to be investigated before any scientific claim about cross-lingual NLP is proposed.

Method: Uses single-shot policy with short, no-context drafts from base model. Computes uncertainty scores from prefix logits: mean token entropy, margin signal (top-1/top-2 logit gap), or small-N variance across stochastic prefixes. Retrieves only when score exceeds threshold.

Result: On NQ-Open, TriviaQA, and PopQA, matches or improves EM/F1 vs Always-RAG while reducing retrieval by 70-90% and cutting end-to-end latency. Close to Never-RAG in overhead. Margin signal works best with modern instruction-tuned LLMs.

Conclusion: TARG shifts accuracy-efficiency frontier, providing explicit budget trade-offs through delta-latency view. Model-agnostic, no training needed, adds minimal draft tokens.

Abstract: Retrieval-Augmented Generation (RAG) improves factuality but retrieving for every query often hurts quality while inflating tokens and latency. We propose Training-free Adaptive Retrieval Gating (TARG), a single-shot policy that decides when to retrieve using only a short, no-context draft from the base model. From the draft’s prefix logits, TARG computes lightweight uncertainty scores: mean token entropy, a margin signal derived from the top-1/top-2 logit gap via a monotone link, or small-N variance across a handful of stochastic prefixes, and triggers retrieval only when the score exceeds a threshold. The gate is model agnostic, adds only tens to hundreds of draft tokens, and requires no additional training or auxiliary heads. On NQ-Open, TriviaQA, and PopQA, TARG consistently shifts the accuracy-efficiency frontier: compared with Always-RAG, TARG matches or improves EM/F1 while reducing retrieval by 70-90% and cutting end-to-end latency, and it remains close to Never-RAG in overhead. A central empirical finding is that under modern instruction-tuned LLMs the margin signal is a robust default (entropy compresses as backbones sharpen), with small-N variance offering a conservative, budget-first alternative. We provide ablations over gate type and prefix length and use a delta-latency view to make budget trade-offs explicit.

Method: Introduces VocalNet-M2 with multi-codebook tokenizer and multi-token prediction strategy, directly generating multi-codebook speech tokens to eliminate flow-matching models.

Result: Achieves substantial reduction in first chunk latency from approximately 725ms to 350ms while maintaining competitive performance across mainstream SLMs.

Conclusion: Provides valuable insights for developing efficient real-time interactive SLMs through comprehensive comparison of single-codebook and multi-codebook strategies.

Abstract: Current end-to-end spoken language models (SLMs) have made notable progress, yet they still encounter considerable response latency. This delay primarily arises from the autoregressive generation of speech tokens and the reliance on complex flow-matching models for speech synthesis. To overcome this, we introduce VocalNet-M2, a novel low-latency SLM that integrates a multi-codebook tokenizer and a multi-token prediction (MTP) strategy. Our model directly generates multi-codebook speech tokens, thus eliminating the need for a latency-inducing flow-matching model. Furthermore, our MTP strategy enhances generation efficiency and improves overall performance. Extensive experiments demonstrate that VocalNet-M2 achieves a substantial reduction in first chunk latency (from approximately 725ms to 350ms) while maintaining competitive performance across mainstream SLMs. This work also provides a comprehensive comparison of single-codebook and multi-codebook strategies, offering valuable insights for developing efficient and high-performance SLMs for real-time interactive applications.

Method: Integrates Khmer subword segmentation, NER, grapheme-to-phoneme conversion, and a language model to identify and rank correction candidates while handling the specific challenges of Khmer language.

Result: Achieves state-of-the-art Khmer spellchecking accuracy of 94.4%, significantly outperforming existing solutions. Benchmark datasets for Khmer spellchecking and NER will be made publicly available.

Conclusion: The holistic approach effectively addresses Khmer-specific spellchecking challenges and sets a new standard for accuracy, with publicly available datasets to support future research.

Abstract: Compared to English and other high-resource languages, spellchecking for Khmer remains an unresolved problem due to several challenges. First, there are misalignments between words in the lexicon and the word segmentation model. Second, a Khmer word can be written in different forms. Third, Khmer compound words are often loosely and easily formed, and these compound words are not always found in the lexicon. Fourth, some proper nouns may be flagged as misspellings due to the absence of a Khmer named-entity recognition (NER) model. Unfortunately, existing solutions do not adequately address these challenges. This paper proposes a holistic approach to the Khmer spellchecking problem by integrating Khmer subword segmentation, Khmer NER, Khmer grapheme-to-phoneme (G2P) conversion, and a Khmer language model to tackle these challenges, identify potential correction candidates, and rank the most suitable candidate. Experimental results show that the proposed approach achieves a state-of-the-art Khmer spellchecking accuracy of up to 94.4%, compared to existing solutions. The benchmark datasets for Khmer spellchecking and NER tasks in this study will be made publicly available.

Method: Combines data mining, collaborative filtering, and machine learning algorithms with user preferences and academic criteria. Developed through iterative prototyping with user feedback for interface optimization.

Result: A refined and optimized recommendation system that provides customized course suggestions aligned with industry trends and individual career goals.

Conclusion: The system serves as a valuable tool for students, instructors, and career advisors to promote lifelong learning and professional progression by improving course selection decisions.

Abstract: This paper aims to address the challenge of selecting relevant courses for students by proposing the design and development of a course recommendation system. The course recommendation system utilises a combination of data analytics techniques and machine learning algorithms to recommend courses that align with current industry trends and requirements. In order to provide customised suggestions, the study entails the design and implementation of an extensive algorithmic framework that combines machine learning methods, user preferences, and academic criteria. The system employs data mining and collaborative filtering techniques to examine past courses and individual career goals in order to provide course recommendations. Moreover, to improve the accessibility and usefulness of the recommendation system, special attention is given to the development of an easy-to-use front-end interface. The front-end design prioritises visual clarity, interaction, and simplicity through iterative prototyping and user input revisions, guaranteeing a smooth and captivating user experience. We refined and optimised the proposed system by incorporating user feedback, ensuring that it effectively meets the needs and preferences of its target users. The proposed course recommendation system could be a useful tool for students, instructors, and career advisers to use in promoting lifelong learning and professional progression as it fills the gap between university learning and industry expectations. We hope that the proposed course recommendation system will help university students in making data-drive and industry-informed course decisions, in turn, improving graduate outcomes for the university sector.

Method: Fine-tune open-source LLM on course data, use RAG with local knowledge base containing course content, and integrate multi chain-of-thought reasoning to reduce hallucination.

Result: The fine-tuned LLM with RAG method shows strong performance on question answering tasks, tested on HotpotQA dataset.

Conclusion: The proposed system effectively mitigates forum challenges by providing timely, accurate responses to student queries using advanced LLM techniques.

Abstract: The course forums are increasingly significant and play vital role in facilitating student discussions and answering their questions related to the course. It provides a platform for students to post their questions related to the content and admin issues related to the course. However, there are several challenges due to the increase in the number of students enrolled in the course. The primary challenge is that students’ queries cannot be responded immediately and the instructors have to face lots of repetitive questions. To mitigate these issues, we propose a question answering system based on large language model with retrieval augmented generation (RAG) method. This work focuses on designing a question answering system with open source Large Language Model (LLM) and fine-tuning it on the relevant course dataset. To further improve the performance, we use a local knowledge base and applied RAG method to retrieve relevant documents relevant to students’ queries, where the local knowledge base contains all the course content. To mitigate the hallucination of LLMs, We also integrate it with multi chain-of-thought reasoning to overcome the challenge of hallucination in LLMs. In this work, we experiment fine-tuned LLM with RAG method on the HotpotQA dataset. The experimental results demonstrate that the fine-tuned LLM with RAG method has a strong performance on question answering task.

Method: Propose TermGPT with multi-level contrastive fine-tuning: construct sentence graphs to capture semantic relations, generate discriminative positive/negative samples using contextual and topological cues, and apply contrastive learning at both sentence and token levels for global context and fine-grained terminology discrimination.

Result: TermGPT outperforms existing baselines in term discrimination tasks within finance and legal domains, with evaluation supported by the first financial terminology dataset from official regulatory documents.

Conclusion: The proposed multi-level contrastive fine-tuning framework effectively addresses terminology-level representation weaknesses in LLMs, demonstrating superior performance in domain-specific term discrimination tasks.

Abstract: Large language models (LLMs) have demonstrated impressive performance in text generation tasks; however, their embedding spaces often suffer from the isotropy problem, resulting in poor discrimination of domain-specific terminology, particularly in legal and financial contexts. This weakness in terminology-level representation can severely hinder downstream tasks such as legal judgment prediction or financial risk analysis, where subtle semantic distinctions are critical. To address this problem, we propose TermGPT, a multi-level contrastive fine-tuning framework designed for terminology adaptation. We first construct a sentence graph to capture semantic and structural relations, and generate semantically consistent yet discriminative positive and negative samples based on contextual and topological cues. We then devise a multi-level contrastive learning approach at both the sentence and token levels, enhancing global contextual understanding and fine-grained terminology discrimination. To support robust evaluation, we construct the first financial terminology dataset derived from official regulatory documents. Experiments show that TermGPT outperforms existing baselines in term discrimination tasks within the finance and legal domains.

Method: Curated a seed dataset of diverse realistic scenarios, evaluated model behavior across three dimensions, and developed a data synthesis methodology based on key findings.

Result: Found that strategic planning quality is the primary performance determinant, interaction design and task complexity significantly influence reasoning, and data quality matters more than diversity for optimal performance.

Conclusion: The developed data synthesis methodology demonstrates significant improvements in open-source LLMs’ analytical reasoning capabilities.

Abstract: Large Language Models (LLMs) hold promise in automating data analysis tasks, yet open-source models face significant limitations in these kinds of reasoning-intensive scenarios. In this work, we investigate strategies to enhance the data analysis capabilities of open-source LLMs. By curating a seed dataset of diverse, realistic scenarios, we evaluate model behavior across three core dimensions: data understanding, code generation, and strategic planning. Our analysis reveals three key findings: (1) Strategic planning quality serves as the primary determinant of model performance; (2) Interaction design and task complexity significantly influence reasoning capabilities; (3) Data quality demonstrates a greater impact than diversity in achieving optimal performance. We leverage these insights to develop a data synthesis methodology, demonstrating significant improvements in open-source LLMs’ analytical reasoning capabilities. Code is available at https://github.com/zjunlp/DataMind.

Method: Cascading domain-specific encoders with LLMs, testing across four speech encoders and three datasets covering four tasks, with component-wise analysis.

Result: Attack achieves 90.76% to 99.41% success rates across different configurations, identifying vulnerable pipeline stages.

Conclusion: Audio backdoor attacks pose serious threat to speech language models, but fine-tuning-based defense can mitigate poisoned pretrained encoders.

Abstract: Large Language Models (LLMs) and their multimodal extensions are becoming increasingly popular. One common approach to enable multimodality is to cascade domain-specific encoders with an LLM, making the resulting model inherit vulnerabilities from all of its components. In this work, we present the first systematic study of audio backdoor attacks against speech language models. We demonstrate its effectiveness across four speech encoders and three datasets, covering four tasks: automatic speech recognition (ASR), speech emotion recognition, and gender and age prediction. The attack consistently achieves high success rates, ranging from 90.76% to 99.41%. To better understand how backdoors propagate, we conduct a component-wise analysis to identify the most vulnerable stages of the pipeline. Finally, we propose a fine-tuning-based defense that mitigates the threat of poisoned pretrained encoders.

Method: InTRO uses correction factors - token-wise importance weights estimated from information discrepancy between generative policy and answer-conditioned policy - for informative next token selection, enabling token-level exploration and self-feedback within a single forward pass.

Result: Across six math-reasoning benchmarks, InTRO consistently outperforms baselines, improving solution accuracy by up to 20% relative to base model, with notably more concise rationales and reduced verbosity.

Conclusion: InTRO demonstrates robust generalization by enabling cross-domain transfer to out-of-domain reasoning tasks beyond mathematics, providing an effective solution for training LLMs for chain-of-thought reasoning.

Abstract: Training Large Language Models (LLMs) for chain-of-thought reasoning presents a significant challenge: supervised fine-tuning on a single “golden” rationale hurts generalization as it penalizes equally valid alternatives, whereas reinforcement learning with verifiable rewards struggles with credit assignment and prohibitive computational cost. To tackle these limitations, we introduce InTRO (In-Token Rationality Optimization), a new framework that enables both token-level exploration and self-feedback for accurate and concise reasoning. Instead of directly optimizing an intractable objective over all valid reasoning paths, InTRO leverages correction factors-token-wise importance weights estimated by the information discrepancy between the generative policy and its answer-conditioned counterpart, for informative next token selection. This approach allows the model to perform token-level exploration and receive self-generated feedback within a single forward pass, ultimately encouraging accurate and concise rationales. Across six math-reasoning benchmarks, InTRO consistently outperforms other baselines, raising solution accuracy by up to 20% relative to the base model. Its chains of thought are also notably more concise, exhibiting reduced verbosity. Beyond this, InTRO enables cross-domain transfer, successfully adapting to out-of-domain reasoning tasks that extend beyond the realm of mathematics, demonstrating robust generalization.

Method: Literature review analyzing scholarly research, digital platforms, and commercial tools from past 20 years, proposing a knowledge-based computational framework using rule-based acoustic modeling based on canonical pronunciation principles and articulation points (Makhraj).

Result: Analysis reveals fundamental flaws in current ASR-based approaches, including biased datasets, demographic disparities, and inability to provide meaningful feedback. The proposed paradigm shift emphasizes linguistic expertise combined with audio processing.

Conclusion: Future automated Quranic recitation assessment requires hybrid systems combining linguistic expertise with advanced audio processing to develop reliable, fair, and pedagogically effective tools for global learners.

Abstract: The art and science of Quranic recitation (Tajweed), a discipline governed by meticulous phonetic, rhythmic, and theological principles, confronts substantial educational challenges in today’s digital age. Although modern technology offers unparalleled opportunities for learning, existing automated systems for evaluating recitation have struggled to gain broad acceptance or demonstrate educational effectiveness. This literature review examines this crucial disparity, offering a thorough analysis of scholarly research, digital platforms, and commercial tools developed over the past twenty years. Our analysis uncovers a fundamental flaw in current approaches that adapt Automatic Speech Recognition (ASR) systems, which emphasize word identification over qualitative acoustic evaluation. These systems suffer from limitations such as reliance on biased datasets, demographic disparities, and an inability to deliver meaningful feedback for improvement. Challenging these data-centric methodologies, we advocate for a paradigm shift toward a knowledge-based computational framework. By leveraging the unchanging nature of the Quranic text and the well-defined rules of Tajweed, we propose that an effective evaluation system should be built upon rule-based acoustic modeling centered on canonical pronunciation principles and articulation points (Makhraj), rather than depending on statistical patterns derived from flawed or biased data. The review concludes that the future of automated Quranic recitation assessment lies in hybrid systems that combine linguistic expertise with advanced audio processing. Such an approach paves the way for developing reliable, fair, and pedagogically effective tools that can authentically assist learners across the globe.

Method: Formulated as a finite-horizon Markov Decision Process (MDP) with a PPO-based reinforcement learning agent that iteratively selects which models to invoke at each stage of multi-hop inference, conditioning on evolving context and accumulated cost.

Result: Experiments with three open-source LLMs across six benchmarks (QA, code generation, mathematical reasoning) show HierRouter improves response quality by up to 2.4x compared to using individual models independently, with only minimal additional inference cost.

Conclusion: Hierarchical routing shows promise for cost-efficient, high-performance LLM inference by dynamically assembling specialized lightweight models.

Abstract: Large Language Models (LLMs) deliver state-of-the-art performance across many tasks but impose high computational and memory costs, limiting their deployment in resource-constrained or real-time settings. To address this, we propose HierRouter, a hierarchical routing approach that dynamically assembles inference pipelines from a pool of specialized, lightweight language models. Formulated as a finite-horizon Markov Decision Process (MDP), our approach trains a Proximal Policy Optimization (PPO)-based reinforcement learning agent to iteratively select which models to invoke at each stage of multi-hop inference. The agent conditions on the evolving context and accumulated cost to make context-aware routing decisions. Experiments with three open-source candidate LLMs across six benchmarks, including QA, code generation, and mathematical reasoning, show that HierRouter improves response quality by up to 2.4x compared to using individual models independently, while incurring only a minimal additional inference cost on average. These results highlight the promise of hierarchical routing for cost-efficient, high-performance LLM inference. All codes can be found here https://github.com/ Nikunj-Gupta/hierouter.

Method: Uses Neural Tangent Kernel-based safety vector distillation to decouple safety constraints from task reasoning, plus interference-aware merging technique to balance safety and utility across different model architectures and sizes.

Result: Achieves significantly lower unsafe rate and higher utility score than baselines, demonstrates robust resistance to jailbreaking attacks, and works across 3 task domains, 3 LLM architectures, and 11 datasets with universal applicability.

Conclusion: EnchTable effectively maintains safety alignment in fine-tuned LLMs without compromising utility, offering seamless integration into deployment pipelines with minimal overhead.

Abstract: Many machine learning models are fine-tuned from large language models (LLMs) to achieve high performance in specialized domains like code generation, biomedical analysis, and mathematical problem solving. However, this fine-tuning process often introduces a critical vulnerability: the systematic degradation of safety alignment, undermining ethical guidelines and increasing the risk of harmful outputs. Addressing this challenge, we introduce EnchTable, a novel framework designed to transfer and maintain safety alignment in downstream LLMs without requiring extensive retraining. EnchTable leverages a Neural Tangent Kernel (NTK)-based safety vector distillation method to decouple safety constraints from task-specific reasoning, ensuring compatibility across diverse model architectures and sizes. Additionally, our interference-aware merging technique effectively balances safety and utility, minimizing performance compromises across various task domains. We implemented a fully functional prototype of EnchTable on three different task domains and three distinct LLM architectures, and evaluated its performance through extensive experiments on eleven diverse datasets, assessing both utility and model safety. Our evaluations include LLMs from different vendors, demonstrating EnchTable’s generalization capability. Furthermore, EnchTable exhibits robust resistance to static and dynamic jailbreaking attacks, outperforming vendor-released safety models in mitigating adversarial prompts. Comparative analyses with six parameter modification methods and two inference-time alignment baselines reveal that EnchTable achieves a significantly lower unsafe rate, higher utility score, and universal applicability across different task domains. Additionally, we validate EnchTable can be seamlessly integrated into various deployment pipelines without significant overhead.

Method: Norm-RAG models utterance-level attributes (communicative intent, speaker roles, interpersonal framing, linguistic cues) and grounds them in structured normative documentation via Semantic Chunking. Introduces MINDS dataset with 31 bilingual multi-turn conversations annotated for norm adherence.

Result: Norm-RAG improves norm detection and generalization, showing enhanced performance for culturally adaptive and socially intelligent dialogue systems.

Conclusion: The framework enables interpretable and context-aware reasoning about norm adherence and violation across multilingual dialogues, addressing limitations of prior approaches.

Abstract: Social norms are implicit, culturally grounded expectations that guide interpersonal communication. Unlike factual commonsense, norm reasoning is subjective, context-dependent, and varies across cultures, posing challenges for computational models. Prior works provide valuable normative annotations but mostly target isolated utterances or synthetic dialogues, limiting their ability to capture the fluid, multi-turn nature of real-world conversations. In this work, we present Norm-RAG, a retrieval-augmented, agentic framework for nuanced social norm inference in multi-turn dialogues. Norm-RAG models utterance-level attributes including communicative intent, speaker roles, interpersonal framing, and linguistic cues and grounds them in structured normative documentation retrieved via a novel Semantic Chunking approach. This enables interpretable and context-aware reasoning about norm adherence and violation across multilingual dialogues. We further introduce MINDS (Multilingual Interactions with Norm-Driven Speech), a bilingual dataset comprising 31 multi-turn Mandarin-English and Spanish-English conversations. Each turn is annotated for norm category and adherence status using multi-annotator consensus, reflecting cross-cultural and realistic norm expression. Our experiments show that Norm-RAG improves norm detection and generalization, demonstrates improved performance for culturally adaptive and socially intelligent dialogue systems.

Method: Used GPT-4o-mini to generate KCs for 646 questions, then applied cosine similarity-based semantic merging to reduce redundant KC labels.

Result: Initial LLM approach performed worse than experts (RMSE 0.4285 vs 0.4206) with 569 vs 101 KCs. Semantic merging at 0.8 threshold reduced KCs to 428 and improved RMSE to 0.4259.

Conclusion: LLM generation alone is insufficient for KC identification, but combining it with semantic merging provides a viable automated approach.

Abstract: Knowledge Components (KCs) are foundational to adaptive learning systems, but their manual identification by domain experts is a significant bottleneck. While Large Language Models (LLMs) offer a promising avenue for automating this process, prior research has been limited to small datasets and has been shown to produce superfluous, redundant KC labels. This study addresses these limitations by first scaling a “simulated textbook” LLM prompting strategy (using GPT-4o-mini) to a larger dataset of 646 multiple-choice questions. We found that this initial automated approach performed significantly worse than an expert-designed KC model (RMSE 0.4285 vs. 0.4206) and generated an excessive number of KCs (569 vs. 101). To address the issue of redundancy, we proposed and evaluated a novel method for merging semantically similar KC labels based on their cosine similarity. This merging strategy significantly improved the model’s performance; a model using a cosine similarity threshold of 0.8 achieved the best result, reducing the KC count to 428 and improving the RMSE to 0.4259. This demonstrates that while scaled LLM generation alone is insufficient, combining it with a semantic merging technique offers a viable path toward automating and refining KC identification.

Method: REAP uses Sub-task Planner (SP) for global reasoning direction and task state evaluation, and Fact Extractor (FE) for fine-grained content analysis. Both modules incrementally build a coherent global knowledge representation through a unified task paradigm for multi-task fine-tuning.

Result: Extensive experiments on multiple public multi-hop datasets show REAP significantly outperforms existing RAG methods in both in-domain and out-of-domain settings.

Conclusion: REAP effectively addresses limitations in multi-hop reasoning by providing global planning, better content utilization, and traceable reasoning processes, demonstrating superior performance in complex reasoning tasks.

Abstract: Retrieval-augmented generation (RAG) has been extensively employed to mitigate hallucinations in large language models (LLMs). However, existing methods for multi-hop reasoning tasks often lack global planning, increasing the risk of falling into local reasoning impasses. Insufficient exploitation of retrieved content and the neglect of latent clues fail to ensure the accuracy of reasoning outcomes. To overcome these limitations, we propose Recursive Evaluation and Adaptive Planning (REAP), whose core idea is to explicitly maintain structured sub-tasks and facts related to the current task through the Sub-task Planner (SP) and Fact Extractor (FE) modules. SP maintains a global perspective, guiding the overall reasoning direction and evaluating the task state based on the outcomes of FE, enabling dynamic optimization of the task-solving trajectory. FE performs fine-grained analysis over retrieved content to extract reliable answers and clues. These two modules incrementally enrich a logically coherent representation of global knowledge, enhancing the reliability and the traceability of the reasoning process. Furthermore, we propose a unified task paradigm design that enables effective multi-task fine-tuning, significantly enhancing SP’s performance on complex, data-scarce tasks. We conduct extensive experiments on multiple public multi-hop datasets, and the results demonstrate that our method significantly outperforms existing RAG methods in both in-domain and out-of-domain settings, validating its effectiveness in complex multi-hop reasoning tasks.

Method: Systematic evaluation using controlled perturbations including label-flipping probes to test model robustness on numerical claims and evidence pairs.

Result: Leading proprietary models experience up to 62% accuracy drops under perturbations; no model is robust across all conditions; extended context generally reduces accuracy but enriched demonstrations help recovery.

Conclusion: Numerical fact-checking has critical limitations and robustness remains an open challenge for current language models.

Abstract: Large language models show strong performance on knowledge intensive tasks such as fact-checking and question answering, yet they often struggle with numerical reasoning. We present a systematic evaluation of state-of-the-art models for veracity prediction on numerical claims and evidence pairs using controlled perturbations, including label-flipping probes, to test robustness. Our results indicate that even leading proprietary systems experience accuracy drops of up to 62% under certain perturbations. No model proves to be robust across all conditions. We further find that increasing context length generally reduces accuracy, but when extended context is enriched with perturbed demonstrations, most models substantially recover. These findings highlight critical limitations in numerical fact-checking and suggest that robustness remains an open challenge for current language models.

Method: ETC utilizes first- and second-order differences of the entropy sequence to detect emerging uncertainty trends, enabling earlier and more precise retrieval timing without requiring training.

Result: Experiments on six QA benchmarks with three LLM backbones show ETC consistently outperforms strong baselines while reducing retrieval frequency, particularly effective in domain-specific scenarios with robust generalization.

Conclusion: Trend-aware uncertainty modeling yields more effective retrieval timing, and ETC is plug-and-play, model-agnostic, and easily integrable into existing decoding pipelines.

Abstract: Dynamic retrieval-augmented generation (RAG) allows large language models (LLMs) to fetch external knowledge on demand, offering greater adaptability than static RAG. A central challenge in this setting lies in determining the optimal timing for retrieval. Existing methods often trigger retrieval based on low token-level confidence, which may lead to delayed intervention after errors have already propagated. We introduce Entropy-Trend Constraint (ETC), a training-free method that determines optimal retrieval timing by modeling the dynamics of token-level uncertainty. Specifically, ETC utilizes first- and second-order differences of the entropy sequence to detect emerging uncertainty trends, enabling earlier and more precise retrieval. Experiments on six QA benchmarks with three LLM backbones demonstrate that ETC consistently outperforms strong baselines while reducing retrieval frequency. ETC is particularly effective in domain-specific scenarios, exhibiting robust generalization capabilities. Ablation studies and qualitative analyses further confirm that trend-aware uncertainty modeling yields more effective retrieval timing. The method is plug-and-play, model-agnostic, and readily integrable into existing decoding pipelines. Implementation code is included in the supplementary materials.

Method: Proposes Soft Constrained Decoding (SCD), a training-free decoding strategy that gently steers generation toward target language by penalizing non-target-language tokens. It’s model-agnostic and works with any generation algorithm.

Result: Experiments across three multilingual datasets and diverse languages show SCD consistently improves language alignment and task performance in multilingual RAG settings.

Conclusion: SCD provides an effective and generalizable solution to language drift in multilingual RAG, addressing decoder-level collapse and English semantic attraction without requiring model modifications or additional data.

Abstract: Multilingual Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to perform knowledge-intensive tasks in multilingual settings by leveraging retrieved documents as external evidence. However, when the retrieved evidence differs in language from the user query and in-context exemplars, the model often exhibits language drift by generating responses in an unintended language. This phenomenon is especially pronounced during reasoning-intensive decoding, such as Chain-of-Thought (CoT) generation, where intermediate steps introduce further language instability. In this paper, we systematically study output language drift in multilingual RAG across multiple datasets, languages, and LLM backbones. Our controlled experiments reveal that the drift results not from comprehension failure but from decoder-level collapse, where dominant token distributions and high-frequency English patterns dominate the intended generation language. We further observe that English serves as a semantic attractor under cross-lingual conditions, emerging as both the strongest interference source and the most frequent fallback language. To mitigate this, we propose Soft Constrained Decoding (SCD), a lightweight, training-free decoding strategy that gently steers generation toward the target language by penalizing non-target-language tokens. SCD is model-agnostic and can be applied to any generation algorithm without modifying the architecture or requiring additional data. Experiments across three multilingual datasets and multiple typologically diverse languages show that SCD consistently improves language alignment and task performance, providing an effective and generalizable solution in multilingual RAG.

Method: Created FinNuE diagnostic dataset with controlled numerical perturbations across earnings calls, regulatory filings, social media, and news articles to test BERTScore’s numerical sensitivity.

Result: BERTScore fails to distinguish semantically critical numerical differences, often assigning high similarity scores to financially divergent text pairs.

Conclusion: Embedding-based metrics have fundamental limitations for finance, motivating the need for numerically-aware evaluation frameworks in financial NLP.

Abstract: BERTScore has become a widely adopted metric for evaluating semantic similarity between natural language sentences. However, we identify a critical limitation: BERTScore exhibits low sensitivity to numerical variation, a significant weakness in finance where numerical precision directly affects meaning (e.g., distinguishing a 2% gain from a 20% loss). We introduce FinNuE, a diagnostic dataset constructed with controlled numerical perturbations across earnings calls, regulatory filings, social media, and news articles. Using FinNuE, demonstrate that BERTScore fails to distinguish semantically critical numerical differences, often assigning high similarity scores to financially divergent text pairs. Our findings reveal fundamental limitations of embedding-based metrics for finance and motivate numerically-aware evaluation frameworks for financial NLP.

Method: Developed NCERT-QA dataset with Factoid, Inferential, and Other question types; evaluated using meta-prompt, few-shot, and CoT-style prompting on various open-source and high-end LLMs.

Result: Comprehensive evaluation of LLM performance on curriculum-aligned content, identifying strengths and limitations of different models as AI learning tools.

Conclusion: Provides framework for assessing LLM suitability in formal education systems, highlighting the need for curriculum-specific adaptations and proper evaluation methodologies.

Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in understanding and generating human-like content. This has revolutionized various sectors such as healthcare, software development, and education. In education, LLMs offer potential for personalized and interactive learning experiences, especially in regions with limited teaching resources. However, adapting these models effectively to curriculum-specific content, such as the National Council of Educational Research and Training (NCERT) syllabus in India, presents unique challenges in terms of accuracy, alignment, and pedagogical relevance. In this paper, we present the framework “PustakAI”\footnote{Pustak means `book’ in many Indian languages.} for the design and evaluation of a novel question-answering dataset “NCERT-QA” aligned with the NCERT curriculum for English and Science subjects of grades 6 to 8. We classify the curated QA pairs as Factoid, Inferential, and Others (evaluative and reasoning). We evaluate the dataset with various prompting techniques, such as meta-prompt, few-shot, and CoT-style prompting, using diverse evaluation metrics to understand which approach aligns more efficiently with the structure and demands of the curriculum. Along with the usability of the dataset, we analyze the strengths and limitations of current open-source LLMs (Gemma3:1b, Llama3.2:3b, and Nemotron-mini:4b) and high-end LLMs (Llama-4-Scout-17B and Deepseek-r1-70B) as AI-based learning tools in formal education systems.

Method: Segments long inputs into overlapping chunks, generates compressed context-aware representations, and uses a novel parameter-free fusion mechanism that enriches chunk boundaries with cumulative context from preceding and succeeding chunks.

Result: Experiments on long-document summarization show the method is highly competitive with and often outperforms state-of-the-art approaches without requiring architectural modifications or external retrieval.

Conclusion: ScaleFormer provides a simple plug-and-play framework that enables pre-trained models to effectively reason over long-form text with linear complexity while maintaining strong performance.

Abstract: The quadratic complexity of standard self-attention severely limits the application of Transformer-based models to long-context tasks. While efficient Transformer variants exist, they often require architectural changes and costly pre-training from scratch. To circumvent this, we propose ScaleFormer(Span Representation Cumulation for Long-Context Transformer) - a simple and effective plug-and-play framework that adapts off-the-shelf pre-trained encoder-decoder models to process long sequences without requiring architectural modifications. Our approach segments long inputs into overlapping chunks and generates a compressed, context-aware representation for the decoder. The core of our method is a novel, parameter-free fusion mechanism that endows each chunk’s representation with structural awareness of its position within the document. It achieves this by enriching each chunk’s boundary representations with cumulative context vectors from all preceding and succeeding chunks. This strategy provides the model with a strong signal of the document’s narrative flow, achieves linear complexity, and enables pre-trained models to reason effectively over long-form text. Experiments on long-document summarization show that our method is highly competitive with and often outperforms state-of-the-art approaches without requiring architectural modifications or external retrieval mechanisms.

Method: Created LEX-ICON dataset with 8,052 words from 4 languages and 2,930 pseudo-words, analyzed MLLMs’ layer-wise processing using phoneme-level attention scores across 25 semantic dimensions in both text and audio inputs.

Result: MLLMs demonstrate phonetic intuitions consistent with linguistic research and show phonosemantic attention patterns focusing on iconic phonemes across multiple semantic dimensions.

Conclusion: This work bridges AI and cognitive linguistics, providing first large-scale quantitative analysis of phonetic iconicity in MLLMs’ interpretability, showing models can capture sound-meaning associations similar to human cognition.

Abstract: Sound symbolism is a linguistic concept that refers to non-arbitrary associations between phonetic forms and their meanings. We suggest that this can be a compelling probe into how Multimodal Large Language Models (MLLMs) interpret auditory information in human languages. We investigate MLLMs’ performance on phonetic iconicity across textual (orthographic and IPA) and auditory forms of inputs with up to 25 semantic dimensions (e.g., sharp vs. round), observing models’ layer-wise information processing by measuring phoneme-level attention fraction scores. To this end, we present LEX-ICON, an extensive mimetic word dataset consisting of 8,052 words from four natural languages (English, French, Japanese, and Korean) and 2,930 systematically constructed pseudo-words, annotated with semantic features applied across both text and audio modalities. Our key findings demonstrate (1) MLLMs’ phonetic intuitions that align with existing linguistic research across multiple semantic dimensions and (2) phonosemantic attention patterns that highlight models’ focus on iconic phonemes. These results bridge domains of artificial intelligence and cognitive linguistics, providing the first large-scale, quantitative analyses of phonetic iconicity in terms of MLLMs’ interpretability.

Method: GraphIF has three components: relation extraction via action-triggered mechanisms to build structured graphs, relation graph prompt generation to convert graphs into natural language prompts, and response rewriting to refine LLM outputs using graph prompts.

Result: Extensive experiments on two long multi-turn dialogue datasets show GraphIF can be seamlessly integrated into instruction-tuned LLMs and leads to significant improvements across all four multi-turn instruction-following evaluation metrics.

Conclusion: GraphIF effectively bridges the gap in leveraging graph structures for multi-turn instruction following, providing a plug-and-play solution that significantly enhances LLMs’ capabilities without requiring model retraining.

Abstract: Multi-turn instruction following is essential for building intelligent conversational systems that can consistently adhere to instructions across dialogue turns. However, existing approaches to enhancing multi-turn instruction following primarily rely on collecting or generating large-scale multi-turn dialogue datasets to fine-tune large language models (LLMs), which treat each response generation as an isolated task and fail to explicitly incorporate multi-turn instruction following into the optimization objectives. As a result, instruction-tuned LLMs often struggle with complex long-distance constraints. In multi-turn dialogues, relational constraints across turns can be naturally modeled as labeled directed edges, making graph structures particularly suitable for modeling multi-turn instruction following. Despite this potential, leveraging graph structures to enhance the multi-turn instruction following capabilities of LLMs remains unexplored. To bridge this gap, we propose GraphIF, a plug-and-play framework that models multi-turn dialogues as directed relation graphs and leverages graph prompts to enhance the instruction following capabilities of LLMs. GraphIF comprises three key components: (1) an agent-based relation extraction module that captures inter-turn semantic relations via action-triggered mechanisms to construct structured graphs; (2) a relation graph prompt generation module that converts structured graph information into natural language prompts; and (3) a response rewriting module that refines initial LLM outputs using the generated graph prompts. Extensive experiments on two long multi-turn dialogue datasets demonstrate that GraphIF can be seamlessly integrated into instruction-tuned LLMs and leads to significant improvements across all four multi-turn instruction-following evaluation metrics.

Method: Fine-tuned pre-trained ECAPA-TDNN models and Whisper encoder blocks with attention pooling and classification layers, investigated training data size and model parameter effects.

Result: Small F1 score decrease when using only 30% of training data, demonstrating data efficiency.

Conclusion: Dataset and models open-sourced to enable reproduction and further ADI research.

Abstract: We present ADI-20, an extension of the previously published ADI-17 Arabic Dialect Identification (ADI) dataset. ADI-20 covers all Arabic-speaking countries’ dialects. It comprises 3,556 hours from 19 Arabic dialects in addition to Modern Standard Arabic (MSA). We used this dataset to train and evaluate various state-of-the-art ADI systems. We explored fine-tuning pre-trained ECAPA-TDNN-based models, as well as Whisper encoder blocks coupled with an attention pooling layer and a classification dense layer. We investigated the effect of (i) training data size and (ii) the model’s number of parameters on identification performance. Our results show a small decrease in F1 score while using only 30% of the original training data. We open-source our collected data and trained models to enable the reproduction of our work, as well as support further research in ADI.

Method: Adapted two existing datasets of scientific papers with annotations for multimodal claim verification, then evaluated 12 multimodal LLMs on their ability to verify claims using both table and chart evidence.

Result: Current models perform better with table-based evidence while struggling with chart-based evidence. Smaller multimodal LLMs (under 8B) show weak correlation between table and chart performance, indicating limited cross-modal generalization. Humans maintain strong performance across both formats.

Conclusion: There is a critical gap in current models’ multimodal reasoning capabilities, particularly in chart understanding. Future multimodal LLMs should focus on improving chart understanding to better support scientific claim verification.

Abstract: With the growing number of submitted scientific papers, there is an increasing demand for systems that can assist reviewers in evaluating research claims. Experimental results are a core component of scientific work, often presented in varying formats such as tables or charts. Understanding how robust current multimodal large language models (multimodal LLMs) are at verifying scientific claims across different evidence formats remains an important and underexplored challenge. In this paper, we design and conduct a series of experiments to assess the ability of multimodal LLMs to verify scientific claims using both tables and charts as evidence. To enable this evaluation, we adapt two existing datasets of scientific papers by incorporating annotations and structures necessary for a multimodal claim verification task. Using this adapted dataset, we evaluate 12 multimodal LLMs and find that current models perform better with table-based evidence while struggling with chart-based evidence. We further conduct human evaluations and observe that humans maintain strong performance across both formats, unlike the models. Our analysis also reveals that smaller multimodal LLMs (under 8B) show weak correlation in performance between table-based and chart-based tasks, indicating limited cross-modal generalization. These findings highlight a critical gap in current models’ multimodal reasoning capabilities. We suggest that future multimodal LLMs should place greater emphasis on improving chart understanding to better support scientific claim verification.

Method: Fine-tuned Whisper-large-v3 encoder with data augmentation for ADI, and fine-tuned SeamlessM4T-v2 Large separately for each of eight Arabic dialects for ASR.

Result: ADI accuracy: 79.83% (1st place), ASR: 38.54% WER and 14.53% CER (2nd place) on official test set.

Conclusion: Large pre-trained speech models with targeted fine-tuning are highly effective for Arabic speech processing tasks.

Abstract: This paper describes Elyadata & LIA’s joint submission to the NADI multi-dialectal Arabic Speech Processing 2025. We participated in the Spoken Arabic Dialect Identification (ADI) and multi-dialectal Arabic ASR subtasks. Our submission ranked first for the ADI subtask and second for the multi-dialectal Arabic ASR subtask among all participants. Our ADI system is a fine-tuned Whisper-large-v3 encoder with data augmentation. This system obtained the highest ADI accuracy score of \textbf{79.83%} on the official test set. For multi-dialectal Arabic ASR, we fine-tuned SeamlessM4T-v2 Large (Egyptian variant) separately for each of the eight considered dialects. Overall, we obtained an average WER and CER of \textbf{38.54%} and \textbf{14.53%}, respectively, on the test set. Our results demonstrate the effectiveness of large pre-trained speech models with targeted fine-tuning for Arabic speech processing.

Method: 1) Interrogated a GPT-based language model (Microsoft Copilot) to reveal its knowledge about potential military applications and critically assessed the information. 2) Studied how commercial cloud services (Microsoft Azure) could be used to build such applications and assessed their feasibility.

Result: The analysis revealed that summarization and generative properties of language models directly facilitate many military applications, and other features may find particular uses. Commercial cloud services were found to be readily usable for building such applications.

Conclusion: Large language models and natural language processing technologies have significant potential for military applications, with summarization and generative capabilities being particularly useful, and these can be implemented using existing commercial cloud infrastructure.

Abstract: In this paper, military use cases or applications and implementation thereof are considered for natural language processing and large language models, which have broken into fame with the invention of the generative pre-trained transformer (GPT) and the extensive foundation model pretraining done by OpenAI for ChatGPT and others. First, we interrogate a GPT-based language model (viz. Microsoft Copilot) to make it reveal its own knowledge about their potential military applications and then critically assess the information. Second, we study how commercial cloud services (viz. Microsoft Azure) could be used readily to build such applications and assess which of them are feasible. We conclude that the summarization and generative properties of language models directly facilitate many applications at large and other features may find particular uses.

Method: Approaches bias mitigation through causal lens, proposing method to reduce event- and domain-related biases.

Result: Outperforms multiple baselines by up to +1.9% F1 score and significantly improves PLM-based classifier across three disaster classification tasks.

Conclusion: Causal approach effectively reduces biases in disaster monitoring systems, enhancing generalization capabilities for future events.

Abstract: Due to the rapid growth of social media platforms, these tools have become essential for monitoring information during ongoing disaster events. However, extracting valuable insights requires real-time processing of vast amounts of data. A major challenge in existing systems is their exposure to event-related biases, which negatively affects their ability to generalize to emerging events. While recent advancements in debiasing and causal learning offer promising solutions, they remain underexplored in the disaster event domain. In this work, we approach bias mitigation through a causal lens and propose a method to reduce event- and domain-related biases, enhancing generalization to future events. Our approach outperforms multiple baselines by up to +1.9% F1 and significantly improves a PLM-based classifier across three disaster classification tasks.

Method: Developed VISTA - a web-based visual interactive system that visualizes context influence, allows interactive conversation history modification for “what-if” analyses, and automatically generates reasoning dependency trees.

Result: VISTA significantly reduces the complexity of analyzing reasoning chains and provides transparent views of models’ step-by-step logical paths across different LLMs.

Conclusion: VISTA facilitates deeper understanding of LLM capabilities and limitations, is open-source, and supports easy integration of custom benchmarks and local models.

Abstract: Recent research has increasingly focused on the reasoning capabilities of Large Language Models (LLMs) in multi-turn interactions, as these scenarios more closely mirror real-world problem-solving. However, analyzing the intricate reasoning processes within these interactions presents a significant challenge due to complex contextual dependencies and a lack of specialized visualization tools, leading to a high cognitive load for researchers. To address this gap, we present VISTA, an web-based Visual Interactive System for Textual Analytics in multi-turn reasoning tasks. VISTA allows users to visualize the influence of context on model decisions and interactively modify conversation histories to conduct “what-if” analyses across different models. Furthermore, the platform can automatically parse a session and generate a reasoning dependency tree, offering a transparent view of the model’s step-by-step logical path. By providing a unified and interactive framework, VISTA significantly reduces the complexity of analyzing reasoning chains, thereby facilitating a deeper understanding of the capabilities and limitations of current LLMs. The platform is open-source and supports easy integration of custom benchmarks and local models.

Method: Proposed Text2SQL-Flow framework with six augmentation dimensions, SQL execution verification, natural language question generation, chain-of-thought reasoning, data classification, and modular Database Manager for cross-database compatibility.

Result: Built SQLFlow dataset with 89,544 annotated examples. Fine-tuning on SQLFlow improves open-source LLM performance, and masked alignment retrieval method with closed-source LLMs outperforms existing methods.

Conclusion: Establishes scalable data-centric foundation for Text-to-SQL systems and demonstrates critical role of high-quality structured data in modern AI.

Abstract: The data-centric paradigm has become pivotal in AI, especially for Text-to-SQL, where performance is limited by scarce, simplistic, and low-diversity datasets. To address this, we propose Text2SQL-Flow, a SQL-aware data augmentation framework that generates large-scale, semantically valid, and structurally diverse Text-to-SQL pairs from minimal seed data. It operates across six augmentation dimensions and integrates an end-to-end pipeline featuring SQL execution verification, natural language question generation, chain-of-thought reasoning traces, and data classification. A modular Database Manager ensures cross-database compatibility and scalability. Using this framework, we build SQLFlow, a high-quality dataset of 89,544 annotated examples. We evaluate SQLFlow in two settings: (1) For open-source LLMs, fine-tuning on SQLFlow consistently improves performance across benchmarks under the same data budget. (2) For closed-source LLMs, we introduce a masked alignment retrieval method that treats SQLFlow as both knowledge base and training data for the retriever. This enables structure-aware example matching by modeling fine-grained alignments between questions and SQL queries. Experiments show our retrieval strategy outperforms existing methods, underscoring the value of SQLFlow’s high-fidelity data and our novel technique. Our work establishes a scalable, data-centric foundation for advancing Text-to-SQL systems and highlights the critical role of high-quality structured data in modern AI.

Method: Created EffiReason-Bench with verified CoT annotations via a pipeline enforcing standardized reasoning structures, comprehensive option-wise analysis, and human verification. Evaluated 7 methods across 6 LLMs (1B-70B) on 4 datasets using the E3-Score metric.

Result: No single method universally dominates; optimal strategies depend on backbone scale, task complexity, and architecture. The E3-Score provides smooth, stable evaluation without discontinuities or heavy reliance on heuristics.

Conclusion: Efficient reasoning methods require context-specific selection based on model scale, task type, and architecture, with EffiReason-Bench enabling rigorous cross-paradigm evaluation.

Abstract: Large language models (LLMs) with Chain-of-Thought (CoT) prompting achieve strong reasoning but often produce unnecessarily long explanations, increasing cost and sometimes reducing accuracy. Fair comparison of efficiency-oriented approaches is hindered by fragmented evaluation practices. We introduce EffiReason-Bench, a unified benchmark for rigorous cross-paradigm evaluation of efficient reasoning methods across three categories: Reasoning Blueprints, Dynamic Execution, and Post-hoc Refinement. To enable step-by-step evaluation, we construct verified CoT annotations for CommonsenseQA and LogiQA via a pipeline that enforces standardized reasoning structures, comprehensive option-wise analysis, and human verification. We evaluate 7 methods across 6 open-source LLMs (1B-70B) on 4 datasets spanning mathematics, commonsense, and logic, and propose the E3-Score, a principled metric inspired by economic trade-off modeling that provides smooth, stable evaluation without discontinuities or heavy reliance on heuristics. Experiments show that no single method universally dominates; optimal strategies depend on backbone scale, task complexity, and architecture.

Method: Two-stage training: Persona-aware Learning and Persona Alignment, with Select then Generate inference strategy to improve persona sensitivity at semantics level.

Result: Outperforms state-of-the-art personalized dialogue methods and large language models in experiments.

Conclusion: Treating persona alignment directly as training objective effectively addresses persona neglect in dialogue generation, producing more persona-relevant responses.

Abstract: Personalized dialogue generation aims to leverage persona profiles and dialogue history to generate persona-relevant and consistent responses. Mainstream models typically rely on token-level language model training with persona dialogue data, such as Next Token Prediction, to implicitly achieve personalization, making these methods tend to neglect the given personas and generate generic responses. To address this issue, we propose a novel Persona-Aware Alignment Framework (PAL), which directly treats persona alignment as the training objective of dialogue generation. Specifically, PAL employs a two-stage training method including Persona-aware Learning and Persona Alignment, equipped with an easy-to-use inference strategy Select then Generate, to improve persona sensitivity and generate more persona-relevant responses at the semantics level. Through extensive experiments, we demonstrate that our framework outperforms many state-of-the-art personalized dialogue methods and large language models.

Method: Two-stage framework: 1) Filter training data based on language separability scores (how distinguishable languages are in representation space), 2) Refine subset using existing selection methods like text quality, diversity, or task relevance.

Result: Improves effectiveness and generalizability of existing selection methods across 6 benchmarks and 22 languages, especially for understanding tasks and low-resource languages. Highly separable samples form clearer language boundaries and support faster adaptation.

Conclusion: Language separability offers a new perspective on data utility in multilingual contexts and can serve as an effective signal for multilingual curriculum learning, supporting development of more linguistically informed LLMs.

Abstract: Joint multilingual instruction tuning is a widely adopted approach to improve the multilingual instruction-following ability and downstream performance of large language models (LLMs), but the resulting multilingual capability remains highly sensitive to the composition and selection of the training data. Existing selection methods, often based on features like text quality, diversity, or task relevance, typically overlook the intrinsic linguistic structure of multilingual data. In this paper, we propose LangGPS, a lightweight two-stage pre-selection framework guided by language separability which quantifies how well samples in different languages can be distinguished in the model’s representation space. LangGPS first filters training data based on separability scores and then refines the subset using existing selection methods. Extensive experiments across six benchmarks and 22 languages demonstrate that applying LangGPS on top of existing selection methods improves their effectiveness and generalizability in multilingual training, especially for understanding tasks and low-resource languages. Further analysis reveals that highly separable samples facilitate the formation of clearer language boundaries and support faster adaptation, while low-separability samples tend to function as bridges for cross-lingual alignment. Besides, we also find that language separability can serve as an effective signal for multilingual curriculum learning, where interleaving samples with diverse separability levels yields stable and generalizable gains. Together, we hope our work offers a new perspective on data utility in multilingual contexts and support the development of more linguistically informed LLMs.

Method: Combines semantic understanding with keyword precision, using two complementary retrieval strategies and consumer-grade hardware acceleration, operating entirely on local infrastructure without internet access.

Result: Achieves competitive accuracy on complex queries across legal, scientific, and conversational documents while keeping all data on local machines. The system delivers reliable answers with minimal errors.

Conclusion: Privacy and performance need not be mutually exclusive in enterprise AI deployment. Organizations can adopt conversational document AI without transmitting proprietary information to external providers.

Abstract: Organizations handling sensitive documents face a critical dilemma: adopt cloud-based AI systems that offer powerful question-answering capabilities but compromise data privacy, or maintain local processing that ensures security but delivers poor accuracy. We present a question-answering system that resolves this trade-off by combining semantic understanding with keyword precision, operating entirely on local infrastructure without internet access. Our approach demonstrates that organizations can achieve competitive accuracy on complex queries across legal, scientific, and conversational documents while keeping all data on their machines. By balancing two complementary retrieval strategies and using consumer-grade hardware acceleration, the system delivers reliable answers with minimal errors, letting banks, hospitals, and law firms adopt conversational document AI without transmitting proprietary information to external providers. This work establishes that privacy and performance need not be mutually exclusive in enterprise AI deployment.

Method: Built OPS benchmark to quantify LLM behavior differences, conducted statistical preference analysis, and proposed perplexity-aware reinforcement learning (Group Relative Policy Optimization) to rectify evaluation preferences.

Result: Extensive experiments on OPS and existing critic benchmarks demonstrate the validity of the proposed method in improving critiquing capability.

Conclusion: The imbalanced evaluation preference is a key factor limiting LLMs’ critiquing performance, and the proposed perplexity-aware reinforcement learning effectively addresses this issue.

Abstract: To improve Multi-step Mathematical Reasoning (MsMR) of Large Language Models (LLMs), it is crucial to obtain scalable supervision from the corpus by automatically critiquing mistakes in the reasoning process of MsMR and rendering a final verdict of the problem-solution. Most existing methods rely on crafting high-quality supervised fine-tuning demonstrations for critiquing capability enhancement and pay little attention to delving into the underlying reason for the poor critiquing performance of LLMs. In this paper, we orthogonally quantify and investigate the potential reason – imbalanced evaluation preference, and conduct a statistical preference analysis. Motivated by the analysis of the reason, a novel perplexity-aware reinforcement learning algorithm is proposed to rectify the evaluation preference, elevating the critiquing capability. Specifically, to probe into LLMs’ critiquing characteristics, a One-to-many Problem-Solution (OPS) benchmark is meticulously constructed to quantify the behavior difference of LLMs when evaluating the problem solutions generated by itself and others. Then, to investigate the behavior difference in depth, we conduct a statistical preference analysis oriented on perplexity and find an intriguing phenomenon – ``LLMs incline to judge solutions with lower perplexity as correct’’, which is dubbed as \textit{imbalanced evaluation preference}. To rectify this preference, we regard perplexity as the baton in the algorithm of Group Relative Policy Optimization, supporting the LLMs to explore trajectories that judge lower perplexity as wrong and higher perplexity as correct. Extensive experimental results on our built OPS and existing available critic benchmarks demonstrate the validity of our method.

Method: Constructed large-scale synthetic dataset using 5 generation techniques, explored grounding in documents/personas/topics, compared translation vs native generation, and developed modular quality evaluation pipeline with script/language detection, metadata checks, n-gram analysis, and perplexity filtering.

Result: Empirical results reveal key trade-offs in generation strategies and highlight best practices for constructing effective multilingual corpora.

Conclusion: The framework enables robust quality control across diverse scripts and linguistic contexts, providing scalable solution for multilingual pretraining data generation.

Abstract: In the context of pretraining of Large Language Models (LLMs), synthetic data has emerged as an alternative for generating high-quality pretraining data at scale. This is particularly beneficial in low-resource language settings where the benefits of recent LLMs have been unevenly distributed across languages. In this work, we present a systematic study on the generation and evaluation of synthetic multilingual pretraining data for Indic languages, where we construct a large-scale synthetic dataset BhashaKritika, comprising 540B tokens using 5 different techniques for 10 languages. We explore the impact of grounding generation in documents, personas, and topics. We analyze how language choice, both in the prompt instructions and document grounding, affects data quality, and we compare translations of English content with native generation in Indic languages. To support scalable and language-sensitive evaluation, we introduce a modular quality evaluation pipeline that integrates script and language detection, metadata consistency checks, n-gram repetition analysis, and perplexity-based filtering using KenLM models. Our framework enables robust quality control across diverse scripts and linguistic contexts. Empirical results through model runs reveal key trade-offs in generation strategies and highlight best practices for constructing effective multilingual corpora.

Method: Five-step methodology combining annotation models, ontological frameworks, and LLM-based extraction: foundational analysis, annotation schema development, pipeline architecture, integration refinement, and evaluation using three LLMs (Claude Sonnet 3.7, Llama 3.3 70B, GPT-4o-mini).

Result: High performance metrics: 0.96-0.99 F1 for metadata extraction, 0.7-0.8 F1 for entity recognition, 0.65-0.75 F1 for hypothesis extraction, 0.95-0.97 for evidence extraction, and 0.62 G-EVAL for discourse representation. Smaller models performed competitively.

Conclusion: ATR4CH provides the first systematic methodology for coordinating LLM-based extraction with Cultural Heritage ontologies, offering a replicable framework adaptable across CH domains, though human oversight is needed during post-processing.

Abstract: Cultural Heritage texts contain rich knowledge that is difficult to query systematically due to the challenges of converting unstructured discourse into structured Knowledge Graphs (KGs). This paper introduces ATR4CH (Adaptive Text-to-RDF for Cultural Heritage), a systematic five-step methodology for Large Language Model-based Knowledge Extraction from Cultural Heritage documents. We validate the methodology through a case study on authenticity assessment debates. Methodology - ATR4CH combines annotation models, ontological frameworks, and LLM-based extraction through iterative development: foundational analysis, annotation schema development, pipeline architecture, integration refinement, and comprehensive evaluation. We demonstrate the approach using Wikipedia articles about disputed items (documents, artifacts…), implementing a sequential pipeline with three LLMs (Claude Sonnet 3.7, Llama 3.3 70B, GPT-4o-mini). Findings - The methodology successfully extracts complex Cultural Heritage knowledge: 0.96-0.99 F1 for metadata extraction, 0.7-0.8 F1 for entity recognition, 0.65-0.75 F1 for hypothesis extraction, 0.95-0.97 for evidence extraction, and 0.62 G-EVAL for discourse representation. Smaller models performed competitively, enabling cost-effective deployment. Originality - This is the first systematic methodology for coordinating LLM-based extraction with Cultural Heritage ontologies. ATR4CH provides a replicable framework adaptable across CH domains and institutional resources. Research Limitations - The produced KG is limited to Wikipedia articles. While the results are encouraging, human oversight is necessary during post-processing. Practical Implications - ATR4CH enables Cultural Heritage institutions to systematically convert textual knowledge into queryable KGs, supporting automated metadata enrichment and knowledge discovery.

Method: TruthfulRAG constructs Knowledge Graphs by extracting triples from retrieved content, uses query-based graph retrieval to identify relevant knowledge, and employs entropy-based filtering mechanisms to precisely locate conflicting elements and mitigate factual inconsistencies.

Result: Extensive experiments show that TruthfulRAG outperforms existing methods, effectively alleviating knowledge conflicts and improving the robustness and trustworthiness of RAG systems.

Conclusion: The proposed TruthfulRAG framework successfully addresses factual-level knowledge conflicts in RAG systems through Knowledge Graph integration, enabling LLMs to generate more faithful and accurate responses.

Abstract: Retrieval-Augmented Generation (RAG) has emerged as a powerful framework for enhancing the capabilities of Large Language Models (LLMs) by integrating retrieval-based methods with generative models. As external knowledge repositories continue to expand and the parametric knowledge within models becomes outdated, a critical challenge for RAG systems is resolving conflicts between retrieved external information and LLMs’ internal knowledge, which can significantly compromise the accuracy and reliability of generated content. However, existing approaches to conflict resolution typically operate at the token or semantic level, often leading to fragmented and partial understanding of factual discrepancies between LLMs’ knowledge and context, particularly in knowledge-intensive tasks. To address this limitation, we propose TruthfulRAG, the first framework that leverages Knowledge Graphs (KGs) to resolve factual-level knowledge conflicts in RAG systems. Specifically, TruthfulRAG constructs KGs by systematically extracting triples from retrieved content, utilizes query-based graph retrieval to identify relevant knowledge, and employs entropy-based filtering mechanisms to precisely locate conflicting elements and mitigate factual inconsistencies, thereby enabling LLMs to generate faithful and accurate responses. Extensive experiments reveal that TruthfulRAG outperforms existing methods, effectively alleviating knowledge conflicts and improving the robustness and trustworthiness of RAG systems.

Method: Analyzing the fundamental mismatch between base LLMs’ pretraining objective (statistical plausibility) and normative reasoning qualities (correctness), showing how valid/invalid conclusions are coincidental byproducts.

Result: Demonstrates that base LLMs generate conclusions as coincidental byproducts of linguistic patterns rather than genuine reasoning attempts, challenging assumptions about their outputs representing bona fide reasoning.

Conclusion: Calls for critical re-examination of existing work and future research to account for methodological pitfalls in evaluating base LLMs’ reasoning capabilities, questioning generalizability to post-trained LLMs.

Abstract: Existing work investigates the reasoning capabilities of large language models (LLMs) to uncover their limitations, human-like biases and underlying processes. Such studies include evaluations of base LLMs (pre-trained on unlabeled corpora only) for this purpose. Our position paper argues that evaluating base LLMs’ reasoning capabilities raises inherent methodological concerns that are overlooked in such existing studies. We highlight the fundamental mismatch between base LLMs’ pretraining objective and normative qualities, such as correctness, by which reasoning is assessed. In particular, we show how base LLMs generate logically valid or invalid conclusions as coincidental byproducts of conforming to purely linguistic patterns of statistical plausibility. This fundamental mismatch challenges the assumptions that (a) base LLMs’ outputs can be assessed as their bona fide attempts at correct answers or conclusions; and (b) conclusions about base LLMs’ reasoning can generalize to post-trained LLMs optimized for successful instruction-following. We call for a critical re-examination of existing work that relies implicitly on these assumptions, and for future work to account for these methodological pitfalls.

Method: DELICATE combines BERT-based encoder with contextual Wikidata information using temporal plausibility and entity type consistency. ENEIDE corpus was semi-automatically extracted from 19th-20th century literary and political texts.

Result: DELICATE outperforms other EL models in historical Italian, even compared to larger architectures with billions of parameters. It provides more explainable and interpretable results through confidence scores and feature sensitivity.

Conclusion: The neuro-symbolic approach effectively addresses EL challenges in humanities, demonstrating superior performance and interpretability for historical Italian texts.

Abstract: In spite of the remarkable advancements in the field of Natural Language Processing, the task of Entity Linking (EL) remains challenging in the field of humanities due to complex document typologies, lack of domain-specific datasets and models, and long-tail entities, i.e., entities under-represented in Knowledge Bases (KBs). The goal of this paper is to address these issues with two main contributions. The first contribution is DELICATE, a novel neuro-symbolic method for EL on historical Italian which combines a BERT-based encoder with contextual information from Wikidata to select appropriate KB entities using temporal plausibility and entity type consistency. The second contribution is ENEIDE, a multi-domain EL corpus in historical Italian semi-automatically extracted from two annotated editions spanning from the 19th to the 20th century and including literary and political texts. Results show how DELICATE outperforms other EL models in historical Italian even if compared with larger architectures with billions of parameters. Moreover, further analyses reveal how DELICATE confidence scores and features sensitivity provide results which are more explainable and interpretable than purely neural methods.

Method: Developed a computational approach implementing three cognitive-inspired principles: analogical structure, contrastive learning, and minimal contextual cues. Tested with structured completion tasks where models identify correct sentence completions from analogical patterns with contrastive alternatives.

Result: Training lightweight models (BERT+CNN, 0.5M parameters) on only 100 structured examples achieved F1=0.95, outperforming zero-shot GPT-3 (F1=0.87). Ablation studies confirmed analogical organization and contrastive structure improve performance, consistently surpassing randomly shuffled baselines. Cross-phenomenon validation replicated efficiency gains.

Conclusion: Analogical paradigm organization enables competitive linguistic rule learning with orders of magnitude less data than conventional approaches require.

Abstract: Large language models achieve strong performance through training on vast datasets. Can analogical paradigm organization enable lightweight models to match this performance with minimal data? We develop a computational approach implementing three cognitive-inspired principles: analogical structure, contrastive learning, and minimal contextual cues. We test this approach with structured completion tasks where models identify correct sentence completions from analogical patterns with contrastive alternatives. Training lightweight models (BERT+CNN, $0.5M$ parameters) on only one hundred structured examples of English causative/inchoative alternations achieves $F1=0.95$, outperforming zero-shot \texttt{GPT-o3} ($F1=0.87$). Ablation studies confirm that analogical organization and contrastive structure improve performance, consistently surpassing randomly shuffled baselines across architectures. Cross-phenomenon validation using unspecified object alternations replicates these efficiency gains, confirming approach robustness. Our results show that analogical paradigm organization enables competitive linguistic rule learning with orders of magnitude less data than conventional approaches require.

Method: Train models with reinforcement learning and customized reward functions using multiple valid answers as supervision to generate multiple interpretation-answer pairs in structured responses.

Result: Achieves higher coverage of valid answers than baselines on conversational QA and semantic parsing; human evaluation shows high alignment between predicted interpretations and answers.

Conclusion: Approach promotes transparency with explicit interpretations, achieves efficiency with single generation step, and supports downstream applications through structured output format.

Abstract: Large language models often respond to ambiguous requests by implicitly committing to one interpretation. Intent misunderstandings can frustrate users and create safety risks. To address this, we propose generating multiple interpretation-answer pairs in a single structured response to ambiguous requests. Our models are trained with reinforcement learning and customized reward functions using multiple valid answers as supervision. Experiments on conversational question answering and semantic parsing demonstrate that our method achieves higher coverage of valid answers than baseline approaches. Human evaluation confirms that predicted interpretations are highly aligned with their answers. Our approach promotes transparency with explicit interpretations, achieves efficiency by requiring only one generation step, and supports downstream applications through its structured output format.

Method: Proposed a benchmark with three well-defined state tracking tasks and analyzed LLM performance across different scenarios, including the use of Chain of Thought mechanisms.

Result: Recent LLMs (GPT-4, Llama3) can effectively track state, especially with Chain of Thought, while older models understand the task initially but fail after several steps.

Conclusion: State tracking is a capability that newer LLMs possess, particularly when enhanced with reasoning mechanisms like Chain of Thought, highlighting the advancement in LLM reasoning abilities.

Abstract: Large Language Models (LLMs) have demonstrated impressive capabilities in solving complex tasks, including those requiring a certain level of reasoning. In this paper, we focus on state tracking, a problem where models need to keep track of the state governing a number of entities. To isolate the state tracking component from other factors, we propose a benchmark based on three well-defined state tracking tasks and analyse the performance of LLMs in different scenarios. The results indicate that the recent generation of LLMs (specifically, GPT-4 and Llama3) are capable of tracking state, especially when integrated with mechanisms such as Chain of Thought. However, models from the former generation, while understanding the task and being able to solve it at the initial stages, often fail at this task after a certain number of steps.

Method: Created LocalBench with 14,782 validated question-answer pairs across 526 US counties, integrating Census statistics, local subreddit discourse, and regional news across physical, cognitive, and relational dimensions of locality. Evaluated 13 state-of-the-art LLMs under closed-book and web-augmented settings.

Result: Critical limitations found: best models achieve only 56.8% accuracy on narrative-style questions and below 15.5% on numerical reasoning. Web augmentation shows inconsistent effects - improves Gemini by +13.6% but reduces GPT-series by -11.4%. Larger model size doesn’t guarantee better performance.

Conclusion: Urgent need for language models that can support equitable, place-aware AI systems capable of engaging with diverse, fine-grained realities of local communities across geographic and cultural contexts.

Abstract: Large language models (LLMs) have been widely evaluated on macro-scale geographic tasks, such as global factual recall, event summarization, and regional reasoning. Yet, their ability to handle hyper-local knowledge remains poorly understood. This gap is increasingly consequential as real-world applications, from civic platforms to community journalism, demand AI systems that can reason about neighborhood-specific dynamics, cultural narratives, and local governance. Existing benchmarks fall short in capturing this complexity, often relying on coarse-grained data or isolated references. We present LocalBench, the first benchmark designed to systematically evaluate LLMs on county-level local knowledge across the United States. Grounded in the Localness Conceptual Framework, LocalBench includes 14,782 validated question-answer pairs across 526 U.S. counties in 49 states, integrating diverse sources such as Census statistics, local subreddit discourse, and regional news. It spans physical, cognitive, and relational dimensions of locality. Using LocalBench, we evaluate 13 state-of-the-art LLMs under both closed-book and web-augmented settings. Our findings reveal critical limitations: even the best-performing models reach only 56.8% accuracy on narrative-style questions and perform below 15.5% on numerical reasoning. Moreover, larger model size and web augmentation do not guarantee better performance, for example, search improves Gemini’s accuracy by +13.6%, but reduces GPT-series performance by -11.4%. These results underscore the urgent need for language models that can support equitable, place-aware AI systems: capable of engaging with the diverse, fine-grained realities of local communities across geographic and cultural contexts.

Method: KPPO introduces three innovations: 1) knowledge gap filling mechanism for identification and remediation, 2) batch-wise candidate evaluation considering performance and distributional stability, 3) adaptive knowledge pruning strategy balancing performance and token efficiency.

Result: Extensive evaluation on 15 knowledge-intensive benchmarks shows KPPO achieves ~6% average performance improvement over strongest baselines while maintaining comparable or lower token consumption, with up to 29% token reduction.

Conclusion: KPPO successfully reformulates prompt optimization as systematic knowledge integration rather than potential elicitation, demonstrating superior performance on knowledge-intensive tasks while maintaining token efficiency.

Abstract: While prompt optimization has emerged as a critical technique for enhancing language model performance, existing approaches primarily focus on elicitation-based strategies that search for optimal prompts to activate models’ capabilities. These methods exhibit fundamental limitations when addressing knowledge-intensive tasks, as they operate within fixed parametric boundaries rather than providing the factual knowledge, terminology precision, and reasoning patterns required in specialized domains. To address these limitations, we propose Knowledge-Provision-based Prompt Optimization (KPPO), a framework that reformulates prompt optimization as systematic knowledge integration rather than potential elicitation. KPPO introduces three key innovations: 1) a knowledge gap filling mechanism for knowledge gap identification and targeted remediation; 2) a batch-wise candidate evaluation approach that considers both performance improvement and distributional stability; 3) an adaptive knowledge pruning strategy that balances performance and token efficiency, reducing up to 29% token usage. Extensive evaluation on 15 knowledge-intensive benchmarks from various domains demonstrates KPPO’s superiority over elicitation-based methods, with an average performance improvement of ~6% over the strongest baseline while achieving comparable or lower token consumption. Code at: https://github.com/xyz9911/KPPO.

Method: Proposes RIFL (Rubric-based Instruction-Following Learning) - a post-training pipeline that uses rubric generation, a finetuned rubric verifier, and reward shaping for reinforcement learning to improve instruction following.

Result: RIFL achieves 6.7% absolute gain on AdvancedIF benchmark and strong results on public benchmarks, with ablation studies confirming the effectiveness of each component.

Conclusion: Rubrics are established as a powerful tool for both training and evaluating advanced instruction following in LLMs, enabling more capable and reliable AI systems.

Abstract: Recent progress in large language models (LLMs) has led to impressive performance on a range of tasks, yet advanced instruction following (IF)-especially for complex, multi-turn, and system-prompted instructions-remains a significant challenge. Rigorous evaluation and effective training for such capabilities are hindered by the lack of high-quality, human-annotated benchmarks and reliable, interpretable reward signals. In this work, we introduce AdvancedIF (we will release this benchmark soon), a comprehensive benchmark featuring over 1,600 prompts and expert-curated rubrics that assess LLMs ability to follow complex, multi-turn, and system-level instructions. We further propose RIFL (Rubric-based Instruction-Following Learning), a novel post-training pipeline that leverages rubric generation, a finetuned rubric verifier, and reward shaping to enable effective reinforcement learning for instruction following. Extensive experiments demonstrate that RIFL substantially improves the instruction-following abilities of LLMs, achieving a 6.7% absolute gain on AdvancedIF and strong results on public benchmarks. Our ablation studies confirm the effectiveness of each component in RIFL. This work establishes rubrics as a powerful tool for both training and evaluating advanced IF in LLMs, paving the way for more capable and reliable AI systems.

Method: LOCA-R (LOgical Chain Augmentation for Reasoning), an improved version of the LOCA framework adapted for complex reasoning.

Result: Achieved 313 out of 320 points on CPhO 2025 theory exam, surpassing highest-scoring human competitor and significantly outperforming all baseline methods.

Conclusion: LOCA-R demonstrates exceptional performance in complex physics problem-solving, establishing new state-of-the-art results in this challenging domain.

Abstract: Olympiad-level physics problem-solving presents a significant challenge for both humans and artificial intelligence (AI), as it requires a sophisticated integration of precise calculation, abstract reasoning, and a fundamental grasp of physical principles. The Chinese Physics Olympiad (CPhO), renowned for its complexity and depth, serves as an ideal and rigorous testbed for these advanced capabilities. In this paper, we introduce LOCA-R (LOgical Chain Augmentation for Reasoning), an improved version of the LOCA framework adapted for complex reasoning, and apply it to the CPhO 2025 theory examination. LOCA-R achieves a near-perfect score of 313 out of 320 points, solidly surpassing the highest-scoring human competitor and significantly outperforming all baseline methods.

Method: Constructed style-augmented jailbreak benchmark by transforming prompts from 3 datasets into 11 distinct linguistic styles using templates and LLM rewrites. Evaluated 16 models and proposed style neutralization preprocessing.

Result: Stylistic reframing increased jailbreak success rates by up to +57 percentage points. Contextualized rewrites outperformed templated variants. Style neutralization significantly reduced success rates.

Conclusion: Linguistic style manipulation represents a systemic and scaling-resistant vulnerability in current LLM safety pipelines that requires specific mitigation strategies.

Abstract: Large Language Models (LLMs) are commonly evaluated for robustness against paraphrased or semantically equivalent jailbreak prompts, yet little attention has been paid to linguistic variation as an attack surface. In this work, we systematically study how linguistic styles such as fear or curiosity can reframe harmful intent and elicit unsafe responses from aligned models. We construct style-augmented jailbreak benchmark by transforming prompts from 3 standard datasets into 11 distinct linguistic styles using handcrafted templates and LLM-based rewrites, while preserving semantic intent. Evaluating 16 open- and close-source instruction-tuned models, we find that stylistic reframing increases jailbreak success rates by up to +57 percentage points. Styles such as fearful, curious and compassionate are most effective and contextualized rewrites outperform templated variants. To mitigate this, we introduce a style neutralization preprocessing step using a secondary LLM to strip manipulative stylistic cues from user inputs, significantly reducing jailbreak success rates. Our findings reveal a systemic and scaling-resistant vulnerability overlooked in current safety pipelines.

Method: Created a benchmark with 75,336 question-answer pairs across diverse categories including user facts, assistant recall, abstention, preferences, temporal changes, and implicit connections. Compared simple full-context approaches with sophisticated RAG-based memory systems like Mem0.

Result: Simple full-context approaches achieve 70-82% accuracy on challenging multi-message evidence cases, while RAG-based memory systems achieve only 30-45% on conversations under 150 interactions. Long context excels for first 30 conversations, remains viable up to 150 conversations, and requires hybrid/RAG approaches beyond that.

Conclusion: The small-corpus advantage of conversational memory—where exhaustive search and complete reranking are feasible—deserves dedicated research attention rather than simply applying general RAG solutions to conversation histories.

Abstract: We introduce a comprehensive benchmark for conversational memory evaluation containing 75,336 question-answer pairs across diverse categories including user facts, assistant recall, abstention, preferences, temporal changes, and implicit connections. While existing benchmarks have advanced the field, our work addresses fundamental challenges in statistical power, data generation consistency, and evaluation flexibility that limit current memory evaluation frameworks. We examine the relationship between conversational memory and retrieval-augmented generation (RAG). While these systems share fundamental architectural patterns–temporal reasoning, implicit extraction, knowledge updates, and graph representations–memory systems have a unique characteristic: they start from zero and grow progressively with each conversation. This characteristic enables naive approaches that would be impractical for traditional RAG. Consistent with recent findings on long context effectiveness, we observe that simple full-context approaches achieve 70-82% accuracy even on our most challenging multi-message evidence cases, while sophisticated RAG-based memory systems like Mem0 achieve only 30-45% when operating on conversation histories under 150 interactions. Our analysis reveals practical transition points: long context excels for the first 30 conversations, remains viable with manageable trade-offs up to 150 conversations, and typically requires hybrid or RAG approaches beyond that point as costs and latencies become prohibitive. These patterns indicate that the small-corpus advantage of conversational memory–where exhaustive search and complete reranking are feasible–deserves dedicated research attention rather than simply applying general RAG solutions to conversation histories.

Method: Used computational analysis on CONLIT dataset with Welch’s ANOVA to compare narrative features by author gender and genre, logistic regression to model literary classification effects, and analyzed stylistic/semantic vector representations.

Result: Found statistically significant formal markers for each literary category and demonstrated that female authorship narrows and blurs the criteria for achieving literary status.

Conclusion: Genre functions as both a formal and institutional designation, with author gender playing a significant moderating role in how works are classified as literary fiction.

Abstract: Though the concept of genre has been a subject of discussion for millennia, the relatively recent emergence of genre fiction has added a new layer to this ongoing conversation. While more traditional perspectives on genre have emphasized form, contemporary scholarship has invoked both formal and institutional characteristics in its taxonomy of genre, genre fiction, and literary fiction. This project uses computational methods to explore the soundness of genre as a formal designation as opposed to an institutional one. Pulling from Andrew Piper’s CONLIT dataset of Contemporary Literature, we assemble a corpus of literary and genre fiction, with the latter category containing romance, mystery, and science fiction novels. We use Welch’s ANOVA to compare the distribution of narrative features according to author gender within each genre and within genre versus literary fiction. Then, we use logistic regression to model the effect that each feature has on literary classification and to measure how author gender moderates these effects. Finally, we analyze stylistic and semantic vector representations of our genre categories to understand the importance of form and content in literary classification. This project finds statistically significant formal markers of each literary category and illustrates how female authorship narrows and blurs the target for achieving literary status.

Method: URaG introduces a lightweight cross-modal retrieval module that converts early Transformer layers into evidence selectors, identifying relevant pages while discarding irrelevant content, allowing deeper layers to focus on pertinent information.

Result: Extensive experiments show URaG achieves state-of-the-art performance while reducing computational overhead by 44-56%.

Conclusion: MLLMs have inherent evidence localization capabilities that can be leveraged for efficient retrieval during reasoning, enabling unified retrieval-generation for long document understanding.

Abstract: Recent multimodal large language models (MLLMs) still struggle with long document understanding due to two fundamental challenges: information interference from abundant irrelevant content, and the quadratic computational cost of Transformer-based architectures. Existing approaches primarily fall into two categories: token compression, which sacrifices fine-grained details; and introducing external retrievers, which increase system complexity and prevent end-to-end optimization. To address these issues, we conduct an in-depth analysis and observe that MLLMs exhibit a human-like coarse-to-fine reasoning pattern: early Transformer layers attend broadly across the document, while deeper layers focus on relevant evidence pages. Motivated by this insight, we posit that the inherent evidence localization capabilities of MLLMs can be explicitly leveraged to perform retrieval during the reasoning process, facilitating efficient long document understanding. To this end, we propose URaG, a simple-yet-effective framework that Unifies Retrieval and Generation within a single MLLM. URaG introduces a lightweight cross-modal retrieval module that converts the early Transformer layers into an efficient evidence selector, identifying and preserving the most relevant pages while discarding irrelevant content. This design enables the deeper layers to concentrate computational resources on pertinent information, improving both accuracy and efficiency. Extensive experiments demonstrate that URaG achieves state-of-the-art performance while reducing computational overhead by 44-56%. The code is available at https://github.com/shi-yx/URaG.

Method: Integrates DeBERTa’s enhanced attention mechanism with LSTM in a dual-channel structure to process both meaning and grammatical patterns, with refined component interaction.

Result: Achieved F1-score improvements of 4.85, 8.36, and 2.42 on standard datasets, with DeBERTa’s attention system helping handle complex sentence structures and distant word relationships.

Conclusion: Upgrading to more advanced language models with thoughtful integration can lead to real improvements in sentiment analysis performance.

Abstract: Fine-grained sentiment analysis faces ongoing challenges in Aspect Sentiment Triple Extraction (ASTE), particularly in accurately capturing the relationships between aspects, opinions, and sentiment polarities. While researchers have made progress using BERT and Graph Neural Networks, the full potential of advanced language models in understanding complex language patterns remains unexplored. We introduce DESS, a new approach that builds upon previous work by integrating DeBERTa’s enhanced attention mechanism to better understand context and relationships in text. Our framework maintains a dual-channel structure, where DeBERTa works alongside an LSTM channel to process both meaning and grammatical patterns in text. We have carefully refined how these components work together, paying special attention to how different types of language information interact. When we tested DESS on standard datasets, it showed meaningful improvements over current methods, with F1-score increases of 4.85, 8.36, and 2.42 in identifying aspect opinion pairs and determining sentiment accurately. Looking deeper into the results, we found that DeBERTa’s sophisticated attention system helps DESS handle complicated sentence structures better, especially when important words are far apart. Our findings suggest that upgrading to more advanced language models when thoughtfully integrated, can lead to real improvements in how well we can analyze sentiments in text. The implementation of our approach is publicly available at: https://github.com/VishalRepos/DESS.

Method: Systematically evaluated three prompting paradigms: one-shot approach, ReAct framework, and multi-step agentic workflow using MedGemma language model.

Result: Simple one-shot prompting achieved highest performance on official validation set, while complex reasoning frameworks introduced noise through “overthinking”.

Conclusion: Direct approaches are more robust and efficient than complex reasoning chains for clinical information extraction from manually annotated transcripts.

Abstract: The accurate extraction of medical orders from doctor-patient conversations is a critical task for reducing clinical documentation burdens and ensuring patient safety. This paper details our team submission to the MEDIQA-OE-2025 Shared Task. We investigate the performance of MedGemma, a new domain-specific open-source language model, for structured order extraction. We systematically evaluate three distinct prompting paradigms: a straightforward one-Shot approach, a reasoning-focused ReAct framework, and a multi-step agentic workflow. Our experiments reveal that while more complex frameworks like ReAct and agentic flows are powerful, the simpler one-shot prompting method achieved the highest performance on the official validation set. We posit that on manually annotated transcripts, complex reasoning chains can lead to “overthinking” and introduce noise, making a direct approach more robust and efficient. Our work provides valuable insights into selecting appropriate prompting strategies for clinical information extraction in varied data conditions.

Method: 1) Mined prompting: retrieve top-k similar examples as few-shot demonstrations. 2) Metadata-guided generation: predict and incorporate four key metadata attributes based on confidence levels.

Result: Mined prompting improves response relevance, while metadata-guided generation further enhances clinical precision in wound care responses.

Conclusion: These complementary methods show promise for developing reliable AI tools to support efficient wound care management in remote settings.

Abstract: The rapid expansion of asynchronous remote care has intensified provider workload, creating demand for AI systems that can assist clinicians in managing patient queries more efficiently. The MEDIQA-WV 2025 shared task addresses this challenge by focusing on generating free-text responses to wound care queries paired with images. In this work, we present two complementary approaches developed for the English track. The first leverages a mined prompting strategy, where training data is embedded and the top-k most similar examples are retrieved to serve as few-shot demonstrations during generation. The second approach builds on a metadata ablation study, which identified four metadata attributes that consistently enhance response quality. We train classifiers to predict these attributes for test cases and incorporate them into the generation pipeline, dynamically adjusting outputs based on prediction confidence. Experimental results demonstrate that mined prompting improves response relevance, while metadata-guided generation further refines clinical precision. Together, these methods highlight promising directions for developing AI-driven tools that can provide reliable and efficient wound care support.

Method: Introduce encoder-augmented causal decoder model architectures that combine encoder and decoder components for superior training efficiency and compression performance compared to causal transformers.

Result: Achieved higher compression than causal transformers even on modest hardware, demonstrated per-token entropy estimation, and showed that models trained to approach estimated entropies exhibit greater generalization.

Conclusion: Training language models to approach but not exceed estimated per-token entropies leads to better generalization than standard loss minimization approaches, providing a principled method for model optimization.

Abstract: Language prediction is constrained by informational entropy intrinsic to language, such that there exists a limit to how accurate any language model can become and equivalently a lower bound to language compression. The most efficient language compression algorithms today are causal (next token prediction) large language models, but the use of these models to form accurate estimates of language entropy is currently computationally infeasible. We introduce encoder-augmented causal decoder model architectures that exhibit superior training efficiency characteristics and achieve higher compression than causal transformers even when trained on modest hardware. We demonstrate how entropy estimates can be obtained on a per-token basis, and show that the generalization of models trained to approach the entropy of their training data necessarily exceeds the generalization of models trained to minimize loss beyond this value. We show empirically that causal models trained to approach but not exceed estimated per-token entropies exhibit greater generalization than models trained without taking entropy into account.

Method: SSR decomposes model responses into verifiable (sub-question, sub-answer) pairs, performs step-level confidence estimation through controlled re-solving and self-consistency checks, and iteratively refines unreliable steps.

Result: Empirical results across five reasoning benchmarks and three LLMs show that SSR consistently outperforms state-of-the-art iterative self-refinement baselines.

Conclusion: SSR provides both performance improvements and a principled black-box approach for evaluating and understanding the internal reasoning processes of LLMs.

Abstract: Large Language Models (LLMs) have demonstrated remarkable reasoning abilities, yet existing test-time frameworks often rely on coarse self-verification and self-correction, limiting their effectiveness on complex tasks. In this paper, we propose Socratic Self-Refine (SSR), a novel framework for fine-grained evaluation and precise refinement of LLM reasoning. Our proposed SSR decomposes model responses into verifiable (sub-question, sub-answer) pairs, enabling step-level confidence estimation through controlled re-solving and self-consistency checks. By pinpointing unreliable steps and iteratively refining them, SSR produces more accurate and interpretable reasoning chains. Empirical results across five reasoning benchmarks and three LLMs show that SSR consistently outperforms state-of-the-art iterative self-refinement baselines. Beyond performance gains, SSR provides a principled black-box approach for evaluating and understanding the internal reasoning processes of LLMs. Code is available at https://github.com/SalesforceAIResearch/socratic-self-refine-reasoning.

Method: Large-scale pre-training on open data using AMD Instinct MI300X GPUs, followed by general-purpose instruction tuning and human preference alignment. Also created specialized variants: Instella-Long (128K context) and Instella-Math (reasoning-focused with SFT and RL).

Result: Achieves state-of-the-art results among fully open models despite using fewer pre-training tokens than contemporaries, and is competitive with leading open-weight models of comparable size.

Conclusion: Instella establishes a transparent, performant, and versatile alternative for the community, advancing open and reproducible language modeling research.

Abstract: Large language models (LLMs) have demonstrated remarkable performance across a wide range of tasks, yet the majority of high-performing models remain closed-source or partially open, limiting transparency and reproducibility. In this work, we introduce Instella, a family of fully open three billion parameter language models trained entirely on openly available data and codebase. Powered by AMD Instinct MI300X GPUs, Instella is developed through large-scale pre-training, general-purpose instruction tuning, and alignment with human preferences. Despite using substantially fewer pre-training tokens than many contemporaries, Instella achieves state-of-the-art results among fully open models and is competitive with leading open-weight models of comparable size. We further release two specialized variants: Instella-Long, capable of handling context lengths up to 128K tokens, and Instella-Math, a reasoning-focused model enhanced through supervised fine-tuning and reinforcement learning on mathematical tasks. Together, these contributions establish Instella as a transparent, performant, and versatile alternative for the community, advancing the goal of open and reproducible language modeling research.

Method: Generative Adversarial Distillation (GAD) frames student LLM as generator and trains discriminator to distinguish student responses from teacher responses, creating a minimax game where discriminator provides on-policy reward feedback.

Result: GAD consistently outperforms sequence-level knowledge distillation. Qwen2.5-14B-Instruct trained with GAD becomes comparable to GPT-5-Chat on LMSYS-Chat evaluation.

Conclusion: GAD establishes itself as a promising and effective paradigm for black-box LLM distillation, enabling student models to achieve teacher-level performance without access to internal teacher parameters.

Abstract: Black-box distillation creates student large language models (LLMs) by learning from a proprietary teacher model’s text outputs alone, without access to its internal logits or parameters. In this work, we introduce Generative Adversarial Distillation (GAD), which enables on-policy and black-box distillation. GAD frames the student LLM as a generator and trains a discriminator to distinguish its responses from the teacher LLM’s, creating a minimax game. The discriminator acts as an on-policy reward model that co-evolves with the student, providing stable, adaptive feedback. Experimental results show that GAD consistently surpasses the commonly used sequence-level knowledge distillation. In particular, Qwen2.5-14B-Instruct (student) trained with GAD becomes comparable to its teacher, GPT-5-Chat, on the LMSYS-Chat automatic evaluation. The results establish GAD as a promising and effective paradigm for black-box LLM distillation.

Method: Combines hardware-efficient independent Givens rotations with channel-wise scaling to equalize magnitude across channels and narrow dynamic range within quantization groups, with co-designed inference kernels for GPU parallelism.

Result: Achieves average 2.4% accuracy improvement over AWQ on reasoning tasks with less than 10% inference overhead.

Conclusion: ParoQuant enables more efficient and accurate deployment of reasoning LLMs by effectively handling outliers while maintaining low runtime costs.

Abstract: Weight-only post-training quantization (PTQ) compresses the weights of Large Language Models (LLMs) into low-precision representations to reduce memory footprint and accelerate inference. However, the presence of outliers in weights and activations often leads to large quantization errors and severe accuracy degradation, especially in recent reasoning LLMs where errors accumulate across long chains of thought. Existing PTQ methods either fail to sufficiently suppress outliers or introduce significant overhead during inference. In this paper, we propose Pairwise Rotation Quantization (ParoQuant), a weight-only PTQ method that combines hardware-efficient and optimizable independent Givens rotations with channel-wise scaling to even out the magnitude across channels and narrow the dynamic range within each quantization group. We further co-design the inference kernel to fully exploit GPU parallelism and keep the rotations and scaling lightweight at runtime. ParoQuant achieves an average 2.4% accuracy improvement over AWQ on reasoning tasks with less than 10% overhead. This paves the way for more efficient and accurate deployment of reasoning LLMs.

Method: Proposes a dual-knowledge infusion framework with MKGD and EXKR, decomposing proofreading into three stages: detection, localization, and correction, mirroring expert radiologists’ systematic review process.

Result: Substantial improvements: up to 31.56% increase in error detection accuracy and 37.4% reduction in processing time across multiple LLM architectures. Human evaluation confirms superior clinical relevance and factual consistency.

Conclusion: The framework effectively enhances LLMs’ capabilities for radiology report proofreading through systematic medical knowledge integration, providing an automated approach that matches expert radiologists’ precision and interpretability.

Abstract: The increasing complexity and workload of clinical radiology leads to inevitable oversights and mistakes in their use as diagnostic tools, causing delayed treatments and sometimes life-threatening harm to patients. While large language models (LLMs) have shown remarkable progress in many tasks, their utilities in detecting and correcting errors in radiology reporting are limited. This paper proposes a novel dual-knowledge infusion framework that enhances LLMs’ capability for radiology report proofreading through systematic integration of medical expertise. Specifically, the knowledge infusion combines medical knowledge graph distillation (MKGD) with external knowledge retrieval (EXKR), enabling an effective automated approach in tackling mistakes in radiology reporting. By decomposing the complex proofreading task into three specialized stages of detection, localization, and correction, our method mirrors the systematic review process employed by expert radiologists, ensuring both precision and clinical interpretability. To perform a robust, clinically relevant evaluation, a comprehensive benchmark is also proposed using real-world radiology reports with real-world error patterns, including speech recognition confusions, terminology ambiguities, and template-related inconsistencies. Extensive evaluations across multiple LLM architectures demonstrate substantial improvements of our approach: up to 31.56% increase in error detection accuracy and 37.4% reduction in processing time. Human evaluation by radiologists confirms superior clinical relevance and factual consistency compared to existing approaches.

Method: Used human-authored essays as benchmark, prompted ChatGPT to generate equivalent-length essays, then analyzed both using Open Brain AI tool to extract phonological, morphological, syntactic, and lexical features.

Result: Found significant differences across multiple linguistic features including consonants, word stress, nouns, verbs, pronouns, direct objects, prepositional modifiers, and use of difficult words.

Conclusion: Automated tools are valuable for efficient language assessment, and enhanced training methodologies are needed to improve AI’s capacity for producing more human-like text.

Abstract: While extensive research has focused on ChatGPT in recent years, very few studies have systematically quantified and compared linguistic features between human-written and Artificial Intelligence (AI)-generated language. This study aims to investigate how various linguistic components are represented in both types of texts, assessing the ability of AI to emulate human writing. Using human-authored essays as a benchmark, we prompted ChatGPT to generate essays of equivalent length. These texts were analyzed using Open Brain AI, an online computational tool, to extract measures of phonological, morphological, syntactic, and lexical constituents. Despite AI-generated texts appearing to mimic human speech, the results revealed significant differences across multiple linguistic features such as consonants, word stress, nouns, verbs, pronouns, direct objects, prepositional modifiers, and use of difficult words among others. These findings underscore the importance of integrating automated tools for efficient language assessment, reducing time and effort in data analysis. Moreover, they emphasize the necessity for enhanced training methodologies to improve the capacity of AI for producing more human-like text.

Method: Interactive-T2S framework with four general tools for proactive information retrieval by LLMs, using direct database interactions and detailed exemplars for step-wise reasoning.

Result: Achieved advanced performance on Spider and BIRD datasets and their variants, obtaining state-of-the-art results on BIRD leaderboard without oracle knowledge.

Conclusion: The framework effectively addresses current limitations in text-to-SQL parsing by leveraging LLM reasoning capabilities through interactive database interactions.

Abstract: This study explores text-to-SQL parsing by leveraging the powerful reasoning capabilities of large language models (LLMs). Despite recent advancements, existing LLM-based methods are still inefficient and struggle to handle cases with wide tables effectively. Furthermore, current interaction-based approaches either lack a step-by-step, interpretable SQL generation process or fail to provide a universally applicable interaction design. To address these challenges, we introduce Interactive-T2S, a framework that generates SQL queries through direct interactions with databases. This framework includes four general tools that facilitate proactive and efficient information retrieval by the LLM. Additionally, we have developed detailed exemplars to demonstrate the step-wise reasoning processes within our framework. Our approach achieves advanced performance on the Spider and BIRD datasets as well as their variants. Notably, we obtain state-of-the-art results on the BIRD leaderboard under the setting without oracle knowledge, demonstrating the effectiveness of our method.

Method: Participant observation, analysis of internal documentation, and collaborative analysis across two Horizon 2020 projects conducted with a European deaf-led NGO.

Result: Identified five key lessons: recognize invisible labor, manage expectations through accessible communication, dismantle ableism, diversify methods to address fatigue, and redistribute power through deaf leadership.

Conclusion: Co-creation requires structural changes beyond token participation, with actionable implications for participatory AI involving minoritized language and disability communities.

Abstract: In the era of AI-driven language technologies, the participation of deaf communities in sign language technology development, often framed as co-creation, is increasingly emphasized. We present a reflexive case study of two Horizon 2020 projects on sign language machine translation (2021- 2023), conducted with a EUD, a European-level deaf-led NGO. Using participant observation, internal documentation, and collaborative analysis among the authors, we interrogate co-creation as both a practice and a discourse. We offer five lessons for making co-creation consequential: 1) recognise and resource deaf partners invisible labor, 2) manage expectations via accessible science communication, 3) crip co-creation by dismantling structural ableism, 4) diversify participatory methods to address co-creation fatigue and intersectionality, and 5) redistribute power through deaf leadership. We contribute an empirically grounded account of how co-creation plays out in multi-partner AI projects, and actionable implications for design that extend to participatory AI with minoritized language and disability communities.

Method: Parsimonious adaptation of phi-3-mini with pipeline additions including chain of thought, ensembling, and fine-tuning, evaluated on MultiMedQA and MedBullets benchmarks.

Result: Achieves 75.7% on MedQA (USMLE), surpassing physician passing mark (~60%) and setting SOTA for models under 5B parameters. Unexpectedly, retrieval augmented generation showed no significant improvements.

Conclusion: MedMobile democratizes access to medical language models with lower compute needs and fast inference, representing a critical step forward for clinically relevant models.

Abstract: Language models (LMs) have demonstrated expert-level reasoning and recall abilities in medicine. However, computational costs and privacy concerns are mounting barriers to wide-scale implementation. To address these significant limitations, we introduce a parsimonious adaptation of phi-3-mini, MedMobile, a 3.8 billion parameter LM capable of running on a mobile device, for medical applications. We perform a careful set of pipeline additions and demonstrate that chain of thought, ensembling, and fine-tuning lead to the greatest performance gains, while unexpectedly retrieval augmented generation fails to demonstrate significant improvements. We evaluate the efficiency of our pipeline on the MultiMedQA and MedBullets. We demonstrate that MedMobile scores 75.7% on the MedQA (USMLE), surpassing the passing mark for licensed physicians (~60%) and rivaling scores of models 100 times its size. Across the entirety of the MultiMedQA, MedMobile achieves SOTA performance for models with less than 5B parameters and represents the smallest model to pass the MedQA (USMLE). MedMobile holds promise to democratize access to language models in medicine, bolstering lower compute needs and fast inference speeds. With the ability to combat the biggest barriers to entry for language models in medicine, we hope that MedMobile is a critical step forward in developing clinically relevant language models.

Method: Develop MMECInstruct dataset and CASLIE framework, then fine-tune multimodal foundation models using the dataset within the framework.

Result: CASLIE models substantially outperform 5 categories of advanced baseline models in in-domain evaluation and show strong generalizability to out-of-domain settings.

Conclusion: MMECInstruct and CASLIE provide effective solutions for multimodal e-commerce challenges, with publicly available resources for further research.

Abstract: Leveraging multimodal data to drive breakthroughs in e-commerce applications through Multimodal Foundation Models (MFMs) is gaining increasing attention from the research community. However, there are significant challenges that hinder the optimal use of multimodal e-commerce data by foundation models: (1) the scarcity of large-scale, high-quality multimodal benchmark datasets; and (2) the lack of effective multimodal information integration methods. To address these challenges, in this paper, we introduce MMECInstruct, the first-ever, large-scale, and high-quality multimodal instruction dataset for e-commerce. We also develop CASLIE, a simple, lightweight, yet effective framework for integrating multimodal information for e-commerce. Leveraging MMECInstruct, we fine-tune a series of e-commerce MFMs within CASLIE, denoted as CASLIE models. Our comprehensive evaluation demonstrates that CASLIE models substantially outperform 5 categories of advanced baseline models in the in-domain evaluation. Moreover, CASLIE models show strong generalizability to out-of-domain settings. MMECInstruct and CASLIE models are publicly accessible through https://ninglab.github.io/CASLIE/.

Method: The study conducts empirical evaluation of multiple scoping techniques: prompting, supervised fine-tuning, preference learning, and Circuit Breakers. Experiments span three LLM families, multiple topics, fine-grained topics, with ablation studies on query diversity, technique layering, and adversarial evaluations.

Result: Results show scoping is achievable. Supervised fine-tuning works best with diverse irrelevant query examples, while Circuit Breakers perform well with low diversity. Layering techniques provides benefits of both methods. The approach works across different model families and task types.

Conclusion: The study provides a practitioner’s guide to scoping LLMs, demonstrating that appropriate method selection depends on available data diversity, and that combining techniques can yield optimal performance for ensuring LLMs stay within their intended scope.

Abstract: Large language models (LLMs) are deployed in a wide variety of user-facing applications. Typically, these deployments have some specific purpose, like answering questions grounded on documentation or acting as coding assistants, but they require general language understanding. In such deployments, LLMs should respond only to queries that align with the intended purpose and reject all other requests, such as generating poetry or answering questions about physics, a task we refer to as `scoping’. We conduct a comprehensive empirical evaluation of various methods, ranging from prompting, fine-tuning to preference learning and the recently proposed general alignment technique known as Circuit Breakers (CB). Across three families of language models and a broad variety of tasks, we show that it is possible to scope language models. We examine scoping for multiple topics, and fine-grained topics. We ablate diversity of irrelevant queries, layer different techniques, conduct adversarial evaluations and more. Among other results, we find that when diverse examples of irrelevant queries are available, simple supervised fine-tuning produces the best results, but when such diversity is low, Circuit Breakers perform quite well. One can often get the benefits of both methods by layering them in succession. We intend our study to serve as a practitioner’s guide to scoping LLMs.

Method: Compare actual human players’ guesses with near-optimal guesses using NLP techniques to analyze semantic, orthographic, and phonological influences.

Result: Human gameplay in Wordle is significantly influenced by the semantics, orthography, and phonology of previous guesses, showing systematic deviations from optimal strategies.

Conclusion: Wordle provides a unique constrained environment to study human language processing, revealing how semantic, orthographic, and phonological factors shape decision-making in word games.

Abstract: We show that human players’ gameplay in the game of Wordle is influenced by the semantics, orthography, and phonology of the player’s previous guesses. We compare actual human players’ guesses with near-optimal guesses using NLP techniques. We study human language use in the constrained environment of Wordle, which is situated between natural language use and the artificial word association task

Method: Hybrid framework combining CNN and BiLSTM with attention mechanism, fine-tuned with early stopping, and integrated with SHAP for explainable AI.

Result: Achieved 92.81% accuracy initially, improved to 94.29% after fine-tuning. SHAP analysis identified key features like mental health-related terms.

Conclusion: Combining powerful ML methods with explainability creates reliable mental health monitoring systems, enhancing credibility and trust for professionals.

Abstract: Suicidal ideation detection is crucial for preventing suicides, a leading cause of death worldwide. Many individuals express suicidal thoughts on social media, offering a vital opportunity for early detection through advanced machine learning techniques. The identification of suicidal ideation in social media text is improved by utilising a hybrid framework that integrates Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM), enhanced with an attention mechanism. To enhance the interpretability of the model’s predictions, Explainable AI (XAI) methods are applied, with a particular focus on SHapley Additive exPlanations (SHAP), are incorporated. At first, the model managed to reach an accuracy of 92.81%. By applying fine-tuning and early stopping techniques, the accuracy improved to 94.29%. The SHAP analysis revealed key features influencing the model’s predictions, such as terms related to mental health struggles. This level of transparency boosts the model’s credibility while helping mental health professionals understand and trust the predictions. This work highlights the potential for improving the accuracy and interpretability of detecting suicidal tendencies, making a valuable contribution to the progress of mental health monitoring systems. It emphasizes the significance of blending powerful machine learning methods with explainability to develop reliable and impactful mental health solutions.

Method: Three-stage framework: (1) MiniMax-driven training set construction with turn-level LLM feedback, (2) post-training attackers with SFT and DPO, and (3) interactions between attacking and target LLMs.

Result: Achieved 90% ASR with LLaMA-3-8B vs Gemini-1.5-Pro and 70% ASR with Mistral-7B vs GPT-4o, outperforming single-turn baselines and matching GPT-4o-based multi-turn performance with smaller models.

Conclusion: Siren demonstrates the effectiveness of learning-based multi-turn attacks and highlights the need for stronger defenses against realistic adversarial strategies.

Abstract: Large language models (LLMs) are widely used in real-world applications, raising concerns about their safety and trustworthiness. While red-teaming with jailbreak prompts exposes the vulnerabilities of LLMs, current efforts focus primarily on single-turn attacks, overlooking the multi-turn strategies used by real-world adversaries. Existing multi-turn methods rely on static patterns or predefined logical chains, failing to account for the dynamic strategies during attacks. We propose Siren, a learning-based multi-turn attack framework designed to simulate real-world human jailbreak behaviors. Siren consists of three stages: (1) MiniMax-driven training set construction utilizing Turn-Level LLM feedback, (2) post-training attackers with supervised fine-tuning (SFT) and direct preference optimization (DPO), and (3) interactions between the attacking and target LLMs. Experiments demonstrate that Siren achieves an attack success rate (ASR) of 90% with LLaMA-3-8B as the attacker against Gemini-1.5-Pro as the target model, and 70% with Mistral-7B against GPT-4o, significantly outperforming single-turn baselines. Moreover, Siren with a 7B-scale model achieves performance comparable to a multi-turn baseline that leverages GPT-4o as the attacker, while requiring fewer turns and employing decomposition strategies that are better semantically aligned with attack goals. We hope Siren inspires the development of stronger defenses against advanced multi-turn jailbreak attacks under realistic scenarios. Code is available at https://github.com/YiyiyiZhao/siren. Warning: This paper contains potentially harmful text.

Method: Chain-of-Associated-Thoughts (CoAT) integrates Monte Carlo Tree Search for structured exploration with associative memory for dynamic knowledge integration, enabling real-time knowledge updates and reasoning pathway refinement.

Result: Achieved over 10% improvement on HotpotQA and MuSiQue datasets, and over 15% gain on proprietary CRB dataset, demonstrating enhanced reasoning capabilities.

Conclusion: CoAT’s synergy between MCTS and associative memory effectively expands LLM search space and enables adaptive knowledge integration, making it a promising approach for complex reasoning tasks.

Abstract: Research on LLM technologies is rapidly emerging, with most of them employ a ‘fast thinking’ approach to inference. Most LLMs generate the final result based solely on a single query and LLM’s reasoning capabilities. However, with the advent of OpenAI-o1, ‘slow thinking’ techniques have garnered increasing attention because its process is closer to the human thought process. Inspired by the human ability to constantly associate and replenish knowledge during thinking, we developed the novel Chain-of-Associated-Thoughts (CoAT) framework, which introduces an innovative synergy between the Monte Carlo Tree Search (MCTS) algorithm and a dynamic mechanism for integrating new key information, termed ‘associative memory’. By combining the structured exploration capabilities of MCTS with the adaptive learning capacity of associative memory, CoAT significantly expands the LLM search space, enabling our framework to explore diverse reasoning pathways and dynamically update its knowledge base in real-time. This allows the framework to not only revisit and refine earlier inferences but also adaptively incorporate evolving information, ensuring that the final output is both accurate and comprehensive. We validate CoAT’s effectiveness across a variety of generative and reasoning tasks. Quantitative experiments show that CoAT achieves over 10% performance improvement on open-source multi-hop reasoning datasets (HotpotQA, MuSiQue) and more than 15% gain on our proprietary CRB dataset.

Method: Developed MMTEB as a community-driven expansion of MTEB, introduced novel downsampling based on inter-task correlation, optimized tasks with hard negative sampling, and created efficient benchmark splits.

Result: Found that multilingual-e5-large-instruct (560M parameters) outperforms larger LLMs; created optimized benchmarks that maintain ranking accuracy while reducing computational costs significantly.

Conclusion: MMTEB provides comprehensive multilingual embedding evaluation, demonstrates efficient smaller models can outperform larger ones, and offers computationally optimized benchmarks for practical use.

Abstract: Text embeddings are typically evaluated on a limited set of tasks, which are constrained by language, domain, and task diversity. To address these limitations and provide a more comprehensive evaluation, we introduce the Massive Multilingual Text Embedding Benchmark (MMTEB) - a large-scale, community-driven expansion of MTEB, covering over 500 quality-controlled evaluation tasks across 250+ languages. MMTEB includes a diverse set of challenging, novel tasks such as instruction following, long-document retrieval, and code retrieval, representing the largest multilingual collection of evaluation tasks for embedding models to date. Using this collection, we develop several highly multilingual benchmarks, which we use to evaluate a representative set of models. We find that while large language models (LLMs) with billions of parameters can achieve state-of-the-art performance on certain language subsets and task categories, the best-performing publicly available model is multilingual-e5-large-instruct with only 560 million parameters. To facilitate accessibility and reduce computational cost, we introduce a novel downsampling method based on inter-task correlation, ensuring a diverse selection while preserving relative model rankings. Furthermore, we optimize tasks such as retrieval by sampling hard negatives, creating smaller but effective splits. These optimizations allow us to introduce benchmarks that drastically reduce computational demands. For instance, our newly introduced zero-shot English benchmark maintains a ranking order similar to the full-scale version but at a fraction of the computational cost.

Method: Decomposes responses into atomic units, retrieves relevant context from external knowledge sources, and models logical relationships (entailment, contradiction) using probabilistic encodings to estimate posterior probability of factual support.

Result: Outperforms state-of-the-art prompt-based methods in factual precision and recall on benchmark datasets.

Conclusion: FactReasoner provides an effective neuro-symbolic approach for factuality assessment that enhances the reliability of LLM-generated content.

Abstract: Large language models (LLMs) have achieved remarkable success in generative tasks, yet they often fall short in ensuring the factual accuracy of their outputs, thus limiting their reliability in real-world applications where correctness is critical. In this paper, we present FactReasoner, a novel neuro-symbolic based factuality assessment framework that employs probabilistic reasoning to evaluate the truthfulness of long-form generated responses. FactReasoner decomposes a response into atomic units, retrieves relevant contextual information from external knowledge sources, and models the logical relationships (e.g., entailment, contradiction) between these units and their contexts using probabilistic encodings. It then estimates the posterior probability that each atomic unit is supported by the retrieved evidence. Our experiments on both labeled and unlabeled benchmark datasets demonstrate that FactReasoner often outperforms state-of-the-art prompt-based methods in terms of factual precision and recall. Our open-source implementation is publicly available at: https://github.com/IBM/FactReasoner.

Method: Proposed extension of SAREF4ENER ontology that integrates complete FlexOffer model support, including advanced devices and uncertainty handling, while maintaining backward compatibility.

Result: Novel ontology module that can accurately describe flexibility for advanced devices (EVs, batteries, heat pumps) and capture uncertainty in flexible loads.

Conclusion: The extended ontology enables standardized data exchange for energy flexibility applications, supporting advanced use cases while maintaining compatibility with existing SAREF4ENER implementations.

Abstract: A key element to support the increased amounts of renewable energy in the energy system is flexibility, i.e., the possibility of changing energy loads in time and amount. Many flexibility models have been designed; however, exact models fail to scale for long time horizons or many devices. Because of this, the FlexOffers model has been designed, to provide device-independent approximations of flexibility with good accuracy, and much better scaling for long time horizons and many devices. An important aspect of the real-life implementation of energy flexibility is enabling flexible data exchange with many smart energy appliances and market systems, e.g., in smart buildings. For this, ontologies standardizing data formats are required. However, the current industry standard ontology for integrating smart devices for energy purposes, SAREF for Energy Flexibility (SAREF4ENER), only has limited support for flexibility and thus cannot support important use cases. In this paper, we propose an extension of SAREF4ENER that integrates full support for the complete FlexOffer model, including advanced use cases, while maintaining backward compatibility. This novel ontology module can accurately describe flexibility for advanced devices such as electric vehicles, batteries, and heat pumps. It can also capture the inherent uncertainty associated with many flexible load types.

Method: Introduces LSKExtractor to benchmark language-specific knowledge in language models and exploit it during inference. Uses three datasets containing cultural and social behavioral knowledge.

Result: LSKExtractor achieves up to 10% relative improvement across datasets and is competitive against strong baselines while being feasible for real-world deployment.

Conclusion: The research contributes to developing more inclusive language models that are better aligned with cultural and linguistic contexts, with open-source implementation available.

Abstract: Often, multilingual language models are trained with the objective to map semantically similar content (in different languages) in the same latent space. In this paper, we show a nuance in this training objective, and find that by changing the language of the input query, we can improve the question answering ability of language models. Our contributions are two-fold. First, we introduce the term Language Specific Knowledge (LSK) to denote queries that are best answered in an “expert language” for a given LLM, thereby enhancing its question-answering ability. We introduce the problem of language selection – for some queries, language models can perform better when queried in languages other than English, sometimes even better in low-resource languages – and the goal is to select the optimal language for the query. Second, we introduce simple to strong baselines to test this problem. Additionally, as a first-pass solution to this novel problem, we design LSKExtractor to benchmark the language-specific knowledge present in a language model and then exploit it during inference. To test our framework, we employ three datasets that contain knowledge about both cultural and social behavioral norms. Overall, LSKExtractor achieves up to 10% relative improvement across datasets, and is competitive against strong baselines, while being feasible in real-world settings. Broadly, our research contributes to the open-source development (https://github.com/agarwalishika/LSKExtractor/tree/main) of language models that are inclusive and more aligned with the cultural and linguistic contexts in which they are deployed.

Method: Two-stage chunk-level compression: segments Long-CoT into manageable chunks, applies LLM-driven inner-chunk compression, and uses inter-chunk search to select short and coherent sequences.

Result: On MATH500, achieves 92.4% accuracy (only 0.6% drop from baseline) while reducing token usage by about 20%. Similar improvements on AIME24 and GPQA-Diamond datasets.

Conclusion: R1-Compress effectively reduces computational costs of Long-CoT reasoning while maintaining comparable reasoning performance, providing a practical solution for efficient step-by-step problem-solving.

Abstract: Chain-of-Thought (CoT) reasoning enhances large language models (LLMs) by enabling step-by-step problem-solving, yet its extension to Long-CoT introduces substantial computational overhead due to increased token length. Existing compression approaches – instance-level and token-level – either sacrifice essential local reasoning signals like reflection or yield incoherent outputs. To address these limitations, we propose R1-Compress, a two-stage chunk-level compression framework that preserves both local information and coherence. Our method segments Long-CoT into manageable chunks, applies LLM-driven inner-chunk compression, and employs an inter-chunk search mechanism to select the short and coherent sequence. Experiments on Qwen2.5-Instruct models across MATH500, AIME24, and GPQA-Diamond demonstrate that R1-Compress significantly reduces token usage while maintaining comparable reasoning accuracy. On MATH500, R1-Compress achieves an accuracy of 92.4%, with only a 0.6% drop compared to the Long-CoT baseline, while reducing token usage by about 20%. Source code will be available at https://github.com/w-yibo/R1-Compress

Method: Textual Stochastic Gradient Descent with Momentum (TSGD-M) that reweights updates through momentum sampling, using bootstrapped minibatch validation accuracy as importance weights over historical prompts, with Gumbel-Top-k sampling for prompt generation.

Result: TSGD-M achieves consistent gains across 5 benchmarks and integrates seamlessly into existing prompt optimization frameworks like TextGrad, DSPy-COPRO, and AdalFlow.

Conclusion: The findings highlight the importance of incorporating probabilistic exploration into textual-gradient-based optimization for more stable and scalable prompt optimization.

Abstract: LLM-based prompt optimization, that uses LLM-provided “textual gradients” (feedback) to refine prompts, has emerged an effective method for automatic prompt engineering. However, its scalability and stability are unclear when using more data in training. We systematically investigate the potential and challenges of scaling training data in textual gradient descent. We show that naively scaling training examples is infeasible due to both explicit context-length limits and an implicit context wall, where long-context degradation yields diminishing returns. Inspired by prior wisdom in stochastic gradient descent, we propose Textual Stochastic Gradient Descent with Momentum (TSGD-M), which reweights updates through momentum sampling, using bootstrapped minibatch validation accuracy as importance weights over historical prompts. We introduce Gumbel-Top-$k$ sampling for prompt generation, balancing exploration–exploitation and improving sampling efficiency while maintaining a low-variance running mean estimator. TSGD-M integrates seamlessly into existing prompt optimization frameworks, including TextGrad, DSPy-COPRO, and AdalFlow, and achieves consistent gains across 5 benchmarks. Our findings highlight the importance of incorporating probabilistic exploration into textual-gradient-based optimization, paving the way for more stable and scalable prompt optimization.

Method: Analyzed token entropy patterns in Chain-of-Thought reasoning, studied entropy evolution during RLVR training, and improved RLVR by restricting policy gradient updates to high-entropy forking tokens (20% of tokens).

Result: RLVR primarily adjusts entropy of high-entropy tokens while largely preserving base model patterns. Using only 20% of tokens (high-entropy ones) maintained performance on Qwen3-8B and significantly improved performance on Qwen3-32B (+11.04 on AIME'25, +7.71 on AIME'24) and Qwen3-14B (+4.79 on AIME'25, +5.21 on AIME'24). Training on low-entropy tokens caused performance decline.

Conclusion: RLVR efficacy stems from optimizing high-entropy tokens that determine reasoning directions, revealing potential to understand and optimize RLVR through token-entropy perspective and leverage minority high-entropy tokens to enhance LLM reasoning.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful approach to enhancing the reasoning capabilities of Large Language Models (LLMs), while its mechanisms are not yet well understood. In this work, we undertake a pioneering exploration of RLVR through the novel perspective of token entropy patterns, comprehensively analyzing how different tokens influence reasoning performance. By examining token entropy patterns in Chain-of-Thought (CoT) reasoning, we observe that only a small fraction of tokens exhibit high entropy, and these tokens act as critical forks that steer the model toward diverse reasoning pathways. Furthermore, studying how entropy patterns evolve during RLVR training reveals that RLVR largely adheres to the base model’s entropy patterns, primarily adjusting the entropy of high-entropy tokens. These findings highlight the significance of high-entropy tokens (i.e., forking tokens) to RLVR. We ultimately improve RLVR by restricting policy gradient updates to forking tokens and uncover a finding even beyond the 80/20 rule: utilizing only 20% of the tokens while maintaining performance comparable to full-gradient updates on the Qwen3-8B base model and significantly surpassing full-gradient updates on the Qwen3-32B (+11.04 on AIME'25 and +7.71 on AIME'24) and Qwen3-14B (+4.79 on AIME'25 and +5.21 on AIME'24) base models, highlighting a strong scaling trend. In contrast, training exclusively on the 80% lowest-entropy tokens leads to a marked decline in performance. These findings indicate that the efficacy of RLVR primarily arises from optimizing the high-entropy tokens that decide reasoning directions. Collectively, our results highlight the potential to understand RLVR through a token-entropy perspective and optimize RLVR by leveraging high-entropy minority tokens to further improve LLM reasoning.

Method: Developed a structural-equation-modeling-aligned framework to quantify internal validity, dimensional separability, and component contributions. Created a Piaget-inspired capability hierarchy stratifying abilities into Perception, Memory, and Reasoning, then reorganized existing tasks under this theoretical framework.

Result: The GOLD benchmark demonstrates superior interpretability, lower indicator redundancy, and clearer cognitive consistency compared to prior benchmarks.

Conclusion: The proposed theoretical framework and GOLD benchmark provide a more structured, interpretable, and diagnostically powerful approach for evaluating multimodal large language models.

Abstract: Evaluating multimodal large language models (MLLMs) is fundamentally challenged by the absence of structured, interpretable, and theoretically grounded benchmarks; current heuristically-grouped tasks have vague cognitive targets, overlapping abilities, redundant indicators, and weak diagnostic power. We therefore propose a structural-equation-modeling-aligned framework that quantifies internal validity, dimensional separability, and component contributions, and introduce a Piaget-inspired capability hierarchy that stratifies MLLM abilities into Perception, Memory, and Reasoning. Reorganizing existing tasks under this theory, we build the GOLD benchmark, whose experiments show superior interpretability, lower indicator redundancy, and clearer cognitive consistency than prior benchmarks.

Method: Train CNNs on LLM attention weights to analyze memorization patterns, develop custom visual interpretability techniques to localize attention regions involved in different memorization forms.

Result: Existing taxonomy performs poorly; proposed new 3-category taxonomy (guessed samples, recalled samples, non-memorized) better aligns with attention weights; significant portion of extractable samples are guessed rather than recalled.

Conclusion: Few-shot verbatim memorization doesn’t correspond to distinct attention mechanisms; extractable samples should be studied separately as many are guessed; new taxonomy provides better framework for analyzing memorization mechanisms.

Abstract: Verbatim memorization in Large Language Models (LLMs) is a multifaceted phenomenon involving distinct underlying mechanisms. We introduce a novel method to analyze the different forms of memorization described by the existing taxonomy. Specifically, we train Convolutional Neural Networks (CNNs) on the attention weights of the LLM and evaluate the alignment between this taxonomy and the attention weights involved in decoding. We find that the existing taxonomy performs poorly and fails to reflect distinct mechanisms within the attention blocks. We propose a new taxonomy that maximizes alignment with the attention weights, consisting of three categories: memorized samples that are guessed using language modeling abilities, memorized samples that are recalled due to high duplication in the training set, and non-memorized samples. Our results reveal that few-shot verbatim memorization does not correspond to a distinct attention mechanism. We also show that a significant proportion of extractable samples are in fact guessed by the model and should therefore be studied separately. Finally, we develop a custom visual interpretability technique to localize the regions of the attention weights involved in each form of memorization.

Method: Manual analysis of recent multilingual evaluation sets in French and Telugu languages, identifying errors and comparing LLM performance differences between original and revised dataset versions.

Result: Identified several errors in both French and Telugu datasets, with performance differences up to almost 10% across multiple LLMs when comparing original vs revised versions.

Conclusion: Test sets should not be considered immutable; they need regular correctness checking, revision, and versioning. Recommendations provided for dataset creators and consumers to address quality issues.

Abstract: Several multilingual benchmark datasets have been developed in a semi-automatic manner in the recent past to measure progress and understand the state-of-the-art in the multilingual capabilities of Large Language Models. However, there is not a lot of attention paid to the quality of the datasets themselves, despite the existence of previous work in identifying errors in even fully human-annotated test sets. In this paper, we manually analyze recent multilingual evaluation sets in two languages - French and Telugu, identifying several errors in the process. We compare the performance difference across several LLMs with the original and revised versions of the datasets and identify large differences (almost 10% in some cases) in both languages). Based on these results, we argue that test sets should not be considered immutable and should be revisited, checked for correctness, and potentially versioned. We end with some recommendations for both the dataset creators as well as consumers on addressing the dataset quality issues.

Method: Proposes SPS metric comparing generated token areas with principal subspace directions, and xCompress framework for sampling, ranking, and compressing retrieval summaries.

Result: Experiments on five QA benchmarks with four LLMs show SPS enhances performance across tasks and provides insights into retrieval-generation interaction.

Conclusion: SPS offers a principled way to measure retrieval relevance and xCompress effectively improves RAG performance through dynamic summary compression.

Abstract: Large Language Models (LLMs) have shown improved generation performance through retrieval-augmented generation (RAG) following the retriever-reader paradigm, which supplements model inputs with externally retrieved knowledge. However, prior work often evaluates RAG holistically, assessing the retriever and reader jointly, making it difficult to isolate the true contribution of retrieval, particularly given the prompt sensitivity of LLMs used as readers. We move beyond perplexity and introduce Spectrum Projection Score (SPS), a lightweight and supervision-free metric that allows the reader to gauge the semantic alignment of a retrieved summary with its hidden representation by comparing the area formed by generated tokens from the summary, and the principal directions of subspace in the reader and to measure the relevance. Building on SPS we present xCompress, an inference-time controller framework that dynamically samples, ranks, and compresses retrieval summary candidates. Extensive experiments on five QA benchmarks with four open-sourced LLMs show that SPS not only enhances performance across a range of tasks but also provides a principled perspective on the interaction between retrieval and generation.

Method: Created EcomMMMU dataset with 406,190 samples and 8.99M images across 8 tasks, then proposed SUMEI method that predicts visual utilities before using images for downstream tasks.

Result: Analysis shows product images don’t consistently improve performance and can degrade it, indicating MLLMs struggle to leverage rich visual content effectively.

Conclusion: SUMEI method demonstrates effectiveness in strategically utilizing multiple images for e-commerce tasks, addressing limitations in current MLLM capabilities.

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

Method: Systematically evaluate agentic frameworks comparing data retrieval strategies (FHIR API vs tools), interaction patterns (single/multi-turn), and reasoning strategies (natural language vs code generation).

Result: Experiments reveal practical challenges in retrieving data from complex FHIR resources and difficulty in reasoning over them, significantly impacting question answering performance.

Conclusion: FHIR-AgentBench dataset and evaluation suite are released to promote reproducible research and development of robust LLM agents for clinical applications.

Abstract: The recent shift toward the Health Level Seven Fast Healthcare Interoperability Resources (HL7 FHIR) standard opens a new frontier for clinical AI, demanding LLM agents to navigate complex, resource-based data models instead of conventional structured health data. However, existing benchmarks have lagged behind this transition, lacking the realism needed to evaluate recent LLMs on interoperable clinical data. To bridge this gap, we introduce FHIR-AgentBench, a benchmark that grounds 2,931 real-world clinical questions in the HL7 FHIR standard. Using this benchmark, we systematically evaluate agentic frameworks, comparing different data retrieval strategies (direct FHIR API calls vs. specialized tools), interaction patterns (single-turn vs. multi-turn), and reasoning strategies (natural language vs. code generation). Our experiments highlight the practical challenges of retrieving data from intricate FHIR resources and the difficulty of reasoning over them, both of which critically affect question answering performance. We publicly release the FHIR-AgentBench dataset and evaluation suite (https://github.com/glee4810/FHIR-AgentBench) to promote reproducible research and the development of robust, reliable LLM agents for clinical applications.

Method: Trained a GPT-style model on 37 nineteenth-century novels by female authors, applied SAEs across layers, and probed with eleven social/moral categories to map sparse features to interpretable concepts.

Result: Revealed stable thematic backbones (especially gender and kinship) and showed how associations expand and entangle with depth in the model representations.

Conclusion: The LLM+SAEs pipeline provides a scalable framework for auditing how cultural assumptions from training data are embedded in model representations.

Abstract: Large Language Models (LLMs) are trained on massive, unstructured corpora, making it unclear which social patterns and biases they absorb and later reproduce. Existing evaluations typically examine outputs or activations, but rarely connect them back to the pre-training data. We introduce a pipeline that couples LLMs with sparse autoencoders (SAEs) to trace how different themes are encoded during training. As a controlled case study, we trained a GPT-style model on 37 nineteenth-century novels by ten female authors, a corpus centered on themes such as gender, marriage, class, and morality. By applying SAEs across layers and probing with eleven social and moral categories, we mapped sparse features to human-interpretable concepts. The analysis revealed stable thematic backbones (most prominently around gender and kinship) and showed how associations expand and entangle with depth. More broadly, we argue that the LLM+SAEs pipeline offers a scalable framework for auditing how cultural assumptions from the data are embedded in model representations.

Method: Used dimensionality reduction techniques, new similarity measures, compared language-trained vs. non-language-trained models, and examined the role of positional encoding in models.

Result: Confirmed previous correlations are robust to dimensionality concerns, validated with new similarity measures, specific to language-trained models, and dependent on positional encoding.

Conclusion: The study strengthens evidence for representational similarity between LLMs and brain states, supporting the biological plausibility and interpretability of state-of-the-art language models.

Abstract: Previous work has shown correlations between the hidden states of large language models and fMRI brain responses, on language tasks. These correlations have been taken as evidence of the representational similarity of these models and brain states. This study tests whether these previous results are robust to several possible concerns. Specifically this study shows: (i) that the previous results are still found after dimensionality reduction, and thus are not attributable to the curse of dimensionality; (ii) that previous results are confirmed when using new measures of similarity; (iii) that correlations between brain representations and those from models are specific to models trained on human language; and (iv) that the results are dependent on the presence of positional encoding in the models. These results confirm and strengthen the results of previous research and contribute to the debate on the biological plausibility and interpretability of state-of-the-art large language models.

Method: Created CCD-Bench with 2,182 open-ended dilemmas across 7 domains, paired with anonymized response options from 10 GLOBE cultural clusters, using stratified Latin square design to mitigate ordering effects. Evaluated 17 non-reasoning LLMs.

Result: Models strongly prefer Nordic Europe (20.2%) and Germanic Europe (12.4%) options, while underrepresenting Eastern Europe and Middle East/North Africa (5.6-5.8%). Although 87.9% of rationales reference multiple GLOBE dimensions, this pluralism is superficial - mainly recombining Future/Performance Orientation, rarely using Assertiveness or Gender Egalitarianism (<3%).

Conclusion: Current alignment pipelines promote consensus-oriented worldviews that underserve scenarios requiring power negotiation, rights-based reasoning, or gender-aware analysis. CCD-Bench shifts evaluation toward pluralistic decision-making and highlights need for alignment strategies that substantively engage diverse worldviews.

Abstract: Although large language models (LLMs) are increasingly implicated in interpersonal and societal decision-making, their ability to navigate explicit conflicts between legitimately different cultural value systems remains largely unexamined. Existing benchmarks predominantly target cultural knowledge (CulturalBench), value prediction (WorldValuesBench), or single-axis bias diagnostics (CDEval); none evaluate how LLMs adjudicate when multiple culturally grounded values directly clash. We address this gap with CCD-Bench, a benchmark that assesses LLM decision-making under cross-cultural value conflict. CCD-Bench comprises 2,182 open-ended dilemmas spanning seven domains, each paired with ten anonymized response options corresponding to the ten GLOBE cultural clusters. These dilemmas are presented using a stratified Latin square to mitigate ordering effects. We evaluate 17 non-reasoning LLMs. Models disproportionately prefer Nordic Europe (mean 20.2 percent) and Germanic Europe (12.4 percent), while options for Eastern Europe and the Middle East and North Africa are underrepresented (5.6 to 5.8 percent). Although 87.9 percent of rationales reference multiple GLOBE dimensions, this pluralism is superficial: models recombine Future Orientation and Performance Orientation, and rarely ground choices in Assertiveness or Gender Egalitarianism (both under 3 percent). Ordering effects are negligible (Cramer’s V less than 0.10), and symmetrized KL divergence shows clustering by developer lineage rather than geography. These patterns suggest that current alignment pipelines promote a consensus-oriented worldview that underserves scenarios demanding power negotiation, rights-based reasoning, or gender-aware analysis. CCD-Bench shifts evaluation beyond isolated bias detection toward pluralistic decision making and highlights the need for alignment strategies that substantively engage diverse worldviews.

Method: Uses cross-modal information flow to apply layer-specific merging strategies and introduces sensitivity-adaptive token matching that evaluates both token similarity and task-critical sensitivity.

Result: Reduces KV cache memory by 80-95% and decoding latency by 1.3-1.8x while maintaining competitive task performance across diverse MLLMs.

Conclusion: FlowMM effectively addresses multimodal KV cache efficiency challenges through adaptive cross-modal information flow guidance and sensitivity-aware merging.

Abstract: Traditional KV cache eviction strategies, which discard less critical KV-pairs based on attention scores, often degrade generation quality, causing context loss or hallucinations. Recent efforts shift toward KV merging, merging eviction tokens with retention tokens based on similarity. However, in multimodal scenarios, distributional biases across modality tokens and attentional biases in cross-modal interactions limit its effectiveness. This work introduces FlowMM, an adaptive framework for cross-modal information flow-guided multimodal KV cache merging. FlowMM leverages cross-modal information flow to dynamically apply layer-specific merging strategies, capturing modality-specific patterns while preserving contextual integrity. Furthermore, we introduce a sensitivity-adaptive token matching mechanism that jointly evaluates token similarity and task-critical sensitivity, merging low-risk tokens while safeguarding high-sensitivity ones. Extensive experiments across diverse leading MLLMs show that FlowMM reduces KV cache memory by 80% to 95% and decoding latency by 1.3-1.8x, while maintaining competitive task performance.

Method: Reviewed RAG applications in medicine, analyzing data sources, retrieval approaches (English-centric embeddings), LLM types (mostly generic), and evaluation methods (automated metrics and human evaluation).

Result: RAG research primarily uses public data with limited private data application. Retrieval relies on English-centric models, LLMs are mostly generic. Evaluation focuses on generation quality, accuracy, completeness, relevance, and fluency, with insufficient attention to bias and safety.

Conclusion: Medical RAG remains at early stage, requiring advances in clinical validation, cross-linguistic adaptation, and low-resource settings support for trustworthy global use.

Abstract: The rapid growth of medical knowledge and increasing complexity of clinical practice pose challenges. In this context, large language models (LLMs) have demonstrated value; however, inherent limitations remain. Retrieval-augmented generation (RAG) technologies show potential to enhance their clinical applicability. This study reviewed RAG applications in medicine. We found that research primarily relied on publicly available data, with limited application in private data. For retrieval, approaches commonly relied on English-centric embedding models, while LLMs were mostly generic, with limited use of medical-specific LLMs. For evaluation, automated metrics evaluated generation quality and task performance, whereas human evaluation focused on accuracy, completeness, relevance, and fluency, with insufficient attention to bias and safety. RAG applications were concentrated on question answering, report generation, text summarization, and information extraction. Overall, medical RAG remains at an early stage, requiring advances in clinical validation, cross-linguistic adaptation, and support for low-resource settings to enable trustworthy and responsible global use.

Method: Developed AlignSurvey benchmark with four tasks aligned with survey stages, created multi-tiered dataset architecture including Social Foundation Corpus and Entire-Pipeline Survey Datasets, and released SurveyLM family through two-stage fine-tuning of open-source LLMs.

Result: Built comprehensive resources including 44K+ interview dialogues, 400K+ structured survey records, and expert-annotated datasets for cross-cultural evaluation, with all datasets, models, and tools publicly available.

Conclusion: AlignSurvey enables transparent and socially responsible research by providing systematic evaluation of LLMs across the full social survey pipeline with focus on demographic diversity and fairness assessment.

Abstract: Understanding human attitudes, preferences, and behaviors through social surveys is essential for academic research and policymaking. Yet traditional surveys face persistent challenges, including fixed-question formats, high costs, limited adaptability, and difficulties ensuring cross-cultural equivalence. While recent studies explore large language models (LLMs) to simulate survey responses, most are limited to structured questions, overlook the entire survey process, and risks under-representing marginalized groups due to training data biases. We introduce AlignSurvey, the first benchmark that systematically replicates and evaluates the full social survey pipeline using LLMs. It defines four tasks aligned with key survey stages: social role modeling, semi-structured interview modeling, attitude stance modeling and survey response modeling. It also provides task-specific evaluation metrics to assess alignment fidelity, consistency, and fairness at both individual and group levels, with a focus on demographic diversity. To support AlignSurvey, we construct a multi-tiered dataset architecture: (i) the Social Foundation Corpus, a cross-national resource with 44K+ interview dialogues and 400K+ structured survey records; and (ii) a suite of Entire-Pipeline Survey Datasets, including the expert-annotated AlignSurvey-Expert (ASE) and two nationally representative surveys for cross-cultural evaluation. We release the SurveyLM family, obtained through two-stage fine-tuning of open-source LLMs, and offer reference models for evaluating domain-specific alignment. All datasets, models, and tools are available at github and huggingface to support transparent and socially responsible research.

Method: Designed shortcut scenarios with spurious associations between user labels and superficial textual cues. Proposed mitigation using large language models and counterfactual data augmentation across data distribution and model causal extraction levels.

Result: Baseline models showed 32% average relative accuracy drop under shortcut scenarios. Proposed strategies achieved 56% average relative performance improvement in robustness.

Conclusion: Shortcut learning significantly impacts social bot detector robustness. Counterfactual data augmentation with LLMs effectively mitigates this vulnerability by addressing spurious correlations and enhancing causal feature extraction.

Abstract: While existing social bot detectors perform well on benchmarks, their robustness across diverse real-world scenarios remains limited due to unclear ground truth and varied misleading cues. In particular, the impact of shortcut learning, where models rely on spurious correlations instead of capturing causal task-relevant features, has received limited attention. To address this gap, we conduct an in-depth study to assess how detectors are influenced by potential shortcuts based on textual features, which are most susceptible to manipulation by social bots. We design a series of shortcut scenarios by constructing spurious associations between user labels and superficial textual cues to evaluate model robustness. Results show that shifts in irrelevant feature distributions significantly degrade social bot detector performance, with an average relative accuracy drop of 32% in the baseline models. To tackle this challenge, we propose mitigation strategies based on large language models, leveraging counterfactual data augmentation. These methods mitigate the problem from data and model perspectives across three levels, including data distribution at both the individual user text and overall dataset levels, as well as the model’s ability to extract causal information. Our strategies achieve an average relative performance improvement of 56% under shortcut scenarios.

Method: Empirical analysis of verbalization quality and monitor reliability across domains, investigation of CoT intervention effects, and proposal of MoME paradigm for structured monitoring.

Result: Provides evidence on correlations between verbalization quality, monitor reliability, and LLM performance, and shows how different CoT interventions affect monitoring effectiveness.

Conclusion: CoT monitoring shows potential for AI safety but faces fundamental challenges; the proposed MoME paradigm offers a structured approach for monitoring model misbehavior through reasoning traces.

Abstract: Large Reasoning Models (LRMs) have demonstrated remarkable performance on complex tasks by engaging in extended reasoning before producing final answers. Beyond improving abilities, these detailed reasoning traces also create a new opportunity for AI safety, CoT Monitorability: monitoring potential model misbehavior, such as the use of shortcuts or sycophancy, through their chain-of-thought (CoT) during decision-making. However, two key fundamental challenges arise when attempting to build more effective monitors through CoT analysis. First, as prior research on CoT faithfulness has pointed out, models do not always truthfully represent their internal decision-making in the generated reasoning. Second, monitors themselves may be either overly sensitive or insufficiently sensitive, and can potentially be deceived by models’ long, elaborate reasoning traces. In this paper, we present the first systematic investigation of the challenges and potential of CoT monitorability. Motivated by two fundamental challenges we mentioned before, we structure our study around two central perspectives: (i) verbalization: to what extent do LRMs faithfully verbalize the true factors guiding their decisions in the CoT, and (ii) monitor reliability: to what extent can misbehavior be reliably detected by a CoT-based monitor? Specifically, we provide empirical evidence and correlation analyses between verbalization quality, monitor reliability, and LLM performance across mathematical, scientific, and ethical domains. Then we further investigate how different CoT intervention methods, designed to improve reasoning efficiency or performance, will affect monitoring effectiveness. Finally, we propose MoME, a new paradigm in which LLMs monitor other models’ misbehavior through their CoT and provide structured judgments along with supporting evidence.

Method: Proposes Bi-RAR framework that evaluates each intermediate step bidirectionally using information distance based on Kolmogorov complexity, approximated via language model generation probabilities. Uses multi-objective reinforcement learning with cascading reward structure for optimization.

Result: Empirical results on seven question answering benchmarks demonstrate that Bi-RAR surpasses previous methods and enables efficient interaction and reasoning with search engines during training and inference.

Conclusion: Bi-RAR provides an effective solution for improving retrieval-augmented reasoning by bidirectional evaluation of intermediate steps, addressing limitations of outcome-based supervision in complex reasoning tasks.

Abstract: Retrieval-augmented generation (RAG) has proven to be effective in mitigating hallucinations in large language models, yet its effectiveness remains limited in complex, multi-step reasoning scenarios. Recent efforts have incorporated search-based interactions into RAG, enabling iterative reasoning with real-time retrieval. Most approaches rely on outcome-based supervision, offering no explicit guidance for intermediate steps. This often leads to reward hacking and degraded response quality. We propose Bi-RAR, a novel retrieval-augmented reasoning framework that evaluates each intermediate step jointly in both forward and backward directions. To assess the information completeness of each step, we introduce a bidirectional information distance grounded in Kolmogorov complexity, approximated via language model generation probabilities. This quantification measures both how far the current reasoning is from the answer and how well it addresses the question. To optimize reasoning under these bidirectional signals, we adopt a multi-objective reinforcement learning framework with a cascading reward structure that emphasizes early trajectory alignment. Empirical results on seven question answering benchmarks demonstrate that Bi-RAR surpasses previous methods and enables efficient interaction and reasoning with the search engine during training and inference.

Method: Jointly learns global and local feature representations with dynamic knowledge distillation, uses anchors to refine feature space (strictly local and non-transmitted for privacy), and provides theoretical convergence analysis for nonconvex objectives.

Result: Outperforms 9 baseline methods on 5 image-classification datasets in effectiveness, stability, scalability, and security, with 4.3% improvement over the best baseline in effectiveness experiments.

Conclusion: FedeCouple effectively addresses the coordination problem between feature extractors and classifiers in federated learning, achieving better balance between global generalization and local adaptability while maintaining privacy and communication efficiency.

Abstract: In privacy-preserving mobile network transmission scenarios with heterogeneous client data, personalized federated learning methods that decouple feature extractors and classifiers have demonstrated notable advantages in enhancing learning capability. However, many existing approaches primarily focus on feature space consistency and classification personalization during local training, often neglecting the local adaptability of the extractor and the global generalization of the classifier. This oversight results in insufficient coordination and weak coupling between the components, ultimately degrading the overall model performance. To address this challenge, we propose FedeCouple, a federated learning method that balances global generalization and local adaptability at a fine-grained level. Our approach jointly learns global and local feature representations while employing dynamic knowledge distillation to enhance the generalization of personalized classifiers. We further introduce anchors to refine the feature space; their strict locality and non-transmission inherently preserve privacy and reduce communication overhead. Furthermore, we provide a theoretical analysis proving that FedeCouple converges for nonconvex objectives, with iterates approaching a stationary point as the number of communication rounds increases. Extensive experiments conducted on five image-classification datasets demonstrate that FedeCouple consistently outperforms nine baseline methods in effectiveness, stability, scalability, and security. Notably, in experiments evaluating effectiveness, FedeCouple surpasses the best baseline by a significant margin of 4.3%.

Method: Proposes MMaDA-Parallel framework with parallel multimodal diffusion for continuous bidirectional text-image interaction, trained with supervised finetuning and optimized by Parallel Reinforcement Learning (ParaRL) with semantic rewards along the denoising trajectory.

Result: Significantly improves cross-modal alignment and semantic consistency, achieving 6.9% improvement in Output Alignment on ParaBench compared to state-of-the-art model Bagel.

Conclusion: Establishes a more robust paradigm for thinking-aware image synthesis by enabling continuous cross-modal interaction throughout the generation process.

Abstract: While thinking-aware generation aims to improve performance on complex tasks, we identify a critical failure mode where existing sequential, autoregressive approaches can paradoxically degrade performance due to error propagation. To systematically analyze this issue, we propose ParaBench, a new benchmark designed to evaluate both text and image output modalities. Our analysis using ParaBench reveals that this performance degradation is strongly correlated with poor alignment between the generated reasoning and the final image. To resolve this, we propose a parallel multimodal diffusion framework, MMaDA-Parallel, that enables continuous, bidirectional interaction between text and images throughout the entire denoising trajectory. MMaDA-Parallel is trained with supervised finetuning and then further optimized by Parallel Reinforcement Learning (ParaRL), a novel strategy that applies semantic rewards along the trajectory to enforce cross-modal consistency. Experiments validate that our model significantly improves cross-modal alignment and semantic consistency, achieving a 6.9% improvement in Output Alignment on ParaBench compared to the state-of-the-art model, Bagel, establishing a more robust paradigm for thinking-aware image synthesis. Our code is open-sourced at https://github.com/tyfeld/MMaDA-Parallel

Method: Created PriVi dataset (424 hours of curated primate videos from 11 research settings and web sources), pretrained V-JEPA model on this data, and evaluated with frozen classifier on four benchmark datasets.

Result: Outperformed prior work including fully finetuned baselines across ChimpACT, BaboonLand, PanAf500, and ChimpBehave datasets, showing better scaling with fewer labels.

Conclusion: Primate-centric pretraining substantially improves data efficiency and generalization, making it promising for low-label applications in primate behavior research.

Abstract: Non-human primates are our closest living relatives, and analyzing their behavior is central to research in cognition, evolution, and conservation. Computer vision could greatly aid this research, but existing methods often rely on human-centric pretrained models and focus on single datasets, which limits generalization. We address this limitation by shifting from a model-centric to a data-centric approach and introduce PriVi, a large-scale primate-centric video pretraining dataset. PriVi contains 424 hours of curated video, combining 174 hours from behavioral research across 11 settings with 250 hours of diverse web-sourced footage, assembled through a scalable data curation pipeline. We pretrain V-JEPA on PriVi to learn primate-specific representations and evaluate it using a lightweight frozen classifier. Across four benchmark datasets, ChimpACT, BaboonLand, PanAf500, and ChimpBehave, our approach consistently outperforms prior work, including fully finetuned baselines, and scales favorably with fewer labels. These results demonstrate that primate-centric pretraining substantially improves data efficiency and generalization, making it a promising approach for low-label applications. Code, models, and the majority of the dataset will be made available.

Method: Hierarchical SUSA architecture with TSU module for local semantic understanding via view-level descriptions and DSP module for global spatial perception via depth exploration maps.

Result: Significantly improves navigation performance on REVERIE, R2R, SOON benchmarks and generalizes better to continuous R2R-CE benchmark compared to baseline.

Conclusion: Hierarchical representation enrichment through semantic understanding and spatial awareness enables more effective environment grounding and navigation in VLN tasks.

Abstract: Navigating unseen environments from natural language instructions remains challenging for egocentric agents in Vision-and-Language Navigation (VLN). Humans naturally ground concrete semantic knowledge within spatial layouts during indoor navigation. Although prior work has introduced diverse environment representations to improve reasoning, auxiliary modalities are often naively concatenated with RGB features, which underutilizes each modality’s distinct contribution. We propose a hierarchical Semantic Understanding and Spatial Awareness (SUSA) architecture to enable agents to perceive and ground environments at multiple scales. Specifically, the Textual Semantic Understanding (TSU) module supports local action prediction by generating view-level descriptions, capturing fine-grained semantics and narrowing the modality gap between instructions and environments. Complementarily, the Depth Enhanced Spatial Perception (DSP) module incrementally builds a trajectory-level depth exploration map, providing a coarse-grained representation of global spatial layout. Extensive experiments show that the hierarchical representation enrichment of SUSA significantly improves navigation performance over the baseline on discrete VLN benchmarks (REVERIE, R2R, and SOON) and generalizes better to the continuous R2R-CE benchmark.

Method: Ordinal regression framework modified with novel soft label loss functions to account for label uncertainty and ordinal nature of severity ratings.

Result: Predictive performance approaches clinical expert level with well-calibrated uncertainty estimates.

Conclusion: Automated phonotrauma severity assessment enables large-scale studies for improved clinical understanding and patient care.

Abstract: Phonotrauma refers to vocal fold tissue damage resulting from exposure to forces during voicing. It occurs on a continuum from mild to severe, and treatment options can vary based on severity. Assessment of severity involves a clinician’s expert judgment, which is costly and can vary widely in reliability. In this work, we present the first method for automatically classifying phonotrauma severity from vocal fold images. To account for the ordinal nature of the labels, we adopt a widely used ordinal regression framework. To account for label uncertainty, we propose a novel modification to ordinal regression loss functions that enables them to operate on soft labels reflecting annotator rating distributions. Our proposed soft ordinal regression method achieves predictive performance approaching that of clinical experts, while producing well-calibrated uncertainty estimates. By providing an automated tool for phonotrauma severity assessment, our work can enable large-scale studies of phonotrauma, ultimately leading to improved clinical understanding and patient care.

Method: Proposes a framework inspired by Structural Causal Models to decouple action semantics from human attributes, enabling text-to-semantics prediction and attribute-controlled generation. Also introduces a new dataset with attribute annotations for evaluation.

Result: The model successfully generates attribute-aware motion aligned with user’s text and attribute inputs. Extensive experiments validate the model’s effectiveness.

Conclusion: This work represents a pilot exploration for bridging the gap in attribute-aware human motion generation, providing a new benchmark and demonstrating the importance of considering human attributes in motion generation.

Abstract: Text-driven human motion generation has recently attracted considerable attention, allowing models to generate human motions based on textual descriptions. However, current methods neglect the influence of human attributes-such as age, gender, weight, and height-which are key factors shaping human motion patterns. This work represents a pilot exploration for bridging this gap. We conceptualize each motion as comprising both attribute information and action semantics, where textual descriptions align exclusively with action semantics. To achieve this, a new framework inspired by Structural Causal Models is proposed to decouple action semantics from human attributes, enabling text-to-semantics prediction and attribute-controlled generation. The resulting model is capable of generating attribute-aware motion aligned with the user’s text and attribute inputs. For evaluation, we introduce a comprehensive dataset containing attribute annotations for text-motion pairs, setting the first benchmark for attribute-aware motion generation. Extensive experiments validate our model’s effectiveness.

Method: Disentangles individual instructions from multi-part prompts and exposes each as a globally trained slider using a single set of low-rank adaptation matrices that generalize across diverse edits, attributes, and compositional instructions.

Result: Substantial improvements in edit controllability, visual consistency, and user steerability when applied to state-of-the-art models like FLUX-Kontext and Qwen-Image-Edit, enabling continuous interpolation along individual edit dimensions.

Conclusion: First framework for continuous, fine-grained instruction control in instruction-based image editing, paving the way for interactive, instruction-driven image manipulation with continuous and compositional control.

Abstract: Instruction-based image editing models have recently achieved impressive performance, enabling complex edits to an input image from a multi-instruction prompt. However, these models apply each instruction in the prompt with a fixed strength, limiting the user’s ability to precisely and continuously control the intensity of individual edits. We introduce SliderEdit, a framework for continuous image editing with fine-grained, interpretable instruction control. Given a multi-part edit instruction, SliderEdit disentangles the individual instructions and exposes each as a globally trained slider, allowing smooth adjustment of its strength. Unlike prior works that introduced slider-based attribute controls in text-to-image generation, typically requiring separate training or fine-tuning for each attribute or concept, our method learns a single set of low-rank adaptation matrices that generalize across diverse edits, attributes, and compositional instructions. This enables continuous interpolation along individual edit dimensions while preserving both spatial locality and global semantic consistency. We apply SliderEdit to state-of-the-art image editing models, including FLUX-Kontext and Qwen-Image-Edit, and observe substantial improvements in edit controllability, visual consistency, and user steerability. To the best of our knowledge, we are the first to explore and propose a framework for continuous, fine-grained instruction control in instruction-based image editing models. Our results pave the way for interactive, instruction-driven image manipulation with continuous and compositional control.

Method: Integrates denoising probabilistic model with diffusion processes, uses narrow Gaussian kernels for multiple density maps, incorporates regression branch for feature extraction, consolidation mechanism for map combination, and event-driven sampling with Farneback optical flow to reduce computational load.

Result: Model effectively captures crowd dynamics in both dense and sparse settings, reduces frame counts while maintaining essential crowd events, and demonstrates efficiency through qualitative/quantitative evaluations including MAE metrics.

Conclusion: The work provides a scalable and efficient framework for real-time crowd monitoring by successfully adapting image-based algorithms to video scenarios with improved computational efficiency.

Abstract: This study enhances a crowd density estimation algorithm originally designed for image-based analysis by adapting it for video-based scenarios. The proposed method integrates a denoising probabilistic model that utilizes diffusion processes to generate high-quality crowd density maps. To improve accuracy, narrow Gaussian kernels are employed, and multiple density map outputs are generated. A regression branch is incorporated into the model for precise feature extraction, while a consolidation mechanism combines these maps based on similarity scores to produce a robust final result. An event-driven sampling technique, utilizing the Farneback optical flow algorithm, is introduced to selectively capture frames showing significant crowd movements, reducing computational load and storage by focusing on critical crowd dynamics. Through qualitative and quantitative evaluations, including overlay plots and Mean Absolute Error (MAE), the model demonstrates its ability to effectively capture crowd dynamics in both dense and sparse settings. The efficiency of the sampling method is further assessed, showcasing its capability to decrease frame counts while maintaining essential crowd events. By addressing the temporal challenges unique to video analysis, this work offers a scalable and efficient framework for real-time crowd monitoring in applications such as public safety, disaster response, and event management.

Method: Modular system that reconstructs scale-aligned 3D point clouds from posed RGB images using Depth Pro foundation model, followed by geometric layout matching via convolution with floor plans to output location/orientation posteriors.

Result: Outperformed PALMS and F3Loc in stationary localization accuracy on Structured3D and custom campus dataset (80 observations across 4 buildings). When integrated with particle filter for sequential localization on 33 real trajectories, achieved lower errors than other methods.

Conclusion: PALMS+ demonstrates robust infrastructure-free indoor localization suitable for camera-free tracking applications, showing potential for accessible indoor navigation without requiring training data.

Abstract: Indoor localization in GPS-denied environments is crucial for applications like emergency response and assistive navigation. Vision-based methods such as PALMS enable infrastructure-free localization using only a floor plan and a stationary scan, but are limited by the short range of smartphone LiDAR and ambiguity in indoor layouts. We propose PALMS$+$, a modular, image-based system that addresses these challenges by reconstructing scale-aligned 3D point clouds from posed RGB images using a foundation monocular depth estimation model (Depth Pro), followed by geometric layout matching via convolution with the floor plan. PALMS$+$ outputs a posterior over the location and orientation, usable for direct or sequential localization. Evaluated on the Structured3D and a custom campus dataset consisting of 80 observations across four large campus buildings, PALMS$+$ outperforms PALMS and F3Loc in stationary localization accuracy – without requiring any training. Furthermore, when integrated with a particle filter for sequential localization on 33 real-world trajectories, PALMS$+$ achieved lower localization errors compared to other methods, demonstrating robustness for camera-free tracking and its potential for infrastructure-free applications. Code and data are available at https://github.com/Head-inthe-Cloud/PALMS-Plane-based-Accessible-Indoor-Localization-Using-Mobile-Smartphones

Method: Proposes a Social LSTM model enhanced with a Dynamic Occupied Space loss function that combines average displacement error with a collision penalty sensitive to scene density and individual spatial occupancy.

Result: The model achieves up to 31% reduction in collision rate and reduces average displacement error by 5% and final displacement error by 6% on average across all datasets compared to baseline, outperforming state-of-the-art models.

Conclusion: The proposed model effectively reduces collisions while maintaining or improving trajectory prediction accuracy across various crowd density conditions, demonstrating the importance of considering physical occupancy space in pedestrian trajectory prediction.

Abstract: In dynamic and crowded environments, realistic pedestrian trajectory prediction remains a challenging task due to the complex nature of human motion and the mutual influences among individuals. Deep learning models have recently achieved promising results by implicitly learning such patterns from 2D trajectory data. However, most approaches treat pedestrians as point entities, ignoring the physical space that each person occupies. To address these limitations, this paper proposes a novel deep learning model that enhances the Social LSTM with a new Dynamic Occupied Space loss function. This loss function guides Social LSTM in learning to avoid realistic collisions without increasing displacement error across different crowd densities, ranging from low to high, in both homogeneous and heterogeneous density settings. Such a function achieves this by combining the average displacement error with a new collision penalty that is sensitive to scene density and individual spatial occupancy. For efficient training and evaluation, five datasets were generated from real pedestrian trajectories recorded during the Festival of Lights in Lyon 2022. Four datasets represent homogeneous crowd conditions – low, medium, high, and very high density – while the fifth corresponds to a heterogeneous density distribution. The experimental findings indicate that the proposed model not only lowers collision rates but also enhances displacement prediction accuracy in each dataset. Specifically, the model achieves up to a 31% reduction in the collision rate and reduces the average displacement error and the final displacement error by 5% and 6%, respectively, on average across all datasets compared to the baseline. Moreover, the proposed model consistently outperforms several state-of-the-art deep learning models across most test sets.

Method: The authors regard soiling detection as a semantic segmentation problem, provide comprehensive comparison of popular segmentation methods, create a new data subset from Woodscape to address data leakage, and make their code and dataset splits publicly available.

Result: Segmentation methods show superiority over tile-level classification approaches. The new smaller data subset enables comparable results in much shorter time despite being smaller than the original dataset.

Conclusion: Semantic segmentation is effective for soiling detection, and addressing data leakage in datasets enables comparable performance with smaller datasets and reduced training time.

Abstract: Soiling detection for automotive cameras is a crucial part of advanced driver assistance systems to make them more robust to external conditions like weather, dust, etc. In this paper, we regard the soiling detection as a semantic segmentation problem. We provide a comprehensive comparison of popular segmentation methods and show their superiority in performance while comparing them to tile-level classification approaches. Moreover, we present an extensive analysis of the Woodscape dataset showing that the original dataset contains a data-leakage and imprecise annotations. To address these problems, we create a new data subset, which, despite being much smaller, provides enough information for the segmentation method to reach comparable results in a much shorter time. All our codes and dataset splits are available at https://github.com/filipberanek/woodscape_revision.

Method: Evaluated vision encoders from 8 medical and general-domain foundation models for chest X-ray analysis using linear probing and fine-tuning on classification (pneumothorax, cardiomegaly) and segmentation (pneumothorax, cardiac boundary) tasks.

Result: Medical pre-training provides significant advantage; medical FMs consistently outperformed general-domain models in linear probing. Pre-trained embeddings were strong for global classification but poor for segmenting complex pathologies without fine-tuning. Image-only and label-supervised FMs performed well, and supervised baselines remained competitive.

Conclusion: Medical pre-training is beneficial but architectural choices are critical; pre-trained features are not universally effective, especially for complex localization tasks where supervised models remain strong alternatives.

Abstract: Foundation models (FMs) promise to generalize medical imaging, but their effectiveness varies. It remains unclear how pre-training domain (medical vs. general), paradigm (e.g., text-guided), and architecture influence embedding quality, hindering the selection of optimal encoders for specific radiology tasks. To address this, we evaluate vision encoders from eight medical and general-domain FMs for chest X-ray analysis. We benchmark classification (pneumothorax, cardiomegaly) and segmentation (pneumothorax, cardiac boundary) using linear probing and fine-tuning. Our results show that domain-specific pre-training provides a significant advantage; medical FMs consistently outperformed general-domain models in linear probing, establishing superior initial feature quality. However, feature utility is highly task-dependent. Pre-trained embeddings were strong for global classification and segmenting salient anatomy (e.g., heart). In contrast, for segmenting complex, subtle pathologies (e.g., pneumothorax), all FMs performed poorly without significant fine-tuning, revealing a critical gap in localizing subtle disease. Subgroup analysis showed FMs use confounding shortcuts (e.g., chest tubes for pneumothorax) for classification, a strategy that fails for precise segmentation. We also found that expensive text-image alignment is not a prerequisite; image-only (RAD-DINO) and label-supervised (Ark+) FMs were among top performers. Notably, a supervised, end-to-end baseline remained highly competitive, matching or exceeding the best FMs on segmentation tasks. These findings show that while medical pre-training is beneficial, architectural choices (e.g., multi-scale) are critical, and pre-trained features are not universally effective, especially for complex localization tasks where supervised models remain a strong alternative.

Method: Traverse generative model’s latent space guided by gradients of specific iris features (sharpness, pupil size, etc.) while preserving identity using GAN inversion for real images.

Result: Enables controlled manipulation of iris image properties while maintaining identity, applicable to both synthetic and real iris images.

Conclusion: The approach provides flexible iris image augmentation that can be extended to manipulate any differentiable attribute, enhancing dataset diversity for iris recognition systems.

Abstract: Developing reliable iris recognition and presentation attack detection methods requires diverse datasets that capture realistic variations in iris features and a wide spectrum of anomalies. Because of the rich texture of iris images, which spans a wide range of spatial frequencies, synthesizing same-identity iris images while controlling specific attributes remains challenging. In this work, we introduce a new iris image augmentation strategy by traversing a generative model’s latent space toward latent codes that represent same-identity samples but with some desired iris image properties manipulated. The latent space traversal is guided by a gradient of specific geometrical, textural, or quality-related iris image features (e.g., sharpness, pupil size, iris size, or pupil-to-iris ratio) and preserves the identity represented by the image being manipulated. The proposed approach can be easily extended to manipulate any attribute for which a differentiable loss term can be formulated. Additionally, our approach can use either randomly generated images using either a pre-train GAN model or real-world iris images. We can utilize GAN inversion to project any given iris image into the latent space and obtain its corresponding latent code.

Method: Three-stage pipeline: 1) Contextual object descriptions guide localization, 2) Self-cross-attention mechanisms identify candidate regions, 3) Segmentation model produces precise masks for pose estimation. Includes automatic re-registration mechanism that monitors feature similarity to detect and recover from tracking failures.

Result: Achieves state-of-the-art accuracy on challenging industrial datasets with multi-object occlusions, high-speed motion, and varying illumination. Operates at real-time speeds without additional training.

Conclusion: STORM provides a practical annotation-free solution that significantly reduces deployment overhead for applications like flexible manufacturing and intelligent quality control, while maintaining robust performance under challenging conditions.

Abstract: Accurate 6D pose estimation and tracking are fundamental capabilities for physical AI systems such as robots. However, existing approaches typically rely on a manually annotated segmentation mask of the target in the first frame, which is labor-intensive and leads to reduced performance when faced with occlusions or rapid movement. To address these limi- tations, we propose STORM (Segment, Track, and Object Re-localization from a single 3D Model), an open-source robust real-time 6D pose estimation system that requires no manual annotation. STORM employs a novel three-stage pipeline combining vision-language understanding with self-supervised feature matching: contextual object descriptions guide localization, self-cross-attention mechanisms identify candidate regions, and a segmentation model produces precise masks for accurate pose estimation. Another key innovation is our automatic re-registration mechanism that detects tracking failures through feature similarity monitoring and recovers from severe occlusions or rapid motion. STORM achieves state-of-the-art accuracy on challenging industrial datasets featuring multi-object occlusions, high-speed motion, and varying illumination, while operating at real-time speeds without additional training. This annotation-free approach significantly reduces deployment overhead, providing a practical solution for modern applications, such as flexible manufacturing and intelligent quality control.

Method: Uses CLIP encoder to identify representative regions from low-frequency classes and transplant them into frequent-class samples within each task. Also incorporates adaptive balancing strategy leveraging prior task distributions to smooth inter-task imbalances.

Result: Extensive experiments show PANDA improves accuracy and reduces catastrophic forgetting when integrated with existing pre-trained model-based continual learning methods.

Conclusion: PANDA effectively addresses dual-level data imbalances in exemplar-free continual learning and can be seamlessly integrated with existing pre-trained model-based methods to enhance performance.

Abstract: Exemplar-Free Continual Learning (EFCL) restricts the storage of previous task data and is highly susceptible to catastrophic forgetting. While pre-trained models (PTMs) are increasingly leveraged for EFCL, existing methods often overlook the inherent imbalance of real-world data distributions. We discovered that real-world data streams commonly exhibit dual-level imbalances, dataset-level distributions combined with extreme or reversed skews within individual tasks, creating both intra-task and inter-task disparities that hinder effective learning and generalization. To address these challenges, we propose PANDA, a Patch-and-Distribution-Aware Augmentation framework that integrates seamlessly with existing PTM-based EFCL methods. PANDA amplifies low-frequency classes by using a CLIP encoder to identify representative regions and transplanting those into frequent-class samples within each task. Furthermore, PANDA incorporates an adaptive balancing strategy that leverages prior task distributions to smooth inter-task imbalances, reducing the overall gap between average samples across tasks and enabling fairer learning with frozen PTMs. Extensive experiments and ablation studies demonstrate PANDA’s capability to work with existing PTM-based CL methods, improving accuracy and reducing catastrophic forgetting.

Method: Extracts spectral subspace from textual embeddings to define principal semantic directions, learns per-sample shift parameters to minimize entropy across augmented views, operates entirely in latent space without backpropagation.

Result: Outperforms state-of-the-art test-time adaptation methods, achieves 8x faster inference speed and 12x smaller memory footprint compared to conventional test-time prompt tuning.

Conclusion: STS provides an efficient and effective alternative to existing test-time adaptation methods, enabling practical deployment of VLMs under domain shifts with minimal computational overhead.

Abstract: Vision-Language Models (VLMs) excel at zero-shot inference but often degrade under test-time domain shifts. For this reason, episodic test-time adaptation strategies have recently emerged as powerful techniques for adapting VLMs to a single unlabeled image. However, existing adaptation strategies, such as test-time prompt tuning, typically require backpropagating through large encoder weights or altering core model components. In this work, we introduce Spectrum-Aware Test-Time Steering (STS), a lightweight adaptation framework that extracts a spectral subspace from the textual embeddings to define principal semantic directions and learns to steer latent representations in a spectrum-aware manner by adapting a small number of per-sample shift parameters to minimize entropy across augmented views. STS operates entirely at inference in the latent space, without backpropagation through or modification of the frozen encoders. Building on standard evaluation protocols, our comprehensive experiments demonstrate that STS largely surpasses or compares favorably against state-of-the-art test-time adaptation methods, while introducing only a handful of additional parameters and achieving inference speeds up to 8x faster with a 12x smaller memory footprint than conventional test-time prompt tuning. The code is available at https://github.com/kdafnis/STS.

Method: Built on a geometry-grounded backbone with 3D Gaussian representation, uses cross-illumination distillation where teacher network learns from normal-light ground truth to transfer geometric information to student model, plus dedicated Lumos loss for photometric consistency.

Result: Achieves high-fidelity low-light 3D scene restoration with accurate geometry and strong generalization to unseen cases, also extends to over-exposure correction.

Conclusion: Lumos3D provides a scalable, generalizable solution for 3D lighting restoration that works without camera poses or per-scene training, demonstrating versatility across diverse lighting correction tasks.

Abstract: Restoring 3D scenes captured under low-light con- ditions remains a fundamental yet challenging problem. Most existing approaches depend on precomputed camera poses and scene-specific optimization, which greatly restricts their scala- bility to dynamic real-world environments. To overcome these limitations, we introduce Lumos3D, a generalizable pose-free framework for 3D low-light scene restoration. Trained once on a single dataset, Lumos3D performs inference in a purely feed- forward manner, directly restoring illumination and structure from unposed, low-light multi-view images without any per- scene training or optimization. Built upon a geometry-grounded backbone, Lumos3D reconstructs a normal-light 3D Gaussian representation that restores illumination while faithfully pre- serving structural details. During training, a cross-illumination distillation scheme is employed, where the teacher network is distilled on normal-light ground truth to transfer accurate geometric information, such as depth, to the student model. A dedicated Lumos loss is further introduced to promote photomet- ric consistency within the reconstructed 3D space. Experiments on real-world datasets demonstrate that Lumos3D achieves high- fidelity low-light 3D scene restoration with accurate geometry and strong generalization to unseen cases. Furthermore, the framework naturally extends to handle over-exposure correction, highlighting its versatility for diverse lighting restoration tasks.

Method: SPOT leverages token embeddings, interactions, and attention dynamics across layers to infer token importance. It uses lightweight predictors that can be plugged into various ViT architectures to derive input-specific token prioritization.

Result: Empirical evaluations show significant efficiency gains of up to 40% compared to standard ViTs, while maintaining or even improving accuracy across various tasks.

Conclusion: SPOT provides a versatile framework for token sparsification in ViTs that improves computational efficiency without sacrificing performance, with adaptable designs for different resource constraints.

Abstract: While Vision Transformers (ViT) have demonstrated remarkable performance across diverse tasks, their computational demands are substantial, scaling quadratically with the number of processed tokens. Compact attention representations, reflecting token interaction distributions, can guide early detection and reduction of less salient tokens prior to attention computation. Motivated by this, we present SParsification with attentiOn dynamics via Token relevance (SPOT), a framework for early detection of redundant tokens within ViTs that leverages token embeddings, interactions, and attention dynamics across layers to infer token importance, resulting in a more context-aware and interpretable relevance detection process. SPOT informs token sparsification and facilitates the elimination of such tokens, improving computational efficiency without sacrificing performance. SPOT employs computationally lightweight predictors that can be plugged into various ViT architectures and learn to derive effective input-specific token prioritization across layers. Its versatile design supports a range of performance levels adaptable to varying resource constraints. Empirical evaluations demonstrate significant efficiency gains of up to 40% compared to standard ViTs, while maintaining or even improving accuracy. Code and models are available at https://github.com/odedsc/SPOT .

Method: Extracts geographic cues from street-view images via web search and LLM location inference, generates satellite queries using geocoding API, and retrieves matches using pretrained vision encoder (DINOv2) with PCA-based whitening feature refinement.

Result: Outperforms prior learning-based approaches on benchmark dataset under zero-shot settings without ground-truth supervision or finetuning. Enables automatic construction of semantically aligned street-to-satellite datasets.

Conclusion: Proposed training-free framework provides effective cross-view retrieval and scalable dataset construction, offering cost-efficient alternative to manual annotation.

Abstract: Cross-view image retrieval, particularly street-to-satellite matching, is a critical task for applications such as autonomous navigation, urban planning, and localization in GPS-denied environments. However, existing approaches often require supervised training on curated datasets and rely on panoramic or UAV-based images, which limits real-world deployment. In this paper, we present a simple yet effective cross-view image retrieval framework that leverages a pretrained vision encoder and a large language model (LLM), requiring no additional training. Given a monocular street-view image, our method extracts geographic cues through web-based image search and LLM-based location inference, generates a satellite query via geocoding API, and retrieves matching tiles using a pretrained vision encoder (e.g., DINOv2) with PCA-based whitening feature refinement. Despite using no ground-truth supervision or finetuning, our proposed method outperforms prior learning-based approaches on the benchmark dataset under zero-shot settings. Moreover, our pipeline enables automatic construction of semantically aligned street-to-satellite datasets, which is offering a scalable and cost-efficient alternative to manual annotation. All source codes will be made publicly available at https://jeonghomin.github.io/street2orbit.github.io/.

Method: Uses 3DGS as the 3D representation for both humans and scenes. Features a Gaussian-aligned motion module that synthesizes motion without explicit scene geometry using opacity-based cues and projected Gaussian structures. Includes human-scene Gaussian refinement optimization for realistic contact and navigation.

Result: Evaluated on scenes from Scannet++ and SuperSplat library, and avatars from sparse/dense multi-view human capture. Enables novel applications like geometry-consistent free-viewpoint rendering of edited monocular RGB videos with new animated humans.

Conclusion: The framework demonstrates the unique advantage of 3DGS for monocular video-based human animation, allowing decoupled rendering from motion synthesis without paired human-scene data.

Abstract: We present a novel framework for animating humans in 3D scenes using 3D Gaussian Splatting (3DGS), a neural scene representation that has recently achieved state-of-the-art photorealistic results for novel-view synthesis but remains under-explored for human-scene animation and interaction. Unlike existing animation pipelines that use meshes or point clouds as the underlying 3D representation, our approach introduces the use of 3DGS as the 3D representation to the problem of animating humans in scenes. By representing humans and scenes as Gaussians, our approach allows for geometry-consistent free-viewpoint rendering of humans interacting with 3D scenes. Our key insight is that the rendering can be decoupled from the motion synthesis and each sub-problem can be addressed independently, without the need for paired human-scene data. Central to our method is a Gaussian-aligned motion module that synthesizes motion without explicit scene geometry, using opacity-based cues and projected Gaussian structures to guide human placement and pose alignment. To ensure natural interactions, we further propose a human-scene Gaussian refinement optimization that enforces realistic contact and navigation. We evaluate our approach on scenes from Scannet++ and the SuperSplat library, and on avatars reconstructed from sparse and dense multi-view human capture. Finally, we demonstrate that our framework allows for novel applications such as geometry-consistent free-viewpoint rendering of edited monocular RGB videos with new animated humans, showcasing the unique advantage of 3DGS for monocular video-based human animation.

Method: Constructs a provably sufficient set of binary masks using a mathematically rigorous coverage strategy that ensures each patch location is covered at least k times, with single-round masking and O(n) complexity.

Result: Improves certified robust accuracy by up to +13.4% over PatchCleanser on ImageNet, ImageNette, and CIFAR-10 while maintaining clean accuracy nearly identical to the vanilla model.

Conclusion: CertMask provides an efficient and certifiably robust defense against adversarial patch attacks with strong theoretical guarantees and practical performance improvements.

Abstract: Adversarial patch attacks inject localized perturbations into images to mislead deep vision models. These attacks can be physically deployed, posing serious risks to real-world applications. In this paper, we propose CertMask, a certifiably robust defense that constructs a provably sufficient set of binary masks to neutralize patch effects with strong theoretical guarantees. While the state-of-the-art approach (PatchCleanser) requires two rounds of masking and incurs $O(n^2)$ inference cost, CertMask performs only a single round of masking with $O(n)$ time complexity, where $n$ is the cardinality of the mask set to cover an input image. Our proposed mask set is computed using a mathematically rigorous coverage strategy that ensures each possible patch location is covered at least $k$ times, providing both efficiency and robustness. We offer a theoretical analysis of the coverage condition and prove its sufficiency for certification. Experiments on ImageNet, ImageNette, and CIFAR-10 show that CertMask improves certified robust accuracy by up to +13.4% over PatchCleanser, while maintaining clean accuracy nearly identical to the vanilla model.

Method: Adapted a foundation model for solar physics to create embeddings, concatenated with spacecraft position and solar magnetic connectivity using Fourier features, forming a neural field-based model fine-tuned with Parker Solar Probe measurements.

Result: Overall classification performance is modest, likely due to coarse labeling, class imbalance, and limited transferability of the pretrained model.

Conclusion: Demonstrates feasibility of leveraging foundation model embeddings for in situ solar wind tasks and lays groundwork for future improvements in space weather predictions.

Abstract: Space weather at Earth, driven by the solar activity, poses growing risks to satellites around our planet as well as to critical ground-based technological infrastructure. Major space weather contributors are the solar wind and coronal mass ejections whose variable density, speed, temperature, and magnetic field make the automated classification of those structures challenging. In this work, we adapt a foundation model for solar physics, originally trained on Solar Dynamics Observatory imagery, to create embeddings suitable for solar wind structure analysis. These embeddings are concatenated with the spacecraft position and solar magnetic connectivity encoded using Fourier features which generates a neural field-based model. The full deep learning architecture is fine-tuned bridging the gap between remote sensing and in situ observations. Labels are derived from Parker Solar Probe measurements, forming a downstream classification task that maps plasma properties to solar wind structures. Although overall classification performance is modest, likely due to coarse labeling, class imbalance, and limited transferability of the pretrained model, this study demonstrates the feasibility of leveraging foundation model embeddings for in situ solar wind tasks. As a first proof-of-concept, it lays the groundwork for future improvements toward more reliable space weather predictions. The code and configuration files used in this study are publicly available to support reproducibility.

Method: Proposed IPCD-Net with point-wise feature aggregation for non-grid processing and Projection-based Luminance Distribution (PLD) with hierarchical feature refinement to capture global-light cues via multi-view projection.

Result: IPCD-Net reduces cast shadows in albedo and enhances color accuracy in shade, demonstrated on a synthetic outdoor-scene dataset. Applications shown in texture editing, relighting, and point-cloud registration under varying illumination.

Conclusion: IPCD-Net effectively decomposes point clouds into albedo and shade, verified for real-world applicability and enabling various downstream applications in AR and robotics.

Abstract: Point clouds are widely used in various fields, including augmented reality (AR) and robotics, where relighting and texture editing are crucial for realistic visualization. Achieving these tasks requires accurately separating albedo from shade. However, performing this separation on point clouds presents two key challenges: (1) the non-grid structure of point clouds makes conventional image-based decomposition models ineffective, and (2) point-cloud models designed for other tasks do not explicitly consider global-light direction, resulting in inaccurate shade. In this paper, we introduce \textbf{Intrinsic Point-Cloud Decomposition (IPCD)}, which extends image decomposition to the direct decomposition of colored point clouds into albedo and shade. To overcome challenge (1), we propose \textbf{IPCD-Net} that extends image-based model with point-wise feature aggregation for non-grid data processing. For challenge (2), we introduce \textbf{Projection-based Luminance Distribution (PLD)} with a hierarchical feature refinement, capturing global-light ques via multi-view projection. For comprehensive evaluation, we create a synthetic outdoor-scene dataset. Experimental results demonstrate that IPCD-Net reduces cast shadows in albedo and enhances color accuracy in shade. Furthermore, we showcase its applications in texture editing, relighting, and point-cloud registration under varying illumination. Finally, we verify the real-world applicability of IPCD-Net.

Method: Three-step Decay Resilience Strategies: (1) SD-DRS amplifies semantically meaningful distant signals, (2) DC-DRS modulates attention weights based on positional distances, and (3) reRD-DRS reinforces remaining informative remote dependencies.

Result: Extensive experiments on Vision Question Answering benchmarks show consistent performance improvements in a training-free manner.

Conclusion: T-DRS effectively mitigates attention decay in LVLMs and enhances global context modeling without requiring retraining.

Abstract: Large Vision-Language Models (LVLMs) have achieved impressive performance across a wide range of multimodal tasks. However, they still face critical challenges in modeling long-range dependencies under the usage of Rotary Positional Encoding (ROPE). Although it can facilitate precise modeling of token positions, it induces progressive attention decay as token distance increases, especially with progressive attention decay over distant token pairs, which severely impairs the model’s ability to remember global context. To alleviate this issue, we propose inference-only Three-step Decay Resilience Strategies (T-DRS), comprising (1) Semantic-Driven DRS (SD-DRS), amplifying semantically meaningful but distant signals via content-aware residuals, (2) Distance-aware Control DRS (DC-DRS), which can purify attention by smoothly modulating weights based on positional distances, suppressing noise while preserving locality, and (3) re-Reinforce Distant DRS (reRD-DRS), consolidating the remaining informative remote dependencies to maintain global coherence. Together, the T-DRS recover suppressed long-range token pairs without harming local inductive biases. Extensive experiments on Vision Question Answering (VQA) benchmarks demonstrate that T-DRS can consistently improve performance in a training-free manner. The code can be accessed in https://github.com/labixiaoq-qq/Remember-me

Method: Uses depth-guided parallel adapters in a multi-modal image encoder for prompt-free fine-tuning, and a query-driven temporal memory module that unifies memory bank and prompt embeddings using frame-level and video-level queries.

Result: Extensive experiments on three RGB-D VSOD datasets show SAM-DAQ consistently outperforms state-of-the-art methods across all evaluation metrics.

Conclusion: SAM-DAQ successfully adapts SAM2 for RGB-D VSOD by effectively integrating depth and temporal information while overcoming the limitations of manual prompts and high computational costs.

Abstract: Recently segment anything model (SAM) has attracted widespread concerns, and it is often treated as a vision foundation model for universal segmentation. Some researchers have attempted to directly apply the foundation model to the RGB-D video salient object detection (RGB-D VSOD) task, which often encounters three challenges, including the dependence on manual prompts, the high memory consumption of sequential adapters, and the computational burden of memory attention. To address the limitations, we propose a novel method, namely Segment Anything Model with Depth-guided Adaptive Queries (SAM-DAQ), which adapts SAM2 to pop-out salient objects from videos by seamlessly integrating depth and temporal cues within a unified framework. Firstly, we deploy a parallel adapter-based multi-modal image encoder (PAMIE), which incorporates several depth-guided parallel adapters (DPAs) in a skip-connection way. Remarkably, we fine-tune the frozen SAM encoder under prompt-free conditions, where the DPA utilizes depth cues to facilitate the fusion of multi-modal features. Secondly, we deploy a query-driven temporal memory (QTM) module, which unifies the memory bank and prompt embeddings into a learnable pipeline. Concretely, by leveraging both frame-level queries and video-level queries simultaneously, the QTM module can not only selectively extract temporal consistency features but also iteratively update the temporal representations of the queries. Extensive experiments are conducted on three RGB-D VSOD datasets, and the results show that the proposed SAM-DAQ consistently outperforms state-of-the-art methods in terms of all evaluation metrics.

Method: Apply semantic-preserving distortions to create degraded images, use both original and degraded images with prompts to condition LMM’s quality perception, and aggregate scores using conditional probability.

Result: Extensive experiments show consistent enhancement of LMM performance across various IQA datasets.

Conclusion: The proposed training-free debiasing framework effectively mitigates semantic bias in LMMs for improved image quality assessment without costly retraining.

Abstract: Despite the impressive performance of large multimodal models (LMMs) in high-level visual tasks, their capacity for image quality assessment (IQA) remains limited. One main reason is that LMMs are primarily trained for high-level tasks (e.g., image captioning), emphasizing unified image semantics extraction under varied quality. Such semantic-aware yet quality-insensitive perception bias inevitably leads to a heavy reliance on image semantics when those LMMs are forced for quality rating. In this paper, instead of retraining or tuning an LMM costly, we propose a training-free debiasing framework, in which the image quality prediction is rectified by mitigating the bias caused by image semantics. Specifically, we first explore several semantic-preserving distortions that can significantly degrade image quality while maintaining identifiable semantics. By applying these specific distortions to the query or test images, we ensure that the degraded images are recognized as poor quality while their semantics mainly remain. During quality inference, both a query image and its corresponding degraded version are fed to the LMM along with a prompt indicating that the query image quality should be inferred under the condition that the degraded one is deemed poor quality. This prior condition effectively aligns the LMM’s quality perception, as all degraded images are consistently rated as poor quality, regardless of their semantic variance. Finally, the quality scores of the query image inferred under different prior conditions (degraded versions) are aggregated using a conditional probability model. Extensive experiments on various IQA datasets show that our debiasing framework could consistently enhance the LMM performance.

Method: Proposes RWKV-PCSSC with RWKV Seed Generator to aggregate features from partial point clouds and RWKV Point Deconvolution modules to progressively restore point-wise features through multiple stages.

Result: Achieves state-of-the-art performance on indoor (SSC-PC, NYUCAD-PC) and outdoor (PointSSC) datasets, plus proposed datasets (NYUCAD-PC-V2, 3D-FRONT-PC), with 4.18× parameter reduction and 1.37× memory efficiency improvement.

Conclusion: The proposed lightweight RWKV-PCSSC network effectively addresses the limitations of dense architectures in semantic scene completion while maintaining high performance.

Abstract: Semantic Scene Completion (SSC) aims to generate a complete semantic scene from an incomplete input. Existing approaches often employ dense network architectures with a high parameter count, leading to increased model complexity and resource demands. To address these limitations, we propose RWKV-PCSSC, a lightweight point cloud semantic scene completion network inspired by the Receptance Weighted Key Value (RWKV) mechanism. Specifically, we introduce a RWKV Seed Generator (RWKV-SG) module that can aggregate features from a partial point cloud to produce a coarse point cloud with coarse features. Subsequently, the point-wise feature of the point cloud is progressively restored through multiple stages of the RWKV Point Deconvolution (RWKV-PD) modules. By leveraging a compact and efficient design, our method achieves a lightweight model representation. Experimental results demonstrate that RWKV-PCSSC reduces the parameter count by 4.18$\times$ and improves memory efficiency by 1.37$\times$ compared to state-of-the-art methods PointSSC. Furthermore, our network achieves state-of-the-art performance on established indoor (SSC-PC, NYUCAD-PC) and outdoor (PointSSC) scene dataset, as well as on our proposed datasets (NYUCAD-PC-V2, 3D-FRONT-PC).

Method: Extract feature embeddings using CNN-Transformer encoder, project into spherical space, approximate geodesic distances with Riemannian distances in bi-invariant SO(4) space, and use differentiable Levenberg-Marquardt optimization during inference.

Result: Experiments on real and synthetic datasets show superior accuracy in both patient-specific and patient-agnostic scenarios compared to existing methods.

Conclusion: The proposed spherical feature space approach with Riemannian distances provides a more expressive and geometrically consistent deep similarity metric, enhancing registration performance and convergence speed.

Abstract: Intraoperative 2D/3D registration aligns preoperative 3D volumes with real-time 2D radiographs, enabling accurate localization of instruments and implants. A recent fully differentiable similarity learning framework approximates geodesic distances on SE(3), expanding the capture range of registration and mitigating the effects of substantial disturbances, but existing Euclidean approximations distort manifold structure and slow convergence. To address these limitations, we explore similarity learning in non-Euclidean spherical feature spaces to better capture and fit complex manifold structure. We extract feature embeddings using a CNN-Transformer encoder, project them into spherical space, and approximate their geodesic distances with Riemannian distances in the bi-invariant SO(4) space. This enables a more expressive and geometrically consistent deep similarity metric, enhancing the ability to distinguish subtle pose differences. During inference, we replace gradient descent with fully differentiable Levenberg-Marquardt optimization to accelerate convergence. Experiments on real and synthetic datasets show superior accuracy in both patient-specific and patient-agnostic scenarios.

Method: Proposes Hierarchical Compensatory Compression with global structure compression (GSC) using global queries and adaptive detail mining (ADM) for selective recompression of salient features.

Result: Achieves approximately 98% compression ratio compared to previous 3D-VLMs while achieving state-of-the-art performance on 3D understanding tasks.

Conclusion: HCC-3D successfully reduces computational overhead while preserving essential 3D information, demonstrating improvements in both efficiency and performance.

Abstract: 3D understanding has drawn significant attention recently, leveraging Vision-Language Models (VLMs) to enable multi-modal reasoning between point cloud and text data. Current 3D-VLMs directly embed the 3D point clouds into 3D tokens, following large 2D-VLMs with powerful reasoning capabilities. However, this framework has a great computational cost limiting its application, where we identify that the bottleneck lies in processing all 3D tokens in the Large Language Model (LLM) part. This raises the question: how can we reduce the computational overhead introduced by 3D tokens while preserving the integrity of their essential information? To address this question, we introduce Hierarchical Compensatory Compression (HCC-3D) to efficiently compress 3D tokens while maintaining critical detail retention. Specifically, we first propose a global structure compression (GSC), in which we design global queries to compress all 3D tokens into a few key tokens while keeping overall structural information. Then, to compensate for the information loss in GSC, we further propose an adaptive detail mining (ADM) module that selectively recompresses salient but under-attended features through complementary scoring. Extensive experiments demonstrate that HCC-3D not only achieves extreme compression ratios (approximately 98%) compared to previous 3D-VLMs, but also achieves new state-of-the-art performance, showing the great improvements on both efficiency and performance.

Method: Proposes SARL with cross-scale spatial-channel attention to enrich features and strengthen cross-layer sharing, and SAL that reshapes IoU-based losses to assign lower weights to larger objects, focusing training on tiny objects.

Result: Boosts generalization ability by 5.5% AP when embedded in YOLOv5 and YOLOx baselines, and achieves 29.0% AP on real-world noisy dataset AI-TOD-v2.0.

Conclusion: The proposed SARL and SAL effectively address tiny object detection challenges in aerial images and are compatible with top-performing frameworks, demonstrating significant performance improvements across multiple benchmarks.

Abstract: Recently, despite the remarkable advancements in object detection, modern detectors still struggle to detect tiny objects in aerial images. One key reason is that tiny objects carry limited features that are inevitably degraded or lost during long-distance network propagation. Another is that smaller objects receive disproportionately greater regression penalties than larger ones during training. To tackle these issues, we propose a Scale-Aware Relay Layer (SARL) and a Scale-Adaptive Loss (SAL) for tiny object detection, both of which are seamlessly compatible with the top-performing frameworks. Specifically, SARL employs a cross-scale spatial-channel attention to progressively enrich the meaningful features of each layer and strengthen the cross-layer feature sharing. SAL reshapes the vanilla IoU-based losses so as to dynamically assign lower weights to larger objects. This loss is able to focus training on tiny objects while reducing the influence on large objects. Extensive experiments are conducted on three benchmarks (\textit{i.e.,} AI-TOD, DOTA-v2.0 and VisDrone2019), and the results demonstrate that the proposed method boosts the generalization ability by 5.5% Average Precision (AP) when embedded in YOLOv5 (anchor-based) and YOLOx (anchor-free) baselines. Moreover, it also promotes the robust performance with 29.0% AP on the real-world noisy dataset (\textit{i.e.,} AI-TOD-v2.0).

Method: Swin-BART encoder-decoder architecture with lightweight regional attention module that amplifies diagnostically salient regions before cross-attention, trained on ROCO dataset with beam search decoding.

Result: Achieves SOTA performance: ROUGE 0.603 (vs 0.356/0.255 baselines), BERTScore 0.807 (vs 0.645/0.623 baselines), with competitive BLEU, CIDEr, and METEOR scores. Provides interpretable heatmaps showing regional attributions.

Conclusion: The proposed system yields accurate, clinically phrased captions with transparent regional attributions, supporting safe research use with human oversight in medical reporting workflows.

Abstract: Automated medical image captioning translates complex radiological images into diagnostic narratives that can support reporting workflows. We present a Swin-BART encoder-decoder system with a lightweight regional attention module that amplifies diagnostically salient regions before cross-attention. Trained and evaluated on ROCO, our model achieves state-of-the-art semantic fidelity while remaining compact and interpretable. We report results as mean$\pm$std over three seeds and include $95%$ confidence intervals. Compared with baselines, our approach improves ROUGE (proposed 0.603, ResNet-CNN 0.356, BLIP2-OPT 0.255) and BERTScore (proposed 0.807, BLIP2-OPT 0.645, ResNet-CNN 0.623), with competitive BLEU, CIDEr, and METEOR. We further provide ablations (regional attention on/off and token-count sweep), per-modality analysis (CT/MRI/X-ray), paired significance tests, and qualitative heatmaps that visualize the regions driving each description. Decoding uses beam search (beam size $=4$), length penalty $=1.1$, $no_repeat_ngram_size$ $=3$, and max length $=128$. The proposed design yields accurate, clinically phrased captions and transparent regional attributions, supporting safe research use with a human in the loop.

Method: Uses controllable Gaussian noise injection and multi-scale Gaussian blurring for initial feature perturbations, followed by temporal modeling and liquid parameter adjustment mechanism to generate adaptive modulation parameters for smooth continuous domain adaptation.

Result: Significant performance improvements on Diverse Weather dataset and Real-to-Art benchmark, demonstrating superior generalization and robustness to unseen domain shifts.

Conclusion: The proposed method effectively bridges source-unknown domain distribution gap by capturing progressive cross-domain feature evolution and dynamically regulating adaptation paths, significantly boosting generalization capabilities.

Abstract: In this paper, we focus on Single-Domain Generalized Object Detection (Single-DGOD), aiming to transfer a detector trained on one source domain to multiple unknown domains. Existing methods for Single-DGOD typically rely on discrete data augmentation or static perturbation methods to expand data diversity, thereby mitigating the lack of access to target domain data. However, in real-world scenarios such as changes in weather or lighting conditions, domain shifts often occur continuously and gradually. Discrete augmentations and static perturbations fail to effectively capture the dynamic variation of feature distributions, thereby limiting the model’s ability to perceive fine-grained cross-domain differences. To this end, we propose a new method, Liquid Temporal Feature Evolution, which simulates the progressive evolution of features from the source domain to simulated latent distributions by incorporating temporal modeling and liquid neural network-driven parameter adjustment. Specifically, we introduce controllable Gaussian noise injection and multi-scale Gaussian blurring to simulate initial feature perturbations, followed by temporal modeling and a liquid parameter adjustment mechanism to generate adaptive modulation parameters, enabling a smooth and continuous adaptation across domains. By capturing progressive cross-domain feature evolution and dynamically regulating adaptation paths, our method bridges the source-unknown domain distribution gap, significantly boosting generalization and robustness to unseen shifts. Significant performance improvements on the Diverse Weather dataset and Real-to-Art benchmark demonstrate the superiority of our method. Our code is available at https://github.com/2490o/LTFE.

Method: Developed DocWeaver, a multi-agent pipeline using Large Language Models to automatically generate MosaicDoc - a bilingual benchmark with 72K images from newspapers/magazines featuring diverse layouts and comprehensive annotations.

Result: Created MosaicDoc with 72K images, 600K+ QA pairs, diverse complex layouts from 196 publishers, and multi-task annotations (OCR, VQA, reading order, localization). Evaluation revealed current models’ limitations in handling real-world document complexity.

Conclusion: MosaicDoc serves as a definitive VRDU benchmark that exposes current model limitations and provides clear direction for future research in handling complex document understanding tasks.

Abstract: Despite the rapid progress of Vision-Language Models (VLMs), their capabilities are inadequately assessed by existing benchmarks, which are predominantly English-centric, feature simplistic layouts, and support limited tasks. Consequently, they fail to evaluate model performance for Visually Rich Document Understanding (VRDU), a critical challenge involving complex layouts and dense text. To address this, we introduce DocWeaver, a novel multi-agent pipeline that leverages Large Language Models to automatically generate a new benchmark. The result is MosaicDoc, a large-scale, bilingual (Chinese and English) resource designed to push the boundaries of VRDU. Sourced from newspapers and magazines, MosaicDoc features diverse and complex layouts (including multi-column and non-Manhattan), rich stylistic variety from 196 publishers, and comprehensive multi-task annotations (OCR, VQA, reading order, and localization). With 72K images and over 600K QA pairs, MosaicDoc serves as a definitive benchmark for the field. Our extensive evaluation of state-of-the-art models on this benchmark reveals their current limitations in handling real-world document complexity and charts a clear path for future research.

Method: Proposes SLDC with linear and weakly nonlinear variants that learn transition operators to align feature distributions across tasks, combined with knowledge distillation to reduce representation drift.

Result: Extensive experiments show SLDC significantly improves SeqFT performance, achieving comparable results to joint training across all evaluated datasets when combined with KD.

Conclusion: SLDC effectively mitigates distribution drift in sequential fine-tuning, making SeqFT a viable alternative to joint training for class-incremental learning.

Abstract: Recent advances have shown that sequential fine-tuning (SeqFT) of pre-trained vision transformers (ViTs), followed by classifier refinement using approximate distributions of class features, can be an effective strategy for class-incremental learning (CIL). However, this approach is susceptible to distribution drift, caused by the sequential optimization of shared backbone parameters. This results in a mismatch between the distributions of the previously learned classes and that of the updater model, ultimately degrading the effectiveness of classifier performance over time. To address this issue, we introduce a latent space transition operator and propose Sequential Learning with Drift Compensation (SLDC). SLDC aims to align feature distributions across tasks to mitigate the impact of drift. First, we present a linear variant of SLDC, which learns a linear operator by solving a regularized least-squares problem that maps features before and after fine-tuning. Next, we extend this with a weakly nonlinear SLDC variant, which assumes that the ideal transition operator lies between purely linear and fully nonlinear transformations. This is implemented using learnable, weakly nonlinear mappings that balance flexibility and generalization. To further reduce representation drift, we apply knowledge distillation (KD) in both algorithmic variants. Extensive experiments on standard CIL benchmarks demonstrate that SLDC significantly improves the performance of SeqFT. Notably, by combining KD to address representation drift with SLDC to compensate distribution drift, SeqFT achieves performance comparable to joint training across all evaluated datasets. Code: https://github.com/raoxuan98-hash/sldc.git.

Method: Two-phase framework: 1) Data balancing using diffusion-model-based resampling for fairness and diversity, 2) Bi-adversarial self-meta defense with metric adversarial training and farthest negative extension softening, plus adversarially-enhanced self-meta mechanism.

Result: Experiments show the method significantly outperforms existing state-of-the-art defenses for person ReID.

Conclusion: The proposed framework effectively addresses adversarial defense challenges in person ReID through debiasing and dual-invariant mechanisms, achieving superior robustness.

Abstract: Person re-identification (ReID) is a fundamental task in many real-world applications such as pedestrian trajectory tracking. However, advanced deep learning-based ReID models are highly susceptible to adversarial attacks, where imperceptible perturbations to pedestrian images can cause entirely incorrect predictions, posing significant security threats. Although numerous adversarial defense strategies have been proposed for classification tasks, their extension to metric learning tasks such as person ReID remains relatively unexplored. Moreover, the several existing defenses for person ReID fail to address the inherent unique challenges of adversarially robust ReID. In this paper, we systematically identify the challenges of adversarial defense in person ReID into two key issues: model bias and composite generalization requirements. To address them, we propose a debiased dual-invariant defense framework composed of two main phases. In the data balancing phase, we mitigate model bias using a diffusion-model-based data resampling strategy that promotes fairness and diversity in training data. In the bi-adversarial self-meta defense phase, we introduce a novel metric adversarial training approach incorporating farthest negative extension softening to overcome the robustness degradation caused by the absence of classifier. Additionally, we introduce an adversarially-enhanced self-meta mechanism to achieve dual-generalization for both unseen identities and unseen attack types. Experiments demonstrate that our method significantly outperforms existing state-of-the-art defenses.

Method: Proposes Adaptive Graph Convolution with a static axial scaffold and dynamic Exponential Decay Gating that selectively weighs long-range connections based on feature similarity, plus a hybrid strategy using gating in early stages and Global Attention in the final stage.

Result: Achieves SOTA trade-off between accuracy and efficiency: AdaptViG-M gets 82.6% top-1 accuracy (outperforming ViG-B by 0.3%) with 80% fewer parameters and 84% fewer GMACs. On downstream tasks, it surpasses EfficientFormer-L7 with 78% fewer parameters.

Conclusion: AdaptViG provides an efficient and powerful hybrid Vision GNN that significantly reduces computational costs while achieving superior performance compared to larger models.

Abstract: Vision Graph Neural Networks (ViGs) offer a new direction for advancements in vision architectures. While powerful, ViGs often face substantial computational challenges stemming from their graph construction phase, which can hinder their efficiency. To address this issue we propose AdaptViG, an efficient and powerful hybrid Vision GNN that introduces a novel graph construction mechanism called Adaptive Graph Convolution. This mechanism builds upon a highly efficient static axial scaffold and a dynamic, content-aware gating strategy called Exponential Decay Gating. This gating mechanism selectively weighs long-range connections based on feature similarity. Furthermore, AdaptViG employs a hybrid strategy, utilizing our efficient gating mechanism in the early stages and a full Global Attention block in the final stage for maximum feature aggregation. Our method achieves a new state-of-the-art trade-off between accuracy and efficiency among Vision GNNs. For instance, our AdaptViG-M achieves 82.6% top-1 accuracy, outperforming ViG-B by 0.3% while using 80% fewer parameters and 84% fewer GMACs. On downstream tasks, AdaptViG-M obtains 45.8 mIoU, 44.8 APbox, and 41.1 APmask, surpassing the much larger EfficientFormer-L7 by 0.7 mIoU, 2.2 APbox, and 2.1 APmask, respectively, with 78% fewer parameters.

Method: Uniformly samples transmittance to model pixel-wise multi-modal distribution of opacity and depth, progressively fuses truncated signed distance functions to reconstruct external and internal surfaces separately.

Result: Significantly improves semi-transparent geometry reconstruction while maintaining performance on opaque scenes, works without extra training overhead.

Conclusion: TSPE-GS provides an effective solution for semi-transparent surface reconstruction in Gaussian Splatting frameworks, resolving cross-surface depth ambiguity through probabilistic modeling.

Abstract: 3D Gaussian Splatting offers a strong speed-quality trade-off but struggles to reconstruct semi-transparent surfaces because most methods assume a single depth per pixel, which fails when multiple surfaces are visible. We propose TSPE-GS (Transparent Surface Probabilistic Extraction for Gaussian Splatting), which uniformly samples transmittance to model a pixel-wise multi-modal distribution of opacity and depth, replacing the prior single-peak assumption and resolving cross-surface depth ambiguity. By progressively fusing truncated signed distance functions, TSPE-GS reconstructs external and internal surfaces separately within a unified framework. The method generalizes to other Gaussian-based reconstruction pipelines without extra training overhead. Extensive experiments on public and self-collected semi-transparent and opaque datasets show TSPE-GS significantly improves semi-transparent geometry reconstruction while maintaining performance on opaque scenes.

Method: Proposes an adaptive fusion framework that extracts absolute CLIP image features, applies Box-Cox transformation for statistical normalization, and uses confidence-guided fusion to combine magnitude cues with cosine similarity.

Result: Extensive experiments on multiple benchmark IQA datasets show consistent outperformance over standard CLIP-based IQA and state-of-the-art baselines.

Conclusion: The magnitude of CLIP image features provides valuable quality cues that, when properly normalized and fused with semantic similarity, significantly improve no-reference image quality assessment without requiring additional training.

Abstract: Recent efforts have repurposed the Contrastive Language-Image Pre-training (CLIP) model for No-Reference Image Quality Assessment (NR-IQA) by measuring the cosine similarity between the image embedding and textual prompts such as “a good photo” or “a bad photo.” However, this semantic similarity overlooks a critical yet underexplored cue: the magnitude of the CLIP image features, which we empirically find to exhibit a strong correlation with perceptual quality. In this work, we introduce a novel adaptive fusion framework that complements cosine similarity with a magnitude-aware quality cue. Specifically, we first extract the absolute CLIP image features and apply a Box-Cox transformation to statistically normalize the feature distribution and mitigate semantic sensitivity. The resulting scalar summary serves as a semantically-normalized auxiliary cue that complements cosine-based prompt matching. To integrate both cues effectively, we further design a confidence-guided fusion scheme that adaptively weighs each term according to its relative strength. Extensive experiments on multiple benchmark IQA datasets demonstrate that our method consistently outperforms standard CLIP-based IQA and state-of-the-art baselines, without any task-specific training.

Method: Leverage VLMs to generate pseudo-labels, then use per-object co-teaching where two YOLO models collaboratively filter noisy bounding boxes based on peer loss values during training.

Result: Achieved mAP@0.5 boost from 31.12% to 46.61% on KITTI dataset while maintaining real-time latency. With 10% ground truth labels, reached 57.97% mAP@0.5. Similar improvements on ACDC and BDD100k datasets.

Conclusion: The pipeline provides an efficient, robust, and scalable approach to train high-performance object detectors for autonomous driving, significantly reducing reliance on costly human annotation.

Abstract: Foundation models, especially vision-language models (VLMs), offer compelling zero-shot object detection for applications like autonomous driving, a domain where manual labelling is prohibitively expensive. However, their detection latency and tendency to hallucinate predictions render them unsuitable for direct deployment. This work introduces a novel pipeline that addresses this challenge by leveraging VLMs to automatically generate pseudo-labels for training efficient, real-time object detectors. Our key innovation is a per-object co-teaching-based training strategy that mitigates the inherent noise in VLM-generated labels. The proposed per-object coteaching approach filters noisy bounding boxes from training instead of filtering the entire image. Specifically, two YOLO models learn collaboratively, filtering out unreliable boxes from each mini-batch based on their peers’ per-object loss values. Overall, our pipeline provides an efficient, robust, and scalable approach to train high-performance object detectors for autonomous driving, significantly reducing reliance on costly human annotation. Experimental results on the KITTI dataset demonstrate that our method outperforms a baseline YOLOv5m model, achieving a significant mAP@0.5 boost ($31.12%$ to $46.61%$) while maintaining real-time detection latency. Furthermore, we show that supplementing our pseudo-labelled data with a small fraction of ground truth labels ($10%$) leads to further performance gains, reaching $57.97%$ mAP@0.5 on the KITTI dataset. We observe similar performance improvements for the ACDC and BDD100k datasets.

Method: Proposed EquS method that imposes equivariant information through dual sampling trajectories. Enhanced with EquS+ that includes Timestep-Aware Schedule (TAS) to prioritize deterministic steps for improved certainty and sampling efficiency.

Result: Extensive experiments on benchmarks show the method is compatible with previous problem-agnostic DMIR methods and significantly boosts their performance without increasing computational costs.

Conclusion: The proposed EquS and EquS+ methods effectively address limitations in current diffusion model-based image restoration approaches by better leveraging diffusion priors through dual sampling trajectories and optimized scheduling.

Abstract: Recent advances in generative models, especially diffusion models, have significantly improved image restoration (IR) performance. However, existing problem-agnostic diffusion model-based image restoration (DMIR) methods face challenges in fully leveraging diffusion priors, resulting in suboptimal performance. In this paper, we address the limitations of current problem-agnostic DMIR methods by analyzing their sampling process and providing effective solutions. We introduce EquS, a DMIR method that imposes equivariant information through dual sampling trajectories. To further boost EquS, we propose the Timestep-Aware Schedule (TAS) and introduce EquS$^+$. TAS prioritizes deterministic steps to enhance certainty and sampling efficiency. Extensive experiments on benchmarks demonstrate that our method is compatible with previous problem-agnostic DMIR methods and significantly boosts their performance without increasing computational costs. Our code is available at https://github.com/FouierL/EquS.

Method: DiVE constrains difference vectors (obtained by subtracting pre-trained and fine-tuned embeddings) to be equal across data samples using two losses: Average Vector Loss (AVL) for global geometric structure preservation and Pairwise Vector Loss (PVL) for local multimodal alignment consistency.

Result: Experiments show DiVE effectively preserves geometric structure and achieves strong performance across in-distribution, out-of-distribution, and zero-shot metrics.

Conclusion: DiVE successfully fine-tunes vision-language models on in-distribution data without compromising their generalization abilities in OOD and zero-shot settings by preserving the geometric structure of embeddings.

Abstract: Contrastive pre-trained vision-language models, such as CLIP, demonstrate strong generalization abilities in zero-shot classification by leveraging embeddings extracted from image and text encoders. This paper aims to robustly fine-tune these vision-language models on in-distribution (ID) data without compromising their generalization abilities in out-of-distribution (OOD) and zero-shot settings. Current robust fine-tuning methods tackle this challenge by reusing contrastive learning, which was used in pre-training, for fine-tuning. However, we found that these methods distort the geometric structure of the embeddings, which plays a crucial role in the generalization of vision-language models, resulting in limited OOD and zero-shot performance. To address this, we propose Difference Vector Equalization (DiVE), which preserves the geometric structure during fine-tuning. The idea behind DiVE is to constrain difference vectors, each of which is obtained by subtracting the embeddings extracted from the pre-trained and fine-tuning models for the same data sample. By constraining the difference vectors to be equal across various data samples, we effectively preserve the geometric structure. Therefore, we introduce two losses: average vector loss (AVL) and pairwise vector loss (PVL). AVL preserves the geometric structure globally by constraining difference vectors to be equal to their weighted average. PVL preserves the geometric structure locally by ensuring a consistent multimodal alignment. Our experiments demonstrate that DiVE effectively preserves the geometric structure, achieving strong results across ID, OOD, and zero-shot metrics.

Method: Proposes STELLAR with language-adaptive glyph encoder, multi-stage training (pre-training on synthetic data then fine-tuning on real images), and constructs STIPLAR dataset for training/evaluation. Introduces Text Appearance Similarity (TAS) metric.

Result: Outperforms state-of-the-art models in visual consistency and recognition accuracy, achieving 2.2% average TAS improvement across languages over baselines.

Conclusion: STELLAR effectively addresses key limitations in scene text editing through its multilingual capabilities, real-world adaptation strategy, and robust evaluation framework.

Abstract: Scene Text Editing (STE) is the task of modifying text content in an image while preserving its visual style, such as font, color, and background. While recent diffusion-based approaches have shown improvements in visual quality, key limitations remain: lack of support for low-resource languages, domain gap between synthetic and real data, and the absence of appropriate metrics for evaluating text style preservation. To address these challenges, we propose STELLAR (Scene Text Editor for Low-resource LAnguages and Real-world data). STELLAR enables reliable multilingual editing through a language-adaptive glyph encoder and a multi-stage training strategy that first pre-trains on synthetic data and then fine-tunes on real images. We also construct a new dataset, STIPLAR(Scene Text Image Pairs of Low-resource lAnguages and Real-world data), for training and evaluation. Furthermore, we propose Text Appearance Similarity (TAS), a novel metric that assesses style preservation by independently measuring font, color, and background similarity, enabling robust evaluation even without ground truth. Experimental results demonstrate that STELLAR outperforms state-of-the-art models in visual consistency and recognition accuracy, achieving an average TAS improvement of 2.2% across languages over the baselines.

Method: Systematic material selection (TiO2 for high reflectivity) and novel simulation pipeline with angle-independent BRDF approximation and distance-aware scaling for realistic LiDAR intensity modeling.

Result: Achieves 93.50% attack success rate on state-of-the-art LiDAR and fusion models, outperforming prior methods by over 41%.

Conclusion: Reveals new physically realizable threats and underscores need for material-aware defenses in real-world environments.

Abstract: LiDAR-based 3D object detection is widely used in safety-critical systems. However, these systems remain vulnerable to backdoor attacks that embed hidden malicious behaviors during training. A key limitation of existing backdoor attacks is their lack of physical realizability, primarily due to the digital-to-physical domain gap. Digital triggers often fail in real-world settings because they overlook material-dependent LiDAR reflection properties. On the other hand, physically constructed triggers are often unoptimized, leading to low effectiveness or easy detectability.This paper introduces Material-Oriented Backdoor Attack (MOBA), a novel framework that bridges the digital-physical gap by explicitly modeling the material properties of real-world triggers. MOBA tackles two key challenges in physical backdoor design: 1) robustness of the trigger material under diverse environmental conditions, 2) alignment between the physical trigger’s behavior and its digital simulation. First, we propose a systematic approach to selecting robust trigger materials, identifying titanium dioxide (TiO_2) for its high diffuse reflectivity and environmental resilience. Second, to ensure the digital trigger accurately mimics the physical behavior of the material-based trigger, we develop a novel simulation pipeline that features: (1) an angle-independent approximation of the Oren-Nayar BRDF model to generate realistic LiDAR intensities, and (2) a distance-aware scaling mechanism to maintain spatial consistency across varying depths. We conduct extensive experiments on state-of-the-art LiDAR-based and Camera-LiDAR fusion models, showing that MOBA achieves a 93.50% attack success rate, outperforming prior methods by over 41%. Our work reveals a new class of physically realizable threats and underscores the urgent need for defenses that account for material-level properties in real-world environments.

Method: Two-stage framework: 1) Dual-Branch Point Grouping module generates pseudo labels using semantic and mask cues from multi-view images, with refinement via Granularity-Aware Instance Merging and Semantic Selection and Propagation. 2) Multi-round self-training on end-to-end instance segmentation network with Instance Mask Filter to handle pseudo label inconsistencies.

Result: Achieves competitive performance compared to sparse-point-level supervised 3D instance segmentation methods and surpasses state-of-the-art scene-level supervised 3D semantic segmentation approaches.

Conclusion: DBGroup provides an effective weakly supervised solution for 3D instance segmentation that reduces annotation costs while maintaining competitive performance through scene-level annotations and sophisticated pseudo-label generation and refinement strategies.

Abstract: Weakly supervised 3D instance segmentation is essential for 3D scene understanding, especially as the growing scale of data and high annotation costs associated with fully supervised approaches. Existing methods primarily rely on two forms of weak supervision: one-thing-one-click annotations and bounding box annotations, both of which aim to reduce labeling efforts. However, these approaches still encounter limitations, including labor-intensive annotation processes, high complexity, and reliance on expert annotators. To address these challenges, we propose \textbf{DBGroup}, a two-stage weakly supervised 3D instance segmentation framework that leverages scene-level annotations as a more efficient and scalable alternative. In the first stage, we introduce a Dual-Branch Point Grouping module to generate pseudo labels guided by semantic and mask cues extracted from multi-view images. To further improve label quality, we develop two refinement strategies: Granularity-Aware Instance Merging and Semantic Selection and Propagation. The second stage involves multi-round self-training on an end-to-end instance segmentation network using the refined pseudo-labels. Additionally, we introduce an Instance Mask Filter strategy to address inconsistencies within the pseudo labels. Extensive experiments demonstrate that DBGroup achieves competitive performance compared to sparse-point-level supervised 3D instance segmentation methods, while surpassing state-of-the-art scene-level supervised 3D semantic segmentation approaches. Code is available at https://github.com/liuxuexun/DBGroup.

Method: LampQ performs layer-wise quantization with fine-grained control, uses a type-aware Fisher-based metric to measure sensitivity, assigns bit-widths optimally through integer linear programming, and iteratively updates them.

Result: Extensive experiments show LampQ achieves state-of-the-art performance in quantizing ViTs for various tasks including image classification, object detection, and zero-shot quantization.

Conclusion: LampQ overcomes key limitations in Vision Transformer quantization by providing fine-grained layer-wise control, accurate sensitivity measurement, and optimal bit allocation, delivering superior compression performance across multiple vision tasks.

Abstract: How can we accurately quantize a pre-trained Vision Transformer model? Quantization algorithms compress Vision Transformers (ViTs) into low-bit formats, reducing memory and computation demands with minimal accuracy degradation. However, existing methods rely on uniform precision, ignoring the diverse sensitivity of ViT components to quantization. Metric-based Mixed Precision Quantization (MPQ) is a promising alternative, but previous MPQ methods for ViTs suffer from three major limitations: 1) coarse granularity, 2) mismatch in metric scale across component types, and 3) quantization-unaware bit allocation. In this paper, we propose LampQ (Layer-wise Mixed Precision Quantization for Vision Transformers), an accurate metric-based MPQ method for ViTs to overcome these limitations. LampQ performs layer-wise quantization to achieve both fine-grained control and efficient acceleration, incorporating a type-aware Fisher-based metric to measure sensitivity. Then, LampQ assigns bit-widths optimally through integer linear programming and further updates them iteratively. Extensive experiments show that LampQ provides the state-of-the-art performance in quantizing ViTs pre-trained on various tasks such as image classification, object detection, and zero-shot quantization.

Method: Integrates self-supervised masked autoencoder (MAE) for visual representation learning, graph attention networks (GAT) for modeling label correlations, constraint-aware optimization using KL divergence and regularization, and imbalance mitigation via asymmetric loss and boosting ensembles.

Result: Achieves state-of-the-art performance on the TongueAtlas-4K benchmark (4,000 images with 22 diagnostic labels), which represents the largest public dataset in tongue analysis.

Conclusion: While optimized for tongue diagnosis, the MIRNet framework demonstrates strong generalization potential for broader diagnostic medical imaging tasks by effectively addressing key challenges in medical image analysis.

Abstract: Automated interpretation of medical images demands robust modeling of complex visual-semantic relationships while addressing annotation scarcity, label imbalance, and clinical plausibility constraints. We introduce MIRNet (Medical Image Reasoner Network), a novel framework that integrates self-supervised pre-training with constrained graph-based reasoning. Tongue image diagnosis is a particularly challenging domain that requires fine-grained visual and semantic understanding. Our approach leverages self-supervised masked autoencoder (MAE) to learn transferable visual representations from unlabeled data; employs graph attention networks (GAT) to model label correlations through expert-defined structured graphs; enforces clinical priors via constraint-aware optimization using KL divergence and regularization losses; and mitigates imbalance using asymmetric loss (ASL) and boosting ensembles. To address annotation scarcity, we also introduce TongueAtlas-4K, a comprehensive expert-curated benchmark comprising 4,000 images annotated with 22 diagnostic labels–representing the largest public dataset in tongue analysis. Validation shows our method achieves state-of-the-art performance. While optimized for tongue diagnosis, the framework readily generalizes to broader diagnostic medical imaging tasks.

Method: AffordBot framework integrates MLLMs with chain-of-thought reasoning, renders surround-view images, projects 3D elements to 2D views, and uses active perception for viewpoint selection.

Result: Achieves state-of-the-art performance on SceneFun3D dataset with strong generalization using only 3D point cloud input and MLLMs.

Conclusion: The approach enables physically grounded reasoning for fine-grained 3D embodied tasks through effective integration of MLLMs and 3D scene understanding.

Abstract: Effective human-agent collaboration in physical environments requires understanding not only what to act upon, but also where the actionable elements are and how to interact with them. Existing approaches often operate at the object level or disjointedly handle fine-grained affordance reasoning, lacking coherent, instruction-driven grounding and reasoning. In this work, we introduce a new task: Fine-grained 3D Embodied Reasoning, which requires an agent to predict, for each referenced affordance element in a 3D scene, a structured triplet comprising its spatial location, motion type, and motion axis, based on a task instruction. To solve this task, we propose AffordBot, a novel framework that integrates Multimodal Large Language Models (MLLMs) with a tailored chain-of-thought (CoT) reasoning paradigm. To bridge the gap between 3D input and 2D-compatible MLLMs, we render surround-view images of the scene and project 3D element candidates into these views, forming a rich visual representation aligned with the scene geometry. Our CoT pipeline begins with an active perception stage, prompting the MLLM to select the most informative viewpoint based on the instruction, before proceeding with step-by-step reasoning to localize affordance elements and infer plausible interaction motions. Evaluated on the SceneFun3D dataset, AffordBot achieves state-of-the-art performance, demonstrating strong generalization and physically grounded reasoning with only 3D point cloud input and MLLMs.

Method: Uses crossmodal prompt encoding to unify visual and textual cues, inpainting-based generation pipeline, and contrastive refinement for precise anomaly-mask alignment. Trained on AnomVerse dataset with 12,987 anomaly-mask-caption triplets.

Result: Generates more realistic and varied anomalies than prior methods, yielding superior improvements in downstream anomaly detection. Can generate anomalies for any normal-category image using user-defined prompts.

Conclusion: Anomagic establishes a versatile foundation model for anomaly generation that outperforms existing methods and enhances anomaly detection performance.

Abstract: We propose Anomagic, a zero-shot anomaly generation method that produces semantically coherent anomalies without requiring any exemplar anomalies. By unifying both visual and textual cues through a crossmodal prompt encoding scheme, Anomagic leverages rich contextual information to steer an inpainting-based generation pipeline. A subsequent contrastive refinement strategy enforces precise alignment between synthesized anomalies and their masks, thereby bolstering downstream anomaly detection accuracy. To facilitate training, we introduce AnomVerse, a collection of 12,987 anomaly-mask-caption triplets assembled from 13 publicly available datasets, where captions are automatically generated by multimodal large language models using structured visual prompts and template-based textual hints. Extensive experiments demonstrate that Anomagic trained on AnomVerse can synthesize more realistic and varied anomalies than prior methods, yielding superior improvements in downstream anomaly detection. Furthermore, Anomagic can generate anomalies for any normal-category image using user-defined prompts, establishing a versatile foundation model for anomaly generation.

Method: Proposes DGFusion with Difficulty-aware Instance Pair Matcher (DIPM) that performs instance-level feature matching based on difficulty to generate easy and hard instance pairs, and Dual-guided Modules that exploit both pair types for effective multi-modal feature fusion.

Result: Outperforms baseline methods with +1.0% mAP, +0.8% NDS, and +1.3% average recall on nuScenes dataset, showing consistent robustness gains for hard instance detection across various scenarios.

Conclusion: The dual-guided paradigm effectively addresses limitations of single-guided approaches for multi-modal 3D object detection, particularly improving performance on challenging hard instances critical for autonomous driving safety.

Abstract: As a critical task in autonomous driving perception systems, 3D object detection is used to identify and track key objects, such as vehicles and pedestrians. However, detecting distant, small, or occluded objects (hard instances) remains a challenge, which directly compromises the safety of autonomous driving systems. We observe that existing multi-modal 3D object detection methods often follow a single-guided paradigm, failing to account for the differences in information density of hard instances between modalities. In this work, we propose DGFusion, based on the Dual-guided paradigm, which fully inherits the advantages of the Point-guide-Image paradigm and integrates the Image-guide-Point paradigm to address the limitations of the single paradigms. The core of DGFusion, the Difficulty-aware Instance Pair Matcher (DIPM), performs instance-level feature matching based on difficulty to generate easy and hard instance pairs, while the Dual-guided Modules exploit the advantages of both pair types to enable effective multi-modal feature fusion. Experimental results demonstrate that our DGFusion outperforms the baseline methods, with respective improvements of +1.0% mAP, +0.8% NDS, and +1.3% average recall on nuScenes. Extensive experiments demonstrate consistent robustness gains for hard instance detection across ego-distance, size, visibility, and small-scale training scenarios.

Method: Propose a 3D VAE framework based on UDFs with local-to-global architecture: partition UDF into uniform subvolumes (UBlocks), use 3D convolutions for local detail and sparse transformers for global coherence, and employ Pad-Average strategy for smooth boundary transitions.

Result: Achieves state-of-the-art performance in reconstruction accuracy and generative quality, with superior surface smoothness and geometric flexibility, enabling scaling to ultra-high resolutions up to 2048^3.

Conclusion: The proposed UDF-based 3D VAE with local-to-global architecture provides a robust and computationally efficient solution for generating high-fidelity 3D content with complex topologies, overcoming limitations of previous methods.

Abstract: Generating high-fidelity 3D contents remains a fundamental challenge due to the complexity of representing arbitrary topologies-such as open surfaces and intricate internal structures-while preserving geometric details. Prevailing methods based on signed distance fields (SDFs) are hampered by costly watertight preprocessing and struggle with non-manifold geometries, while point-cloud representations often suffer from sampling artifacts and surface discontinuities. To overcome these limitations, we propose a novel 3D variational autoencoder (VAE) framework built upon unsigned distance fields (UDFs)-a more robust and computationally efficient representation that naturally handles complex and incomplete shapes. Our core innovation is a local-to-global (LoG) architecture that processes the UDF by partitioning it into uniform subvolumes, termed UBlocks. This architecture couples 3D convolutions for capturing local detail with sparse transformers for enforcing global coherence. A Pad-Average strategy further ensures smooth transitions at subvolume boundaries during reconstruction. This modular design enables seamless scaling to ultra-high resolutions up to 2048^3-a regime previously unattainable for 3D VAEs. Experiments demonstrate state-of-the-art performance in both reconstruction accuracy and generative quality, yielding superior surface smoothness and geometric flexibility.

Method: Proposes multimodal frequency domain attention (MFDA) to mine complementary information, frequency domain feed-forward layer (FDFFL) for feature enhancement, cross-modal global modeling module (CGMM) for pixel-wise inter-modal interaction, and local feature enhancement module (LFEM) for multimodal local feature representation.

Result: Extensive experiments show FreDFT achieves excellent performance on multiple public datasets compared with state-of-the-art methods.

Conclusion: The proposed FreDFT effectively addresses information imbalance in visible-infrared object detection through frequency domain fusion and cross-modal interaction, demonstrating superior detection performance.

Abstract: Visible-infrared object detection has gained sufficient attention due to its detection performance in low light, fog, and rain conditions. However, visible and infrared modalities captured by different sensors exist the information imbalance problem in complex scenarios, which can cause inadequate cross-modal fusion, resulting in degraded detection performance. \textcolor{red}{Furthermore, most existing methods use transformers in the spatial domain to capture complementary features, ignoring the advantages of developing frequency domain transformers to mine complementary information.} To solve these weaknesses, we propose a frequency domain fusion transformer, called FreDFT, for visible-infrared object detection. The proposed approach employs a novel multimodal frequency domain attention (MFDA) to mine complementary information between modalities and a frequency domain feed-forward layer (FDFFL) via a mixed-scale frequency feature fusion strategy is designed to better enhance multimodal features. To eliminate the imbalance of multimodal information, a cross-modal global modeling module (CGMM) is constructed to perform pixel-wise inter-modal feature interaction in a spatial and channel manner. Moreover, a local feature enhancement module (LFEM) is developed to strengthen multimodal local feature representation and promote multimodal feature fusion by using various convolution layers and applying a channel shuffle. Extensive experimental results have verified that our proposed FreDFT achieves excellent performance on multiple public datasets compared with other state-of-the-art methods. The code of our FreDFT is linked at https://github.com/WenCongWu/FreDFT.

Method: Proposes Mutual Scoring framework (MuSc-V2) with: Iterative Point Grouping (IPG) for better 3D representation, Similarity Neighborhood Aggregation with Multi-Degrees (SNAMD) for multi-scale patch features, Mutual Scoring Mechanism (MSM) for intra-modal scoring, Cross-modal Anomaly Enhancement (CAE) for fusing 2D/3D scores, and Re-scoring with Constrained Neighborhood (RsCon) to suppress false positives.

Result: Achieves significant performance improvements: +23.7% AP gain on MVTec 3D-AD dataset and +19.3% boost on Eyecandies dataset, surpassing previous zero-shot benchmarks and outperforming most few-shot methods.

Conclusion: MuSc-V2 provides a flexible framework for zero-shot anomaly classification and segmentation that works effectively on both full datasets and smaller subsets, demonstrating robust performance across diverse product lines by explicitly leveraging the similarity patterns of normal patches versus isolated anomalies.

Abstract: Zero-shot anomaly classification (AC) and segmentation (AS) methods aim to identify and outline defects without using any labeled samples. In this paper, we reveal a key property that is overlooked by existing methods: normal image patches across industrial products typically find many other similar patches, not only in 2D appearance but also in 3D shapes, while anomalies remain diverse and isolated. To explicitly leverage this discriminative property, we propose a Mutual Scoring framework (MuSc-V2) for zero-shot AC/AS, which flexibly supports single 2D/3D or multimodality. Specifically, our method begins by improving 3D representation through Iterative Point Grouping (IPG), which reduces false positives from discontinuous surfaces. Then we use Similarity Neighborhood Aggregation with Multi-Degrees (SNAMD) to fuse 2D/3D neighborhood cues into more discriminative multi-scale patch features for mutual scoring. The core comprises a Mutual Scoring Mechanism (MSM) that lets samples within each modality to assign score to each other, and Cross-modal Anomaly Enhancement (CAE) that fuses 2D and 3D scores to recover modality-specific missing anomalies. Finally, Re-scoring with Constrained Neighborhood (RsCon) suppresses false classification based on similarity to more representative samples. Our framework flexibly works on both the full dataset and smaller subsets with consistently robust performance, ensuring seamless adaptability across diverse product lines. In aid of the novel framework, MuSc-V2 achieves significant performance improvements: a $\textbf{+23.7%}$ AP gain on the MVTec 3D-AD dataset and a $\textbf{+19.3%}$ boost on the Eyecandies dataset, surpassing previous zero-shot benchmarks and even outperforming most few-shot methods. The code will be available at The code will be available at \href{https://github.com/HUST-SLOW/MuSc-V2}{https://github.com/HUST-SLOW/MuSc-V2}.

Method: Created a large dataset (128k samples, 640k images) covering four aesthetic aspects, and developed HCM-GRPO method combining Hard Cases Mining with Dynamic Proportional Accuracy reward in GRPO framework.

Result: HCM-GRPO enables smaller models to surpass both large-scale open-source and leading closed-source models in image aesthetic reasoning, overcoming the random-guessing performance of current SOTA MLLMs.

Conclusion: The proposed data collection approach and HCM-GRPO method effectively address the image screening challenge, demonstrating that proper training methodology can enable smaller models to outperform much larger ones in aesthetic reasoning tasks.

Abstract: The performance of image generation has been significantly improved in recent years. However, the study of image screening is rare and its performance with Multimodal Large Language Models (MLLMs) is unsatisfactory due to the lack of data and the weak image aesthetic reasoning ability in MLLMs. In this work, we propose a complete solution to address these problems in terms of data and methodology. For data, we collect a comprehensive image screening dataset with over 128k samples, about 640k images. Each sample consists of an original image, four generated images. The dataset evaluates the image aesthetic reasoning ability under four aspects: appearance deformation, physical shadow, placement layout, and extension rationality. Regarding data annotation, we investigate multiple approaches, including purely manual, fully automated, and answer-driven annotations, to acquire high-quality chains of thought (CoT) data in the most cost-effective manner. Methodologically, we introduce a Hard Cases Mining (HCM) strategy with a Dynamic Proportional Accuracy (DPA) reward into the Group Relative Policy Optimization (GRPO) framework, called HCM-GRPO. This enhanced method demonstrates superior image aesthetic reasoning capabilities compared to the original GRPO. Our experimental results reveal that even state-of-the-art closed-source MLLMs, such as GPT4o and Qwen-VL-Max, exhibit performance akin to random guessing in image aesthetic reasoning. In contrast, by leveraging the HCM-GRPO, we are able to surpass the scores of both large-scale open-source and leading closed-source models with a much smaller model.

Method: Proposes RL-CoMM with two stages: 1) Uses a Large Audio Language Model as reference to generate audio-only reasoning and a Step-wise Reasoning Reward function for self-improvement, 2) Implements Answer-centered Confidence Optimization to reduce uncertainty from heterogeneous reasoning differences.

Result: RL-CoMM improves accuracy by 10-30% over baseline models on audio-visual question answering and audio-visual hallucination tasks with limited training data.

Conclusion: The proposed RL-CoMM framework effectively addresses visually dominated reasoning in MLLMs and significantly enhances their ability to handle audio-visual confusion scenarios.

Abstract: Can Multimodal Large Language Models (MLLMs) discern confused objects that are visually present but audio-absent? To study this, we introduce a new benchmark, AV-ConfuseBench, which simulates an ``Audio-Visual Confusion’’ scene by modifying the corresponding sound of an object in the video, e.g., mute the sounding object and ask MLLMs Is there a/an muted-object sound’’. Experimental results reveal that MLLMs, such as Qwen2.5-Omni and Gemini 2.5, struggle to discriminate non-existent audio due to visually dominated reasoning. Motivated by this observation, we introduce RL-CoMM, a Reinforcement Learning-based Collaborative Multi-MLLM that is built upon the Qwen2.5-Omni foundation. RL-CoMM includes two stages: 1) To alleviate visually dominated ambiguities, we introduce an external model, a Large Audio Language Model (LALM), as the reference model to generate audio-only reasoning. Then, we design a Step-wise Reasoning Reward function that enables MLLMs to self-improve audio-visual reasoning with the audio-only reference. 2) To ensure an accurate answer prediction, we introduce Answer-centered Confidence Optimization to reduce the uncertainty of potential heterogeneous reasoning differences. Extensive experiments on audio-visual question answering and audio-visual hallucination show that RL-CoMM improves the accuracy by 10~30% over the baseline model with limited training data. Follow: https://github.com/rikeilong/AVConfusion.

Method: GaussMedAct framework with multivariate Gaussian representation that projects joint motions to scaled space and decomposes actions into adaptive 3D Gaussian tokens. Uses hybrid spatial encoding with Cartesian and Vector dual-stream strategy.

Result: Achieves 92.1% Top-1 accuracy on CPREval-6k benchmark, outperforming ST-GCN baseline by +5.9% with only 10% FLOPs. Cross-dataset experiments confirm robustness.

Conclusion: The proposed method effectively addresses medical action evaluation challenges through adaptive spatiotemporal representation learning and achieves superior performance with computational efficiency.

Abstract: Fine-grained action evaluation in medical vision faces unique challenges due to the unavailability of comprehensive datasets, stringent precision requirements, and insufficient spatiotemporal dynamic modeling of very rapid actions. To support development and evaluation, we introduce CPREval-6k, a multi-view, multi-label medical action benchmark containing 6,372 expert-annotated videos with 22 clinical labels. Using this dataset, we present GaussMedAct, a multivariate Gaussian encoding framework, to advance medical motion analysis through adaptive spatiotemporal representation learning. Multivariate Gaussian Representation projects the joint motions to a temporally scaled multi-dimensional space, and decomposes actions into adaptive 3D Gaussians that serve as tokens. These tokens preserve motion semantics through anisotropic covariance modeling while maintaining robustness to spatiotemporal noise. Hybrid Spatial Encoding, employing a Cartesian and Vector dual-stream strategy, effectively utilizes skeletal information in the form of joint and bone features. The proposed method achieves 92.1% Top-1 accuracy with real-time inference on the benchmark, outperforming the ST-GCN baseline by +5.9% accuracy with only 10% FLOPs. Cross-dataset experiments confirm the superiority of our method in robustness.

Method: CABIN uses a closed-loop learning process: 1) perceptual awareness through epistemic uncertainty estimation, 2) uncertainty-guided dual sampling strategy (exploring uncertain samples, anchoring confident ones), and 3) fine-grained dynamic assignment strategy categorizing pseudo-labels into reliable, ambiguous, and noisy subsets with tailored losses.

Result: Experimental results show that various state-of-the-art methods benefit from CABIN integration, achieving improved labeling efficiency and performance in hyperspectral image classification.

Conclusion: CABIN effectively addresses overconfidence issues in hyperspectral image classification by incorporating uncertainty awareness and dynamic correction mechanisms, leading to better generalization and reduced confirmation bias.

Abstract: Confidence alone is often misleading in hyperspectral image classification, as models tend to mistake high predictive scores for correctness while lacking awareness of uncertainty. This leads to confirmation bias, especially under sparse annotations or class imbalance, where models overfit confident errors and fail to generalize. We propose CABIN (Cognitive-Aware Behavior-Informed learNing), a semi-supervised framework that addresses this limitation through a closed-loop learning process of perception, action, and correction. CABIN first develops perceptual awareness by estimating epistemic uncertainty, identifying ambiguous regions where errors are likely to occur. It then acts by adopting an Uncertainty-Guided Dual Sampling Strategy, selecting uncertain samples for exploration while anchoring confident ones as stable pseudo-labels to reduce bias. To correct noisy supervision, CABIN introduces a Fine-Grained Dynamic Assignment Strategy that categorizes pseudo-labeled data into reliable, ambiguous, and noisy subsets, applying tailored losses to enhance generalization. Experimental results show that a wide range of state-of-the-art methods benefit from the integration of CABIN, with improved labeling efficiency and performance.

Method: Uses a pre-trained vision-language model to encode source images into vision-language semantic features (VLF), which are transmitted. At receiver, decoder-based language model and diffusion-based image generator both use VLF to generate text and images.

Result: Outperforms text-only and image-only baselines, achieving higher semantic accuracy for both modalities under low SNR with significantly reduced bandwidth.

Conclusion: VLF-MSC provides an efficient unified framework for multimodal semantic communication that eliminates modality-specific streams while maintaining semantic fidelity and robustness to channel noise.

Abstract: We propose Vision-Language Feature-based Multimodal Semantic Communication (VLF-MSC), a unified system that transmits a single compact vision-language representation to support both image and text generation at the receiver. Unlike existing semantic communication techniques that process each modality separately, VLF-MSC employs a pre-trained vision-language model (VLM) to encode the source image into a vision-language semantic feature (VLF), which is transmitted over the wireless channel. At the receiver, a decoder-based language model and a diffusion-based image generator are both conditioned on the VLF to produce a descriptive text and a semantically aligned image. This unified representation eliminates the need for modality-specific streams or retransmissions, improving spectral efficiency and adaptability. By leveraging foundation models, the system achieves robustness to channel noise while preserving semantic fidelity. Experiments demonstrate that VLF-MSC outperforms text-only and image-only baselines, achieving higher semantic accuracy for both modalities under low SNR with significantly reduced bandwidth.

Method: Proposes GlobalDiff with multi-level constraints: joint structure constraint using virtual anchor points, skeleton structure constraint for angular consistency, and temporal structure constraint with multi-scale variational encoder.

Result: Improves performance by 46.0% compared to state-of-the-art methods on standard co-speech benchmarks, generating smooth and accurate motions across multiple speaker identities.

Conclusion: Operating directly in global rotation space with structural constraints effectively addresses hierarchical error accumulation and produces more reliable co-speech motion generation.

Abstract: Reliable co-speech motion generation requires precise motion representation and consistent structural priors across all joints. Existing generative methods typically operate on local joint rotations, which are defined hierarchically based on the skeleton structure. This leads to cumulative errors during generation, manifesting as unstable and implausible motions at end-effectors. In this work, we propose GlobalDiff, a diffusion-based framework that operates directly in the space of global joint rotations for the first time, fundamentally decoupling each joint’s prediction from upstream dependencies and alleviating hierarchical error accumulation. To compensate for the absence of structural priors in global rotation space, we introduce a multi-level constraint scheme. Specifically, a joint structure constraint introduces virtual anchor points around each joint to better capture fine-grained orientation. A skeleton structure constraint enforces angular consistency across bones to maintain structural integrity. A temporal structure constraint utilizes a multi-scale variational encoder to align the generated motion with ground-truth temporal patterns. These constraints jointly regularize the global diffusion process and reinforce structural awareness. Extensive evaluations on standard co-speech benchmarks show that GlobalDiff generates smooth and accurate motions, improving the performance by 46.0 % compared to the current SOTA under multiple speaker identities.

Method: GridPrune splits pruning into two steps: 1) text-conditional guidance to dynamically allocate token budgets across spatial zones, and 2) local selection within each budgeted zone, replacing the global Top-K mechanism.

Result: On LLaVA-NeXT-7B, GridPrune retains 96.98% of full performance while using only 11.1% of tokens, outperforming the best baseline by 2.34% at the same pruning rate across various MLLM architectures.

Conclusion: The two-stage “guide-globally, select-locally” approach effectively addresses spatial allocation issues in visual token pruning, demonstrating superior efficiency and performance compared to existing methods.

Abstract: Multimodal large language models (MLLMs) have shown remarkable capabilities in a wide range of vision-language tasks. However, the large number of visual tokens introduces significant computational overhead. To address this issue, visual token pruning has emerged as a key technique for enhancing the efficiency of MLLMs. In cognitive science, humans tend to first determine which regions of a scene to attend to (“where to look”) before deciding which specific elements within those regions to process in detail (“what to select”). This two-stage strategy enables the visual system to efficiently allocate attention at a coarse spatial level before performing fine-grained selection. However, existing pruning methods primarily focus on directly optimizing “what to select”, typically using attention scores or similarity metrics. They rarely consider “where to look”, which has been shown to lead to inefficient spatial allocation, positional bias, and the retention of irrelevant or redundant tokens. In this paper, we propose GridPrune, a method that replaces the global Top-K mechanism with a “guide-globally, select-locally” zonal selection system. GridPrune splits the pruning process into two steps: first, it uses text-conditional guidance to dynamically allocate a token budget across spatial zones; and then, it performs local selection within each budgeted zone. Experimental results demonstrate that GridPrune achieves superior performance across various MLLM architectures. On LLaVA-NeXT-7B, GridPrune retains 96.98% of the full performance while using 11.1% of the tokens, outperforming the best-performing baseline by 2.34% at the same pruning rate.

Method: SUGAR pipeline: 1) Use large-scale video models to generate visual-motion knowledge, 2) Supervise skeleton learning with prior knowledge to create discrete representations, 3) Use pre-trained LLMs to understand representations and generate outputs, 4) Temporal Query Projection module for long sequence modeling.

Result: Achieves strong performance on skeleton-based action classification benchmarks and shows better versatility than linear-based methods in zero-shot scenarios.

Conclusion: SUGAR effectively bridges LLMs with skeleton data for action recognition, demonstrating robust performance and generalization capabilities without requiring LLM fine-tuning.

Abstract: Large Language Models (LLMs) hold rich implicit knowledge and powerful transferability. In this paper, we explore the combination of LLMs with the human skeleton to perform action classification and description. However, when treating LLM as a recognizer, two questions arise: 1) How can LLMs understand skeleton? 2) How can LLMs distinguish among actions? To address these problems, we introduce a novel paradigm named learning Skeleton representation with visUal-motion knowledGe for Action Recognition (SUGAR). In our pipeline, we first utilize off-the-shelf large-scale video models as a knowledge base to generate visual, motion information related to actions. Then, we propose to supervise skeleton learning through this prior knowledge to yield discrete representations. Finally, we use the LLM with untouched pre-training weights to understand these representations and generate the desired action targets and descriptions. Notably, we present a Temporal Query Projection (TQP) module to continuously model the skeleton signals with long sequences. Experiments on several skeleton-based action classification benchmarks demonstrate the efficacy of our SUGAR. Moreover, experiments on zero-shot scenarios show that SUGAR is more versatile than linear-based methods.

Method: Proposes MTAttack with two optimization constraints: Proxy Space Partitioning and Trigger Prototype Anchoring, which jointly optimize multiple triggers in latent space to ensure each trigger independently maps to unique proxy classes while maintaining separability.

Result: Achieves high success rate for multi-target attacks, substantially outperforms existing methods, shows strong generalizability across datasets, and demonstrates robustness against backdoor defense strategies.

Conclusion: LVLMs are vulnerable to sophisticated multi-target backdoor attacks, highlighting urgent need for defense mechanisms against such threats in real-world applications.

Abstract: Recent advances in Large Visual Language Models (LVLMs) have demonstrated impressive performance across various vision-language tasks by leveraging large-scale image-text pretraining and instruction tuning. However, the security vulnerabilities of LVLMs have become increasingly concerning, particularly their susceptibility to backdoor attacks. Existing backdoor attacks focus on single-target attacks, i.e., targeting a single malicious output associated with a specific trigger. In this work, we uncover multi-target backdoor attacks, where multiple independent triggers corresponding to different attack targets are added in a single pass of training, posing a greater threat to LVLMs in real-world applications. Executing such attacks in LVLMs is challenging since there can be many incorrect trigger-target mappings due to severe feature interference among different triggers. To address this challenge, we propose MTAttack, the first multi-target backdoor attack framework for enforcing accurate multiple trigger-target mappings in LVLMs. The core of MTAttack is a novel optimization method with two constraints, namely Proxy Space Partitioning constraint and Trigger Prototype Anchoring constraint. It jointly optimizes multiple triggers in the latent space, with each trigger independently mapping clean images to a unique proxy class while at the same time guaranteeing their separability. Experiments on popular benchmarks demonstrate a high success rate of MTAttack for multi-target attacks, substantially outperforming existing attack methods. Furthermore, our attack exhibits strong generalizability across datasets and robustness against backdoor defense strategies. These findings highlight the vulnerability of LVLMs to multi-target backdoor attacks and underscore the urgent need for mitigating such threats. Code is available at https://github.com/mala-lab/MTAttack.

Method: Integrates Attend-and-Excite Mixture-of-Experts (AttEx-MoE) for dynamic input routing via lightweight self-attention, and Robust AdaptBN Teacher for dense pseudo-supervision using PEFT-based approach.

Result: Achieves superior adaptation performance across dynamic target domains while maintaining computational efficiency, demonstrating improved generalization under domain shift.

Conclusion: RobIA provides an effective solution for continual test-time adaptation in stereo depth estimation with input-specific flexibility and broad supervision coverage.

Abstract: Stereo Depth Estimation in real-world environments poses significant challenges due to dynamic domain shifts, sparse or unreliable supervision, and the high cost of acquiring dense ground-truth labels. While recent Test-Time Adaptation (TTA) methods offer promising solutions, most rely on static target domain assumptions and input-invariant adaptation strategies, limiting their effectiveness under continual shifts. In this paper, we propose RobIA, a novel Robust, Instance-Aware framework for Continual Test-Time Adaptation (CTTA) in stereo depth estimation. RobIA integrates two key components: (1) Attend-and-Excite Mixture-of-Experts (AttEx-MoE), a parameter-efficient module that dynamically routes input to frozen experts via lightweight self-attention mechanism tailored to epipolar geometry, and (2) Robust AdaptBN Teacher, a PEFT-based teacher model that provides dense pseudo-supervision by complementing sparse handcrafted labels. This strategy enables input-specific flexibility, broad supervision coverage, improving generalization under domain shift. Extensive experiments demonstrate that RobIA achieves superior adaptation performance across dynamic target domains while maintaining computational efficiency.

Method: CACMI framework with two components: Cross-modal Frame Aggregation for temporally coherent event-aligned features, and Context-aware Feature Enhancement that integrates visual dynamics with pseudo-event semantics using query-guided attention.

Result: Extensive experiments on ActivityNet Captions and YouCook2 datasets demonstrate state-of-the-art performance on dense video captioning task.

Conclusion: The proposed explicit temporal-semantic modeling framework effectively addresses limitations of implicit modeling approaches and achieves superior dense video captioning performance.

Abstract: Dense video captioning jointly localizes and captions salient events in untrimmed videos. Recent methods primarily focus on leveraging additional prior knowledge and advanced multi-task architectures to achieve competitive performance. However, these pipelines rely on implicit modeling that uses frame-level or fragmented video features, failing to capture the temporal coherence across event sequences and comprehensive semantics within visual contexts. To address this, we propose an explicit temporal-semantic modeling framework called Context-Aware Cross-Modal Interaction (CACMI), which leverages both latent temporal characteristics within videos and linguistic semantics from text corpus. Specifically, our model consists of two core components: Cross-modal Frame Aggregation aggregates relevant frames to extract temporally coherent, event-aligned textual features through cross-modal retrieval; and Context-aware Feature Enhancement utilizes query-guided attention to integrate visual dynamics with pseudo-event semantics. Extensive experiments on the ActivityNet Captions and YouCook2 datasets demonstrate that CACMI achieves the state-of-the-art performance on dense video captioning task.

Method: Analysis across three core logical primitives (negation, counting, spatial relations) by examining training data limitations, architectural suitability of continuous attention for discrete logic, and evaluation metric biases.

Result: Models show dramatic performance collapse when primitives are combined, with severe interference effects. Current solutions and scaling cannot bridge this gap due to fundamental architectural and data limitations.

Conclusion: Achieving genuine compositionality requires fundamental advances in representation and reasoning rather than incremental adjustments to existing architectures.

Abstract: The architectural blueprint of today’s leading text-to-image models contains a fundamental flaw: an inability to handle logical composition. This survey investigates this breakdown across three core primitives-negation, counting, and spatial relations. Our analysis reveals a dramatic performance collapse: models that are accurate on single primitives fail precipitously when these are combined, exposing severe interference. We trace this failure to three key factors. First, training data show a near-total absence of explicit negations. Second, continuous attention architectures are fundamentally unsuitable for discrete logic. Third, evaluation metrics reward visual plausibility over constraint satisfaction. By analyzing recent benchmarks and methods, we show that current solutions and simple scaling cannot bridge this gap. Achieving genuine compositionality, we conclude, will require fundamental advances in representation and reasoning rather than incremental adjustments to existing architectures.

Method: Proposes split-layer MLP construction that divides each layer into multiple parallel branches and integrates their outputs using Hadamard product, effectively creating high-degree polynomial spaces.

Result: Extensive experiments show split-layer significantly improves INR performance, surpassing existing methods in 2D image fitting, 2D CT reconstruction, 3D shape representation, and 5D novel view synthesis tasks.

Conclusion: The split-layer approach effectively enhances INR’s representational capacity by expanding feature space dimensionality without prohibitive computational overhead, demonstrating superior performance across multiple applications.

Abstract: Implicit neural representation (INR) models signals as continuous functions using neural networks, offering efficient and differentiable optimization for inverse problems across diverse disciplines. However, the representational capacity of INR defined by the range of functions the neural network can characterize, is inherently limited by the low-dimensional feature space in conventional multilayer perceptron (MLP) architectures. While widening the MLP can linearly increase feature space dimensionality, it also leads to a quadratic growth in computational and memory costs. To address this limitation, we propose the split-layer, a novel reformulation of MLP construction. The split-layer divides each layer into multiple parallel branches and integrates their outputs via Hadamard product, effectively constructing a high-degree polynomial space. This approach significantly enhances INR’s representational capacity by expanding the feature space dimensionality without incurring prohibitive computational overhead. Extensive experiments demonstrate that the split-layer substantially improves INR performance, surpassing existing methods across multiple tasks, including 2D image fitting, 2D CT reconstruction, 3D shape representation, and 5D novel view synthesis.

Method: Proposes a dual-mechanism collaborative optimization framework that integrates structural fairness decoupling (decoupling channels sensitive to demographic groups at model architecture level) and global distribution alignment (reducing distance between overall sample distribution and distributions of each demographic group at feature level).

Result: Experimental results show the framework improves both inter-group and intra-group fairness while maintaining overall detection accuracy across domains, outperforming other methods.

Conclusion: The proposed framework successfully addresses the fairness-accuracy trade-off in deepfake detection, enabling more equitable and trustworthy deployment in digital identity security applications.

Abstract: Fairness is a core element in the trustworthy deployment of deepfake detection models, especially in the field of digital identity security. Biases in detection models toward different demographic groups, such as gender and race, may lead to systemic misjudgments, exacerbating the digital divide and social inequities. However, current fairness-enhanced detectors often improve fairness at the cost of detection accuracy. To address this challenge, we propose a dual-mechanism collaborative optimization framework. Our proposed method innovatively integrates structural fairness decoupling and global distribution alignment: decoupling channels sensitive to demographic groups at the model architectural level, and subsequently reducing the distance between the overall sample distribution and the distributions corresponding to each demographic group at the feature level. Experimental results demonstrate that, compared with other methods, our framework improves both inter-group and intra-group fairness while maintaining overall detection accuracy across domains.

Method: Two parallel modules: (1) Text-Guided Token Enhancement (TGTE) uses diffusion-generated images as intermediate semantic representations; (2) Generative Intermediate Fusion (GIF) combines cross-attention between generated images, original images, and text features with triplet alignment loss.

Result: Extensive experiments on CUHK-PEDES, RSTPReid, and ICFG-PEDES datasets demonstrate the effectiveness of the proposed method.

Conclusion: GEA successfully addresses cross-modal alignment challenges in text-to-image person retrieval by leveraging generative approaches to bridge the modality gap.

Abstract: Text-to-Image Person Retrieval (TIPR) aims to retrieve person images based on natural language descriptions. Although many TIPR methods have achieved promising results, sometimes textual queries cannot accurately and comprehensively reflect the content of the image, leading to poor cross-modal alignment and overfitting to limited datasets. Moreover, the inherent modality gap between text and image further amplifies these issues, making accurate cross-modal retrieval even more challenging. To address these limitations, we propose the Generation-Enhanced Alignment (GEA) from a generative perspective. GEA contains two parallel modules: (1) Text-Guided Token Enhancement (TGTE), which introduces diffusion-generated images as intermediate semantic representations to bridge the gap between text and visual patterns. These generated images enrich the semantic representation of text and facilitate cross-modal alignment. (2) Generative Intermediate Fusion (GIF), which combines cross-attention between generated images, original images, and text features to generate a unified representation optimized by triplet alignment loss. We conduct extensive experiments on three public TIPR datasets, CUHK-PEDES, RSTPReid, and ICFG-PEDES, to evaluate the performance of GEA. The results justify the effectiveness of our method. More implementation details and extended results are available at https://github.com/sugelamyd123/Sup-for-GEA.

Method: Deep unfolding network mapping iterative optimization algorithm into learnable structure, using improved second-order semi-smooth Newton algorithm for physical interpretability. Includes explainable convolution module inspired by human brain’s flexible information processing and image characteristics.

Result: InterIR demonstrates excellent performance in multi-degradation restoration while remaining highly competitive on single-degradation tasks.

Conclusion: The proposed interpretability-driven approach successfully addresses multi-degradation image restoration with clear physical interpretability and maintains competitive performance across both multi- and single-degradation scenarios.

Abstract: Although image restoration has advanced significantly, most existing methods target only a single type of degradation. In real-world scenarios, images often contain multiple degradations simultaneously, such as rain, noise, and haze, requiring models capable of handling diverse degradation types. Moreover, methods that improve performance through module stacking often suffer from limited interpretability. In this paper, we propose a novel interpretability-driven approach for multi-degradation image restoration, built upon a deep unfolding network that maps the iterative process of a mathematical optimization algorithm into a learnable network structure. Specifically, we employ an improved second-order semi-smooth Newton algorithm to ensure that each module maintains clear physical interpretability. To further enhance interpretability and adaptability, we design an explainable convolution module inspired by the human brain’s flexible information processing and the intrinsic characteristics of images, allowing the network to flexibly leverage learned knowledge and autonomously adjust parameters for different input. The resulting tightly integrated architecture, named InterIR, demonstrates excellent performance in multi-degradation restoration while remaining highly competitive on single-degradation tasks.

Method: Multi-head residual convolutional network integrating residual encoding, dual-attention mechanisms, and multi-head decoders to enhance contextual reasoning and anatomical precision.

Result: Achieved mean radial error (MRE) of 1.23 mm and success detection rate (SDR) @ 2.0 mm of 85.5%, outperforming all evaluated models including AFPF-Net (MRE = 1.25 mm, SDR @ 2.0 mm = 84.1%) while using less than 25% of its parameters.

Conclusion: CephRes-MHNet attains state-of-the-art accuracy through architectural efficiency, providing a practical solution for real-world orthodontic analysis.

Abstract: Accurate localization of cephalometric landmarks from 2D lateral skull X-rays is vital for orthodontic diagnosis and treatment. Manual annotation is time-consuming and error-prone, whereas automated approaches often struggle with low contrast and anatomical complexity. This paper introduces CephRes-MHNet, a multi-head residual convolutional network for robust and efficient cephalometric landmark detection. The architecture integrates residual encoding, dual-attention mechanisms, and multi-head decoders to enhance contextual reasoning and anatomical precision. Trained on the Aariz Cephalometric dataset of 1,000 radiographs, CephRes-MHNet achieved a mean radial error (MRE) of 1.23 mm and a success detection rate (SDR) @ 2.0 mm of 85.5%, outperforming all evaluated models. In particular, it exceeded the strongest baseline, the attention-driven AFPF-Net (MRE = 1.25 mm, SDR @ 2.0 mm = 84.1%), while using less than 25% of its parameters. These results demonstrate that CephRes-MHNet attains state-of-the-art accuracy through architectural efficiency, providing a practical solution for real-world orthodontic analysis.

Method: Curated and labeled 225 multispectral images of Maldivian islands, fine-tuned both 300M and 600M parameter versions of Prithvi model on training subsets ranging from 5 to 181 images.

Result: Models achieved high performance even with only 5 training images (F1 score of 0.94, IoU of 0.79), demonstrating strong transfer learning capability.

Conclusion: Prithvi geospatial foundation models show significant potential for supporting coastal monitoring in data-poor regions due to their strong transfer learning performance with minimal training data.

Abstract: We present an initial evaluation of NASA and IBM’s Prithvi-EO-2.0 geospatial foundation model on shoreline delineation of small sandy islands using satellite images. We curated and labeled a dataset of 225 multispectral images of two Maldivian islands, which we publicly release, and fine-tuned both the 300M and 600M parameter versions of Prithvi on training subsets ranging from 5 to 181 images. Our experiments show that even with as few as 5 training images, the models achieve high performance (F1 of 0.94, IoU of 0.79). Our results demonstrate the strong transfer learning capability of Prithvi, underscoring the potential of such models to support coastal monitoring in data-poor regions.

Method: Evaluated SmolVLM2 variants across AVCaps and Charades datasets using two novel frameworks: Multi-Context BLV Framework (spatial orientation, social interaction, action events, ambience) and Navigational Assistance Framework. Tested four prompt strategies and deployed on smartphones with FP32/INT8 precision.

Result: Not specified in abstract - evaluation results of model variants and frameworks would be presented in full paper.

Conclusion: The study provides systematic evaluation of model size effects on accessibility-focused description quality and real-world mobile deployment constraints for BLV users.

Abstract: Large Vision-Language Models (VLMs) excel at understanding and generating video descriptions but their high memory, computation, and deployment demands hinder practical use particularly for blind and low-vision (BLV) users who depend on detailed, context-aware descriptions. To study the effect of model size on accessibility-focused description quality, we evaluate SmolVLM2 variants with 500M and 2.2B parameters across two diverse datasets: AVCaps (outdoor), and Charades (indoor). In this work, we introduce two novel evaluation frameworks specifically designed for BLV accessibility assessment: the Multi-Context BLV Framework evaluating spatial orientation, social interaction, action events, and ambience contexts; and the Navigational Assistance Framework focusing on mobility-critical information. Additionally, we conduct a systematic evaluation of four different prompt design strategies and deploy both models on a smartphone, evaluating FP32 and INT8 precision variants to assess real-world performance constraints on resource-limited mobile devices.

Method: VISTA uses: (i) cross-attention fusion module for long-horizon intent integration, (ii) social-token attention for flexible interaction modeling, and (iii) pairwise attention maps for interpretable social influence patterns.

Result: Achieves state-of-the-art accuracy on MADRAS and SDD benchmarks, reducing collision rate from 2.14% to 0.03% on MADRAS and attaining zero collisions on SDD while improving ADE, FDE, and minFDE metrics.

Conclusion: VISTA generates socially compliant, goal-aware, and interpretable trajectories, making it promising for safety-critical autonomous systems.

Abstract: Multi-agent trajectory prediction is crucial for autonomous systems operating in dense, interactive environments. Existing methods often fail to jointly capture agents’ long-term goals and their fine-grained social interactions, which leads to unrealistic multi-agent futures. We propose VISTA, a recursive goal-conditioned transformer for multi-agent trajectory forecasting. VISTA combines (i) a cross-attention fusion module that integrates long-horizon intent with past motion, (ii) a social-token attention mechanism for flexible interaction modeling across agents, and (iii) pairwise attention maps that make social influence patterns interpretable at inference time. Our model turns single-agent goal-conditioned prediction into a coherent multi-agent forecasting framework. Beyond standard displacement metrics, we evaluate trajectory collision rates as a measure of joint realism. On the high-density MADRAS benchmark and on SDD, VISTA achieves state-of-the-art accuracy and substantially fewer collisions. On MADRAS, it reduces the average collision rate of strong baselines from 2.14 to 0.03 percent, and on SDD it attains zero collisions while improving ADE, FDE, and minFDE. These results show that VISTA generates socially compliant, goal-aware, and interpretable trajectories, making it promising for safety-critical autonomous systems.

Method: Proposes LiNeXt with Noise-to-Coarse Module for single-pass denoising, Refine Module for precise refinement, and Distance-aware Selected Repeat strategy to handle LiDAR’s distance-dependent spatial distribution.

Result: On SemanticKITTI, achieves 199.8x inference speedup, 50.7% reduction in Chamfer Distance, and uses only 6.1% of parameters compared to LiDiff.

Conclusion: LiNeXt demonstrates superior efficiency and effectiveness for real-time scene completion, making it suitable for autonomous vehicle perception systems.

Abstract: 3D LiDAR scene completion from point clouds is a fundamental component of perception systems in autonomous vehicles. Previous methods have predominantly employed diffusion models for high-fidelity reconstruction. However, their multi-step iterative sampling incurs significant computational overhead, limiting its real-time applicability. To address this, we propose LiNeXt-a lightweight, non-diffusion network optimized for rapid and accurate point cloud completion. Specifically, LiNeXt first applies the Noise-to-Coarse (N2C) Module to denoise the input noisy point cloud in a single pass, thereby obviating the multi-step iterative sampling of diffusion-based methods. The Refine Module then takes the coarse point cloud and its intermediate features from the N2C Module to perform more precise refinement, further enhancing structural completeness. Furthermore, we observe that LiDAR point clouds exhibit a distance-dependent spatial distribution, being densely sampled at proximal ranges and sparsely sampled at distal ranges. Accordingly, we propose the Distance-aware Selected Repeat strategy to generate a more uniformly distributed noisy point cloud. On the SemanticKITTI dataset, LiNeXt achieves a 199.8x speedup in inference, reduces Chamfer Distance by 50.7%, and uses only 6.1% of the parameters compared with LiDiff. These results demonstrate the superior efficiency and effectiveness of LiNeXt for real-time scene completion.

Method: Proposes HeatV2X with two-stage fine-tuning: Local Heterogeneous Fine-Tuning using Hetero-Aware Adapters for modality-specific alignment, and Global Collaborative Fine-Tuning using Multi-Cognitive Adapter for cross-agent interaction.

Result: Achieves superior perception performance on OPV2V-H and DAIR-V2X datasets with significantly reduced training overhead compared to state-of-the-art methods.

Conclusion: HeatV2X provides an effective and scalable solution for heterogeneous V2X collaborative perception, enabling substantial performance improvements with minimal training costs.

Abstract: Vehicle-to-Everything (V2X) collaborative perception extends sensing beyond single vehicle limits through transmission. However, as more agents participate, existing frameworks face two key challenges: (1) the participating agents are inherently multi-modal and heterogeneous, and (2) the collaborative framework must be scalable to accommodate new agents. The former requires effective cross-agent feature alignment to mitigate heterogeneity loss, while the latter renders full-parameter training impractical, highlighting the importance of scalable adaptation. To address these issues, we propose Heterogeneous Adaptation (HeatV2X), a scalable collaborative framework. We first train a high-performance agent based on heterogeneous graph attention as the foundation for collaborative learning. Then, we design Local Heterogeneous Fine-Tuning and Global Collaborative Fine-Tuning to achieve effective alignment and interaction among heterogeneous agents. The former efficiently extracts modality-specific differences using Hetero-Aware Adapters, while the latter employs the Multi-Cognitive Adapter to enhance cross-agent collaboration and fully exploit the fusion potential. These designs enable substantial performance improvement of the collaborative framework with minimal training cost. We evaluate our approach on the OPV2V-H and DAIR-V2X datasets. Experimental results demonstrate that our method achieves superior perception performance with significantly reduced training overhead, outperforming existing state-of-the-art approaches. Our implementation will be released soon.

Method: Proposes a single-stage training framework incorporating next-frame prediction for both uni-modal and cross-modal features, with a window-level attention mechanism to capture discrepancies between predicted and actual frames.

Result: The model demonstrates strong generalization and precise temporal localization when evaluated on multiple benchmark datasets.

Conclusion: The proposed approach effectively addresses limitations of existing methods by detecting local artifacts around every frame, enabling accurate classification of fully manipulated videos and effective localization of deepfake segments in partially spoofed samples.

Abstract: Recent multimodal deepfake detection methods designed for generalization conjecture that single-stage supervised training struggles to generalize across unseen manipulations and datasets. However, such approaches that target generalization require pretraining over real samples. Additionally, these methods primarily focus on detecting audio-visual inconsistencies and may overlook intra-modal artifacts causing them to fail against manipulations that preserve audio-visual alignment. To address these limitations, we propose a single-stage training framework that enhances generalization by incorporating next-frame prediction for both uni-modal and cross-modal features. Additionally, we introduce a window-level attention mechanism to capture discrepancies between predicted and actual frames, enabling the model to detect local artifacts around every frame, which is crucial for accurately classifying fully manipulated videos and effectively localizing deepfake segments in partially spoofed samples. Our model, evaluated on multiple benchmark datasets, demonstrates strong generalization and precise temporal localization.

Method: Proposes TubeRMC framework that: 1) generates text-conditioned candidate tubes using pre-trained visual grounding models, 2) refines them via tube-conditioned reconstruction with three strategies (temporal, spatial, spatio-temporal), 3) uses Tube-conditioned Reconstructors to reconstruct query clues from tubes, 4) applies mutual constraints between spatial and temporal proposals.

Result: Outperforms existing methods on VidSTG and HCSTVG benchmarks. Visualization shows effective mitigation of target identification errors and inconsistent tracking.

Conclusion: TubeRMC successfully addresses limitations of previous weakly-supervised STVG methods by incorporating text-conditioned tube generation and comprehensive reconstruction strategies with mutual constraints.

Abstract: Spatio-Temporal Video Grounding (STVG) aims to localize a spatio-temporal tube that corresponds to a given language query in an untrimmed video. This is a challenging task since it involves complex vision-language understanding and spatiotemporal reasoning. Recent works have explored weakly-supervised setting in STVG to eliminate reliance on fine-grained annotations like bounding boxes or temporal stamps. However, they typically follow a simple late-fusion manner, which generates tubes independent of the text description, often resulting in failed target identification and inconsistent target tracking. To address this limitation, we propose a Tube-conditioned Reconstruction with Mutual Constraints (\textbf{TubeRMC}) framework that generates text-conditioned candidate tubes with pre-trained visual grounding models and further refine them via tube-conditioned reconstruction with spatio-temporal constraints. Specifically, we design three reconstruction strategies from temporal, spatial, and spatio-temporal perspectives to comprehensively capture rich tube-text correspondences. Each strategy is equipped with a Tube-conditioned Reconstructor, utilizing spatio-temporal tubes as condition to reconstruct the key clues in the query. We further introduce mutual constraints between spatial and temporal proposals to enhance their quality for reconstruction. TubeRMC outperforms existing methods on two public benchmarks VidSTG and HCSTVG. Further visualization shows that TubeRMC effectively mitigates both target identification errors and inconsistent tracking.

Method: Segments action videos into stages, uses stage-aware feature enhancement and fusion to handle camera viewpoint changes, and employs a knowledge-based grade-aware decoder to incorporate deduction items as prior knowledge.

Result: Achieves state-of-the-art performance on the new aerial skiing AQA benchmark with fine-grained sub-score and deduction annotations.

Conclusion: JudgeMind significantly enhances AQA performance and reliability by mimicking professional referees’ judgment process through stage segmentation and deduction knowledge integration.

Abstract: Action Quality Assessment (AQA) aims to evaluate and score sports actions, which has attracted widespread interest in recent years. Existing AQA methods primarily predict scores based on features extracted from the entire video, resulting in limited interpretability and reliability. Meanwhile, existing AQA datasets also lack fine-grained annotations for action scores, especially for deduction items and sub-score annotations. In this paper, we construct the first AQA dataset containing fine-grained sub-score and deduction annotations for aerial skiing, which will be released as a new benchmark. For the technical challenges, we propose a novel AQA method, named JudgeMind, which significantly enhances performance and reliability by simulating the judgment and scoring mindset of professional referees. Our method segments the input action video into different stages and scores each stage to enhance accuracy. Then, we propose a stage-aware feature enhancement and fusion module to boost the perception of stage-specific key regions and enhance the robustness to visual changes caused by frequent camera viewpoints switching. In addition, we propose a knowledge-based grade-aware decoder to incorporate possible deduction items as prior knowledge to predict more accurate and reliable scores. Experimental results demonstrate that our method achieves state-of-the-art performance.

Method: Three-stage framework: 1) Instruction fine-tuning for basic emotional reasoning, 2) Reinforcement training guided by emotion and AU labels as rewards to align reasoning with predictions, 3) Iterative data synthesis pipeline to expand training data for scalable self-improvement.

Result: Achieved state-of-the-art performance across eight standard FEA benchmarks, with strong generalization and robust interpretability. Introduced FEA-20K dataset with 17,737 training and 1,688 test samples.

Conclusion: Facial-R1 effectively addresses hallucinated reasoning and misalignment issues in FEA with minimal supervision, demonstrating superior performance and interpretability through its three-stage alignment framework and iterative data synthesis approach.

Abstract: Facial Emotion Analysis (FEA) extends traditional facial emotion recognition by incorporating explainable, fine-grained reasoning. The task integrates three subtasks: emotion recognition, facial Action Unit (AU) recognition, and AU-based emotion reasoning to model affective states jointly. While recent approaches leverage Vision-Language Models (VLMs) and achieve promising results, they face two critical limitations: (1) hallucinated reasoning, where VLMs generate plausible but inaccurate explanations due to insufficient emotion-specific knowledge; and (2) misalignment between emotion reasoning and recognition, caused by fragmented connections between observed facial features and final labels. We propose Facial-R1, a three-stage alignment framework that effectively addresses both challenges with minimal supervision. First, we employ instruction fine-tuning to establish basic emotional reasoning capability. Second, we introduce reinforcement training guided by emotion and AU labels as reward signals, which explicitly aligns the generated reasoning process with the predicted emotion. Third, we design a data synthesis pipeline that iteratively leverages the prior stages to expand the training dataset, enabling scalable self-improvement of the model. Built upon this framework, we introduce FEA-20K, a benchmark dataset comprising 17,737 training and 1,688 test samples with fine-grained emotion analysis annotations. Extensive experiments across eight standard benchmarks demonstrate that Facial-R1 achieves state-of-the-art performance in FEA, with strong generalization and robust interpretability.

Method: Proposes H3Former with Semantic-Aware Aggregation Module (SAAM) that constructs weighted hypergraphs among tokens using multi-scale contextual cues, and Hyperbolic Hierarchical Contrastive Loss (HHCL) that enforces hierarchical constraints in non-Euclidean space.

Result: Comprehensive experiments on four standard FGVC benchmarks validate the superiority of the H3Former framework over existing approaches.

Conclusion: H3Former effectively addresses limitations of existing FGVC methods by leveraging high-order semantic relations and hierarchical contrastive learning, achieving superior performance on fine-grained visual classification tasks.

Abstract: Fine-Grained Visual Classification (FGVC) remains a challenging task due to subtle inter-class differences and large intra-class variations. Existing approaches typically rely on feature-selection mechanisms or region-proposal strategies to localize discriminative regions for semantic analysis. However, these methods often fail to capture discriminative cues comprehensively while introducing substantial category-agnostic redundancy. To address these limitations, we propose H3Former, a novel token-to-region framework that leverages high-order semantic relations to aggregate local fine-grained representations with structured region-level modeling. Specifically, we propose the Semantic-Aware Aggregation Module (SAAM), which exploits multi-scale contextual cues to dynamically construct a weighted hypergraph among tokens. By applying hypergraph convolution, SAAM captures high-order semantic dependencies and progressively aggregates token features into compact region-level representations. Furthermore, we introduce the Hyperbolic Hierarchical Contrastive Loss (HHCL), which enforces hierarchical semantic constraints in a non-Euclidean embedding space. The HHCL enhances inter-class separability and intra-class consistency while preserving the intrinsic hierarchical relationships among fine-grained categories. Comprehensive experiments conducted on four standard FGVC benchmarks validate the superiority of our H3Former framework.

Method: PROPA combines Monte Carlo Tree Search with GRPO to generate dense process-level rewards, interleaves GRPO updates with SFT to overcome cold-start problems, and trains a Process Reward Model to guide inference-time search.

Result: PROPA outperforms SFT- and RLVR-based baselines across seven benchmarks and four VLM backbones, achieving up to 17.0% gains on in-domain tasks and 21.0% gains on out-of-domain tasks compared to state-of-the-art methods.

Conclusion: PROPA establishes strong reasoning and generalization capabilities for visual reasoning tasks by optimizing reasoning at each intermediate step without human annotations, demonstrating superior performance over existing approaches.

Abstract: Despite significant progress, Vision-Language Models (VLMs) still struggle with complex visual reasoning, where multi-step dependencies cause early errors to cascade through the reasoning chain. Existing post-training paradigms are limited: Supervised Fine-Tuning (SFT) relies on costly step-level annotations, while Reinforcement Learning with Verifiable Rewards (RLVR) methods like GRPO provide only sparse, outcome-level feedback, hindering stable optimization. We introduce PROPA (Process-level Reasoning Optimization with interleaved Policy Alignment), a novel framework that integrates Monte Carlo Tree Search (MCTS) with GRPO to generate dense, process-level rewards and optimize reasoning at each intermediate step without human annotations. To overcome the cold-start problem, PROPA interleaves GRPO updates with SFT, enabling the model to learn from both successful and failed reasoning trajectories. A Process Reward Model (PRM) is further trained to guide inference-time search, aligning the test-time search with the training signal. Across seven benchmarks and four VLM backbones, PROPA consistently outperforms both SFT- and RLVR-based baselines. It achieves up to 17.0% gains on in-domain tasks and 21.0% gains on out-of-domain tasks compared to existing state-of-the-art, establishing a strong reasoning and generalization capability for visual reasoning tasks. The code isavailable at: https://github.com/YanbeiJiang/PROPA.

Method: Uses two key innovations: 1) CARD vector - a per-sample visual evidence vector extracted from residual updates of self-attention layers during standard forward passes, and 2) Bayesian-inspired adaptive gate for token-wise injection of corrective signals based on model’s deviation from visual context.

Result: Achieves performance comparable to state-of-the-art methods on hallucination benchmarks (POPE and CHAIR) while introducing negligible computational latency.

Conclusion: RUDDER provides a pragmatic and effective approach for improving LVLM reliability without compromising efficiency, making it suitable for latency-sensitive real-world deployments.

Abstract: Large Vision-Language Models (LVLMs) often suffer from object hallucination, generating text inconsistent with visual inputs, which can critically undermine their reliability. Existing inference-time interventions to mitigate this issue present a challenging trade-off: while methods that steer internal states or adjust output logits can be effective, they often incur substantial computational overhead, typically requiring extra forward passes. This efficiency bottleneck can limit their practicality for real-world, latency-sensitive deployments. In this work, we aim to address this trade-off with Residual-Update Directed DEcoding Regulation (RUDDER), a low-overhead framework that steers LVLMs towards visually-grounded generation. RUDDER is built on two key innovations: (1) Contextual Activation Residual Direction (CARD) vector, a per-sample visual evidence vector extracted from the residual update of a self-attention layer during a single, standard forward pass. (2) A Bayesian-inspired adaptive gate that performs token-wise injection, applying a corrective signal whose strength is conditioned on the model’s deviation from the visual context. Extensive experiments on key hallucination benchmarks, including POPE and CHAIR, indicate that RUDDER achieves performance comparable to state-of-the-art methods while introducing negligible computational latency, validating RUDDER as a pragmatic and effective approach for improving LVLMs’ reliability without a significant compromise on efficiency.

Method: Two-phase test-time adaptation framework using Mixture-of-Experts architecture with shared backbone to capture region-specific characteristics while learning universal features. Uses prediction consistency across experts to filter unreliable pseudo-labels during adaptation.

Result: GRAM outperforms state-of-the-art baselines in low-resource settings like African cities, using a million-scale dataset from 12 cities across four continents.

Conclusion: GRAM provides a scalable and label-efficient solution for global slum mapping and data-driven urban planning by enabling robust segmentation without requiring labeled data from target regions.

Abstract: Satellite-based slum segmentation holds significant promise in generating global estimates of urban poverty. However, the morphological heterogeneity of informal settlements presents a major challenge, hindering the ability of models trained on specific regions to generalize effectively to unseen locations. To address this, we introduce a large-scale high-resolution dataset and propose GRAM (Generalized Region-Aware Mixture-of-Experts), a two-phase test-time adaptation framework that enables robust slum segmentation without requiring labeled data from target regions. We compile a million-scale satellite imagery dataset from 12 cities across four continents for source training. Using this dataset, the model employs a Mixture-of-Experts architecture to capture region-specific slum characteristics while learning universal features through a shared backbone. During adaptation, prediction consistency across experts filters out unreliable pseudo-labels, allowing the model to generalize effectively to previously unseen regions. GRAM outperforms state-of-the-art baselines in low-resource settings such as African cities, offering a scalable and label-efficient solution for global slum mapping and data-driven urban planning.

Method: Three key modifications: (1) separate QKV projections for vision modality, (2) bidirectional attention on visual tokens, (3) incorporation of both global and local visual representations.

Result: Significantly outperforms baseline LLaVA method on multiple benchmarks, even surpassing models with double the parameter count.

Conclusion: LLaViT establishes a more effective approach to vision-language modeling by enabling LLMs to simultaneously function as vision encoders.

Abstract: Despite the remarkable success of the LLaVA architecture for vision-language tasks, its design inherently struggles to effectively integrate visual features due to the inherent mismatch between text and vision modalities. We tackle this issue from a novel perspective in which the LLM not only serves as a language model but also a powerful vision encoder. To this end, we present LLaViT - Large Language Models as extended Vision Transformers - which enables the LLM to simultaneously function as a vision encoder through three key modifications: (1) learning separate QKV projections for vision modality, (2) enabling bidirectional attention on visual tokens, and (3) incorporating both global and local visual representations. Through extensive controlled experiments on a wide range of LLMs, we demonstrate that LLaViT significantly outperforms the baseline LLaVA method on a multitude of benchmarks, even surpassing models with double its parameter count, establishing a more effective approach to vision-language modeling.

Method: Used image segmentation to automatically distinguish between eight different error categories in pedestrian detection, and proposed new metrics for performance comparison along these categories. Applied this to compare various backbones for a simplified version of APD.

Result: Achieved state-of-the-art performance on CityPersons-reasonable dataset without extra training data using a simple architecture. The new metrics provide more fine-grained and robust model comparison, especially for safety-critical performance.

Conclusion: The proposed error categorization and metrics enable more realistic evaluation of pedestrian detectors, particularly for safety-critical applications in automated driving systems.

Abstract: Reliable pedestrian detection represents a crucial step towards automated driving systems. However, the current performance benchmarks exhibit weaknesses. The currently applied metrics for various subsets of a validation dataset prohibit a realistic performance evaluation of a DNN for pedestrian detection. As image segmentation supplies fine-grained information about a street scene, it can serve as a starting point to automatically distinguish between different types of errors during the evaluation of a pedestrian detector. In this work, eight different error categories for pedestrian detection are proposed and new metrics are proposed for performance comparison along these error categories. We use the new metrics to compare various backbones for a simplified version of the APD, and show a more fine-grained and robust way to compare models with each other especially in terms of safety-critical performance. We achieve SOTA on CityPersons-reasonable (without extra training data) by using a rather simple architecture.

Method: Proposes three components: 1) Text Semantic Generation (TSG) for visible images to enable visible-text alignment, 2) Infrared Feature Embedding (IFE) to rectify infrared features using text semantics, and 3) High-level Semantic Alignment (HSA) to refine id-related semantic alignment.

Result: Extensive experiments show CLIP4VI-ReID achieves superior performance compared to state-of-the-art methods on widely used VI-ReID datasets.

Conclusion: The proposed method effectively bridges the modality gap between visible and infrared images using text as an intermediate representation, achieving accurate cross-modal alignment and discriminative shared representations.

Abstract: This paper proposes a novel CLIP-driven modality-shared representation learning network named CLIP4VI-ReID for VI-ReID task, which consists of Text Semantic Generation (TSG), Infrared Feature Embedding (IFE), and High-level Semantic Alignment (HSA). Specifically, considering the huge gap in the physical characteristics between natural images and infrared images, the TSG is designed to generate text semantics only for visible images, thereby enabling preliminary visible-text modality alignment. Then, the IFE is proposed to rectify the feature embeddings of infrared images using the generated text semantics. This process injects id-related semantics into the shared image encoder, enhancing its adaptability to the infrared modality. Besides, with text serving as a bridge, it enables indirect visible-infrared modality alignment. Finally, the HSA is established to refine the high-level semantic alignment. This process ensures that the fine-tuned text semantics only contain id-related information, thereby achieving more accurate cross-modal alignment and enhancing the discriminability of the learned modal-shared representations. Extensive experimental results demonstrate that the proposed CLIP4VI-ReID achieves superior performance than other state-of-the-art methods on some widely used VI-ReID datasets.

Method: Two core components: (1) Depth-of-field Supervision using scale-recovered monocular depth estimator, defocus convolution, and depth-of-field loss; (2) Multi-View Consistency Supervision using LoFTR-based feature matching and least squares optimization.

Result: Achieves superior depth fidelity with 0.8 dB PSNR improvement over state-of-the-art on Waymo Open Dataset.

Conclusion: This framework bridges physical imaging principles and learning-based depth regularization, offering a scalable solution for complex depth stratification in urban environments.

Abstract: Three-dimensional reconstruction in scenes with extreme depth variations remains challenging due to inconsistent supervisory signals between near-field and far-field regions. Existing methods fail to simultaneously address inaccurate depth estimation in distant areas and structural degradation in close-range regions. This paper proposes a novel computational framework that integrates depth-of-field supervision and multi-view consistency supervision to advance 3D Gaussian Splatting. Our approach comprises two core components: (1) Depth-of-field Supervision employs a scale-recovered monocular depth estimator (e.g., Metric3D) to generate depth priors, leverages defocus convolution to synthesize physically accurate defocused images, and enforces geometric consistency through a novel depth-of-field loss, thereby enhancing depth fidelity in both far-field and near-field regions; (2) Multi-View Consistency Supervision employing LoFTR-based semi-dense feature matching to minimize cross-view geometric errors and enforce depth consistency via least squares optimization of reliable matched points. By unifying defocus physics with multi-view geometric constraints, our method achieves superior depth fidelity, demonstrating a 0.8 dB PSNR improvement over the state-of-the-art method on the Waymo Open Dataset. This framework bridges physical imaging principles and learning-based depth regularization, offering a scalable solution for complex depth stratification in urban environments.

Method: Proposes DSANet with: 1) Self-guided normality modeling branch that reconstructs video features using learned normal prototypes to separate normal patterns from anomalies; 2) Decoupled contrastive semantic alignment that temporally decomposes videos into event-centric and background-centric components for visual-language contrastive learning.

Result: Comprehensive experiments on XD-Violence and UCF-Crime benchmarks show DSANet outperforms existing state-of-the-art methods.

Conclusion: The proposed disentangled semantic alignment approach effectively enhances the distinguishability between normal and abnormal patterns at both coarse and fine granularities, achieving superior video anomaly detection performance.

Abstract: Recent advancements in weakly-supervised video anomaly detection have achieved remarkable performance by applying the multiple instance learning paradigm based on multimodal foundation models such as CLIP to highlight anomalous instances and classify categories. However, their objectives may tend to detect the most salient response segments, while neglecting to mine diverse normal patterns separated from anomalies, and are prone to category confusion due to similar appearance, leading to unsatisfactory fine-grained classification results. Therefore, we propose a novel Disentangled Semantic Alignment Network (DSANet) to explicitly separate abnormal and normal features from coarse-grained and fine-grained aspects, enhancing the distinguishability. Specifically, at the coarse-grained level, we introduce a self-guided normality modeling branch that reconstructs input video features under the guidance of learned normal prototypes, encouraging the model to exploit normality cues inherent in the video, thereby improving the temporal separation of normal patterns and anomalous events. At the fine-grained level, we present a decoupled contrastive semantic alignment mechanism, which first temporally decomposes each video into event-centric and background-centric components using frame-level anomaly scores and then applies visual-language contrastive learning to enhance class-discriminative representations. Comprehensive experiments on two standard benchmarks, namely XD-Violence and UCF-Crime, demonstrate that DSANet outperforms existing state-of-the-art methods.

Method: Developed DermAI - a lightweight smartphone app that performs real-time capture, annotation, and classification of skin lesions with on-device quality checks and local model adaptation using diverse clinical data.

Result: Models trained on public datasets failed to generalize to clinical samples, but fine-tuning with local data improved performance, demonstrating the value of standardized diverse data collection.

Conclusion: Standardized, diverse data collection aligned with healthcare needs is crucial for effective machine learning development in dermatology.

Abstract: AI-based dermatology adoption remains limited by biased datasets, variable image quality, and limited validation. We introduce DermAI, a lightweight, smartphone-based application that enables real-time capture, annotation, and classification of skin lesions during routine consultations. Unlike prior dermoscopy-focused tools, DermAI performs on-device quality checks, and local model adaptation. The DermAI clinical dataset, encompasses a wide range of skin tones, ethinicity and source devices. In preliminary experiments, models trained on public datasets failed to generalize to our samples, while fine-tuning with local data improved performance. These results highlight the importance of standardized, diverse data collection aligned with healthcare needs and oriented to machine learning development.

Method: Integrates von Mises-Fisher distribution to model directional features and Fourier transformation for amplitude/phase perturbation, combined with CLIP-guided semantic augmentation.

Result: Outperforms state-of-the-art methods on the diverse weather-driving benchmark, demonstrating improved cross-domain generalization.

Conclusion: The proposed framework effectively enhances SDG object detection by capturing domain-invariant semantic structures and improving feature robustness through frequency-domain augmentation.

Abstract: Single Domain Generalization (SDG) for object detection aims to train a model on a single source domain that can generalize effectively to unseen target domains. While recent methods like CLIP-based semantic augmentation have shown promise, they often overlook the underlying structure of feature distributions and frequency-domain characteristics that are critical for robustness. In this paper, we propose a novel framework that enhances SDG object detection by integrating the von Mises-Fisher (vMF) distribution and Fourier transformation into a CLIP-guided pipeline. Specifically, we model the directional features of object representations using vMF to better capture domain-invariant semantic structures in the embedding space. Additionally, we introduce a Fourier-based augmentation strategy that perturbs amplitude and phase components to simulate domain shifts in the frequency domain, further improving feature robustness. Our method not only preserves the semantic alignment benefits of CLIP but also enriches feature diversity and structural consistency across domains. Extensive experiments on the diverse weather-driving benchmark demonstrate that our approach outperforms the existing state-of-the-art method.

Method: Integrates three signals: OOD detection in input space, OOD detection in embedding space, and task-specific predictive uncertainty for burn scar segmentation.

Result: OOD scores correlate with lower performance in specific environmental conditions, uncertainty flags help discard poor predictions, and failures concentrate in underrepresented geographies like low-elevation zones and large river areas.

Conclusion: SHRUG-FM enables safer, more interpretable deployment of geospatial foundation models by bridging the gap between benchmark performance and real-world reliability.

Abstract: Geospatial foundation models for Earth observation often fail to perform reliably in environments underrepresented during pretraining. We introduce SHRUG-FM, a framework for reliability-aware prediction that integrates three complementary signals: out-of-distribution (OOD) detection in the input space, OOD detection in the embedding space and task-specific predictive uncertainty. Applied to burn scar segmentation, SHRUG-FM shows that OOD scores correlate with lower performance in specific environmental conditions, while uncertainty-based flags help discard many poorly performing predictions. Linking these flags to land cover attributes from HydroATLAS shows that failures are not random but concentrated in certain geographies, such as low-elevation zones and large river areas, likely due to underrepresentation in pretraining data. SHRUG-FM provides a pathway toward safer and more interpretable deployment of GFMs in climate-sensitive applications, helping bridge the gap between benchmark performance and real-world reliability.

Method: Introduces M3DSG with image-based relational edges, Key Subgraph Selection for efficient reasoning, Adaptive Vocabulary Update for open vocabulary support, Closed-Loop Reasoning for exploration, and Visibility-based Viewpoint Decision for last-mile navigation.

Result: Achieves state-of-the-art performance on GOAT-Bench and HM3D-OVON datasets, demonstrating superior navigation capabilities compared to existing methods.

Conclusion: MSGNav effectively addresses key limitations in zero-shot navigation by preserving visual evidence through multi-modal scene graphs and specialized modules, enabling robust open-vocabulary generalization with low training overhead.

Abstract: Embodied navigation is a fundamental capability for robotic agents operating. Real-world deployment requires open vocabulary generalization and low training overhead, motivating zero-shot methods rather than task-specific RL training. However, existing zero-shot methods that build explicit 3D scene graphs often compress rich visual observations into text-only relations, leading to high construction cost, irreversible loss of visual evidence, and constrained vocabularies. To address these limitations, we introduce the Multi-modal 3D Scene Graph (M3DSG), which preserves visual cues by replacing textual relational edges with dynamically assigned images. Built on M3DSG, we propose MSGNav, a zero-shot navigation system that includes a Key Subgraph Selection module for efficient reasoning, an Adaptive Vocabulary Update module for open vocabulary support, and a Closed-Loop Reasoning module for accurate exploration reasoning. Additionally, we further identify the last-mile problem in zero-shot navigation - determining the feasible target location with a suitable final viewpoint, and propose a Visibility-based Viewpoint Decision module to explicitly resolve it. Comprehensive experimental results demonstrate that MSGNav achieves state-of-the-art performance on GOAT-Bench and HM3D-OVON datasets. The open-source code will be publicly available.

Method: Two-stage parsing pipeline: 1) Large multimodal model jointly predicts document layout and reading order using visual information; 2) Localized recognition of text, formulas, and tables within detected regions. Specialized modules include visual consistency-based reinforcement learning for table structures, Image-Decoupled Table Parsing for tables with embedded images, and Type-Guided Table Merging for cross-page/column table reconstruction.

Result: Comprehensive experiments on OmniDocBench v1.5 show MonkeyOCR v1.5 achieves state-of-the-art performance, outperforming PPOCR-VL and MinerU 2.5 while demonstrating exceptional robustness in visually complex document scenarios.

Conclusion: MonkeyOCR v1.5 provides an effective unified framework for parsing complex document layouts, addressing key challenges in document intelligence through enhanced layout understanding and content recognition capabilities.

Abstract: Document parsing is a core task in document intelligence, supporting applications such as information extraction, retrieval-augmented generation, and automated document analysis. However, real-world documents often feature complex layouts with multi-level tables, embedded images or formulas, and cross-page structures, which remain challenging for existing OCR systems. We introduce MonkeyOCR v1.5, a unified vision-language framework that enhances both layout understanding and content recognition through a two-stage parsing pipeline. The first stage employs a large multimodal model to jointly predict document layout and reading order, leveraging visual information to ensure structural and sequential consistency. The second stage performs localized recognition of text, formulas, and tables within detected regions, maintaining high visual fidelity while reducing error propagation. To address complex table structures, we propose a visual consistency-based reinforcement learning scheme that evaluates recognition quality via render-and-compare alignment, improving structural accuracy without manual annotations. Additionally, two specialized modules, Image-Decoupled Table Parsing and Type-Guided Table Merging, are introduced to enable reliable parsing of tables containing embedded images and reconstruction of tables crossing pages or columns. Comprehensive experiments on OmniDocBench v1.5 demonstrate that MonkeyOCR v1.5 achieves state-of-the-art performance, outperforming PPOCR-VL and MinerU 2.5 while showing exceptional robustness in visually complex document scenarios.

Method: Proposed AntiDB_Purify evaluation framework to systematically test existing defense methods under realistic purification threats including traditional image filters and adversarial purification techniques.

Result: All current defense methods fail to maintain protective effectiveness when subjected to purification attacks, revealing their fragility and the false security they provide.

Conclusion: Current privacy defenses for personalized AI generation are inadequate and there is an urgent need for more imperceptible and robust protection mechanisms to truly safeguard user identity.

Abstract: Personalized AI applications such as DreamBooth enable the generation of customized content from user images, but also raise significant privacy concerns, particularly the risk of facial identity leakage. Recent defense mechanisms like Anti-DreamBooth attempt to mitigate this risk by injecting adversarial perturbations into user photos to prevent successful personalization. However, we identify two critical yet overlooked limitations of these methods. First, the adversarial examples often exhibit perceptible artifacts such as conspicuous patterns or stripes, making them easily detectable as manipulated content. Second, the perturbations are highly fragile, as even a simple, non-learned filter can effectively remove them, thereby restoring the model’s ability to memorize and reproduce user identity. To investigate this vulnerability, we propose a novel evaluation framework, AntiDB_Purify, to systematically evaluate existing defenses under realistic purification threats, including both traditional image filters and adversarial purification. Results reveal that none of the current methods maintains their protective effectiveness under such threats. These findings highlight that current defenses offer a false sense of security and underscore the urgent need for more imperceptible and robust protections to safeguard user identity in personalized generation.

Method: Proposed SAMIRO (Spatial Attention Mutual Information Regularization with Oracle) - a plug-and-play method that transfers knowledge from a pretrained model while preserving domain-agnostic spatial information through mutual information regularization.

Result: Extensive experiments on CULane, Tusimple, and LLAMAS benchmarks show SAMIRO consistently improves performance across different models and datasets when integrated into various state-of-the-art lane detection approaches.

Conclusion: SAMIRO effectively enhances lane detection performance by knowledge transfer from pretrained models while maintaining spatial information, demonstrating consistent improvements across multiple models and datasets.

Abstract: Lane detection is an important topic in the future mobility solutions. Real-world environmental challenges such as background clutter, varying illumination, and occlusions pose significant obstacles to effective lane detection, particularly when relying on data-driven approaches that require substantial effort and cost for data collection and annotation. To address these issues, lane detection methods must leverage contextual and global information from surrounding lanes and objects. In this paper, we propose a Spatial Attention Mutual Information Regularization with a pre-trained model as an Oracle, called SAMIRO. SAMIRO enhances lane detection performance by transferring knowledge from a pretrained model while preserving domain-agnostic spatial information. Leveraging SAMIRO’s plug-and-play characteristic, we integrate it into various state-of-the-art lane detection approaches and conduct extensive experiments on major benchmarks such as CULane, Tusimple, and LLAMAS. The results demonstrate that SAMIRO consistently improves performance across different models and datasets. The code will be made available upon publication.

Method: Transformer-VAE architecture that incorporates PROSAIL radiative transfer model as differentiable physical decoder, trained exclusively on simulated data without real imagery.

Result: Achieves comparable performance to state-of-the-art methods on real-world datasets (FRM4Veg and BelSAR) for retrieving LAI and canopy chlorophyll content, without requiring in-situ labels or real image calibration.

Conclusion: Integrating physical models with deep networks enables cost-effective, self-supervised global vegetation monitoring without real training data, opening prospects for large-scale physically-constrained remote sensing.

Abstract: Accurate retrieval of vegetation biophysical variables from satellite imagery is crucial for ecosystem monitoring and agricultural management. In this work, we propose a physics-informed Transformer-VAE architecture to invert the PROSAIL radiative transfer model for simultaneous estimation of key canopy parameters from Sentinel-2 data. Unlike previous hybrid approaches that require real satellite images for self-supevised training. Our model is trained exclusively on simulated data, yet achieves performance on par with state-of-the-art methods that utilize real imagery. The Transformer-VAE incorporates the PROSAIL model as a differentiable physical decoder, ensuring that inferred latent variables correspond to physically plausible leaf and canopy properties. We demonstrate retrieval of leaf area index (LAI) and canopy chlorophyll content (CCC) on real-world field datasets (FRM4Veg and BelSAR) with accuracy comparable to models trained with real Sentinel-2 data. Our method requires no in-situ labels or calibration on real images, offering a cost-effective and self-supervised solution for global vegetation monitoring. The proposed approach illustrates how integrating physical models with advanced deep networks can improve the inversion of RTMs, opening new prospects for large-scale, physically-constrained remote sensing of vegetation traits.

Method: Proposes Ground Diffusion (GrounDiff) using diffusion models for denoising-based filtering, with gated design and confidence-guided generation. Introduces Prior-Guided Stitching (PrioStitch) for scalability using downsampled global priors to guide local predictions.

Result: Reduces RMSE by up to 93% on ALS2DTM and 47% on USGS benchmarks. For road reconstruction, achieves up to 81% lower distance error on GeRoD benchmark while maintaining competitive surface smoothness without task-specific optimization.

Conclusion: GrounDiff demonstrates superior performance over existing methods for DSM-to-DTM translation and road reconstruction, offering a scalable diffusion-based solution that eliminates the need for manual parameter tuning or complex post-processing.

Abstract: Digital Terrain Models (DTMs) represent the bare-earth elevation and are important in numerous geospatial applications. Such data models cannot be directly measured by sensors and are typically generated from Digital Surface Models (DSMs) derived from LiDAR or photogrammetry. Traditional filtering approaches rely on manually tuned parameters, while learning-based methods require well-designed architectures, often combined with post-processing. To address these challenges, we introduce Ground Diffusion (GrounDiff), the first diffusion-based framework that iteratively removes non-ground structures by formulating the problem as a denoising task. We incorporate a gated design with confidence-guided generation that enables selective filtering. To increase scalability, we further propose Prior-Guided Stitching (PrioStitch), which employs a downsampled global prior automatically generated using GrounDiff to guide local high-resolution predictions. We evaluate our method on the DSM-to-DTM translation task across diverse datasets, showing that GrounDiff consistently outperforms deep learning-based state-of-the-art methods, reducing RMSE by up to 93% on ALS2DTM and up to 47% on USGS benchmarks. In the task of road reconstruction, which requires both high precision and smoothness, our method achieves up to 81% lower distance error compared to specialized techniques on the GeRoD benchmark, while maintaining competitive surface smoothness using only DSM inputs, without task-specific optimization. Our variant for road reconstruction, GrounDiff+, is specifically designed to produce even smoother surfaces, further surpassing state-of-the-art methods. The project page is available at https://deepscenario.github.io/GrounDiff/.

Method: Used four deep learning models to automatically segment pancreas in 584 patients, automatically detect PSL, and develop multivariate models with CT biomarkers for T2DM prediction.

Result: PSL was significantly higher in diabetic patients (4.26±8.32 vs 3.19±3.62, p=0.01). PancAP model achieved best segmentation (Dice: 0.79±0.17). Multivariate model attained 0.90 AUC, 66.7% sensitivity, and 91.9% specificity for T2DM prediction.

Conclusion: PSL is useful for T2DM screening and could potentially help predict early onset of Type 2 Diabetes.

Abstract: Type 2 Diabetes Mellitus (T2DM) is a chronic metabolic disease that affects millions of people worldwide. Early detection is crucial as it can alter pancreas function through morphological changes and increased deposition of ectopic fat, eventually leading to organ damage. While studies have shown an association between T2DM and pancreas volume and fat content, the role of increased pancreatic surface lobularity (PSL) in patients with T2DM has not been fully investigated. In this pilot work, we propose a fully automated approach to delineate the pancreas and other abdominal structures, derive CT imaging biomarkers, and opportunistically screen for T2DM. Four deep learning-based models were used to segment the pancreas in an internal dataset of 584 patients (297 males, 437 non-diabetic, age: 45$\pm$15 years). PSL was automatically detected and it was higher for diabetic patients (p=0.01) at 4.26 $\pm$ 8.32 compared to 3.19 $\pm$ 3.62 for non-diabetic patients. The PancAP model achieved the highest Dice score of 0.79 $\pm$ 0.17 and lowest ASSD error of 1.94 $\pm$ 2.63 mm (p$<$0.05). For predicting T2DM, a multivariate model trained with CT biomarkers attained 0.90 AUC, 66.7% sensitivity, and 91.9% specificity. Our results suggest that PSL is useful for T2DM screening and could potentially help predict the early onset of T2DM.

Method: Integrated YOLOMS with multi-scale detection and sliding-window cropping for fault feature extraction, plus a lightweight KV mapping module to convert visual outputs to structured text, followed by domain-tuned LLM for semantic reasoning.

Result: Achieved 90.6% fault detection accuracy and generated maintenance reports with 89% average accuracy on real-world datasets.

Conclusion: The framework improves diagnostic interpretability and provides practical decision support for wind turbine operation and maintenance.

Abstract: The health condition of wind turbine (WT) components is crucial for ensuring stable and reliable operation. However, existing fault detection methods are largely limited to visual recognition, producing structured outputs that lack semantic interpretability and fail to support maintenance decision-making. To address these limitations, this study proposes an integrated framework that combines YOLOMS with a large language model (LLM) for intelligent fault analysis and diagnosis. Specifically, YOLOMS employs multi-scale detection and sliding-window cropping to enhance fault feature extraction, while a lightweight key-value (KV) mapping module bridges the gap between visual outputs and textual inputs. This module converts YOLOMS detection results into structured textual representations enriched with both qualitative and quantitative attributes. A domain-tuned LLM then performs semantic reasoning to generate interpretable fault analyses and maintenance recommendations. Experiments on real-world datasets demonstrate that the proposed framework achieves a fault detection accuracy of 90.6% and generates maintenance reports with an average accuracy of 89%, thereby improving the interpretability of diagnostic results and providing practical decision support for the operation and maintenance of wind turbines.

Method: GT++ estimates facial planes from 3D US volumes using annotated anatomical landmarks, while 3DFETUS is a deep learning model that automates plane localization in 3D fetal US volumes.

Result: Achieved mean translation error of 4.13 mm and mean rotation error of 7.93 degrees per plane, outperforming state-of-the-art methods. Clinical assessments confirmed statistically significant improvements in plane estimation accuracy.

Conclusion: The proposed approach effectively addresses challenges in fetal facial plane acquisition, providing automated and standardized localization with improved accuracy compared to existing methods.

Abstract: Acquiring standard facial planes during routine fetal ultrasound (US) examinations is often challenging due to fetal movement, variability in orientation, and operator-dependent expertise. These factors contribute to inconsistencies, increased examination time, and potential diagnostic bias. To address these challenges in the context of facial assessment, we present: 1) GT++, a robust algorithm that estimates standard facial planes from 3D US volumes using annotated anatomical landmarks; and 2) 3DFETUS, a deep learning model that automates and standardizes their localization in 3D fetal US volumes. We evaluated our methods both qualitatively, through expert clinical review, and quantitatively. The proposed approach achieved a mean translation error of 4.13 mm and a mean rotation error of 7.93 degrees per plane, outperforming other state-of-the-art methods on 3D US volumes. Clinical assessments further confirmed the effectiveness of both GT++ and 3DFETUS, demonstrating statistically significant improvements in plane estimation accuracy.

Method: Collected 13,053 video clips (10,101 negative, 2,952 positive) from 19 subjects, used five-fold cross-validation with subject-wise partitioning, and employed TimeSformer transformer-based video classifier.

Result: Achieved 97% average F1-score for discriminating between seizure and normal activity in rodents.

Conclusion: The dataset and baseline code are publicly released to enable reproducible research on video-based seizure monitoring in preclinical epilepsy studies.

Abstract: We introduce a curated video dataset of laboratory rodents for automatic detection of convulsive events. The dataset contains short (10~s) top-down and side-view video clips of individual rodents, labeled at clip level as normal activity or seizure. It includes 10,101 negative samples and 2,952 positive samples collected from 19 subjects. We describe the data curation, annotation protocol and preprocessing pipeline, and report baseline experiments using a transformer-based video classifier (TimeSformer). Experiments employ five-fold cross-validation with strict subject-wise partitioning to prevent data leakage (no subject appears in more than one fold). Results show that the TimeSformer architecture enables discrimination between seizure and normal activity with an average F1-score of 97%. The dataset and baseline code are publicly released to support reproducible research on non-invasive, video-based monitoring in preclinical epilepsy research. RodEpil Dataset access - DOI: 10.5281/zenodo.17601357

Method: Semantic segmentation to extract glands from whole slide images as input patches for multiple-instance learning (MIL) and phenotyping.

Result: Achieved 0.83 Dice score for gland segmentation, improved MIL model AUCs by 10% for detecting copy number variations, and identified 15 gland clusters associated with cancer relapse and mutations.

Conclusion: HIT improves accuracy and interpretability of MIL predictions while streamlining computations by focusing on biologically meaningful structures.

Abstract: Histopathologists establish cancer grade by assessing histological structures, such as glands in prostate cancer. Yet, digital pathology pipelines often rely on grid-based tiling that ignores tissue architecture. This introduces irrelevant information and limits interpretability. We introduce histology-informed tiling (HIT), which uses semantic segmentation to extract glands from whole slide images (WSIs) as biologically meaningful input patches for multiple-instance learning (MIL) and phenotyping. Trained on 137 samples from the ProMPT cohort, HIT achieved a gland-level Dice score of 0.83 +/- 0.17. By extracting 380,000 glands from 760 WSIs across ICGC-C and TCGA-PRAD cohorts, HIT improved MIL models AUCs by 10% for detecting copy number variation (CNVs) in genes related to epithelial-mesenchymal transitions (EMT) and MYC, and revealed 15 gland clusters, several of which were associated with cancer relapse, oncogenic mutations, and high Gleason. Therefore, HIT improved the accuracy and interpretability of MIL predictions, while streamlining computations by focussing on biologically meaningful structures during feature extraction.

Method: Train a discrete style codebook from image collections to extract style embeddings, use these as conditions for text-to-image diffusion models, then train an autoregressive style generator on discrete embeddings to synthesize novel style embeddings from numerical codes.

Result: CoTyle effectively turns numerical codes into style controllers, demonstrating unparalleled simplicity and diversity while unlocking a vast space of reproducible styles from minimal input.

Conclusion: A style is worth one numerical code, as CoTyle proves that numerical codes can effectively control visual style generation, offering a simple yet powerful approach to creative visual stylization.

Abstract: Innovative visual stylization is a cornerstone of artistic creation, yet generating novel and consistent visual styles remains a significant challenge. Existing generative approaches typically rely on lengthy textual prompts, reference images, or parameter-efficient fine-tuning to guide style-aware image generation, but often struggle with style consistency, limited creativity, and complex style representations. In this paper, we affirm that a style is worth one numerical code by introducing the novel task, code-to-style image generation, which produces images with novel, consistent visual styles conditioned solely on a numerical style code. To date, this field has only been primarily explored by the industry (e.g., Midjourney), with no open-source research from the academic community. To fill this gap, we propose CoTyle, the first open-source method for this task. Specifically, we first train a discrete style codebook from a collection of images to extract style embeddings. These embeddings serve as conditions for a text-to-image diffusion model (T2I-DM) to generate stylistic images. Subsequently, we train an autoregressive style generator on the discrete style embeddings to model their distribution, allowing the synthesis of novel style embeddings. During inference, a numerical style code is mapped to a unique style embedding by the style generator, and this embedding guides the T2I-DM to generate images in the corresponding style. Unlike existing methods, our method offers unparalleled simplicity and diversity, unlocking a vast space of reproducible styles from minimal input. Extensive experiments validate that CoTyle effectively turns a numerical code into a style controller, demonstrating a style is worth one code.

Method: Provides a modular implementation of SRGAN-style models with generators, discriminators, loss functions, and training schedules exposed through concise configuration files. Includes built-in hooks for logging, validation, and large-scene inference.

Result: A practical tool and benchmark implementation that ships with ready-to-use configurations for common remote sensing scenarios and sensible default settings for adversarial training on multispectral satellite data like Sentinel-2.

Conclusion: The framework enables easier experimentation with SRGANs, reproducible model comparisons, and deployment of super-resolution pipelines across diverse Earth-observation datasets through its configuration-driven workflow.

Abstract: We present OpenSR-SRGAN, an open and modular framework for single-image super-resolution in Earth Observation. The software provides a unified implementation of SRGAN-style models that is easy to configure, extend, and apply to multispectral satellite data such as Sentinel-2. Instead of requiring users to modify model code, OpenSR-SRGAN exposes generators, discriminators, loss functions, and training schedules through concise configuration files, making it straightforward to switch between architectures, scale factors, and band setups. The framework is designed as a practical tool and benchmark implementation rather than a state-of-the-art model. It ships with ready-to-use configurations for common remote sensing scenarios, sensible default settings for adversarial training, and built-in hooks for logging, validation, and large-scene inference. By turning GAN-based super-resolution into a configuration-driven workflow, OpenSR-SRGAN lowers the entry barrier for researchers and practitioners who wish to experiment with SRGANs, compare models in a reproducible way, and deploy super-resolution pipelines across diverse Earth-observation datasets.

Method: Couples unrolled TV solver with data-driven Lambda Mapping Network (LambdaNet) that predicts per-pixel regularization maps, trained end-to-end for joint optimization of reconstruction and regularization.

Result: Experiments on DeepLesion dataset with realistic noise model show +2.9 dB PSNR and +6% SSIM improvements over classical TV and FBP+U-Net methods.

Conclusion: LTV provides interpretable alternative to black-box CNNs and serves as basis for 3D and data-consistency-driven reconstruction.

Abstract: Although Total Variation (TV) performs well in noise reduction and edge preservation on images, its dependence on the lambda parameter limits its efficiency and makes it difficult to use effectively. In this study, we present a Learnable Total Variation (LTV) framework that couples an unrolled TV solver with a data-driven Lambda Mapping Network (LambdaNet) predicting a per-pixel regularization map. The pipeline is trained end-to-end so that reconstruction and regularization are optimized jointly, yielding spatially adaptive smoothing: strong in homogeneous regions, relaxed near anatomical boundaries. Experiments on the DeepLesion dataset, using a realistic noise model adapted from the LoDoPaB-CT methodology, show consistent gains over classical TV and FBP+U-Net: +2.9 dB PSNR and +6% SSIM on average. LTV provides an interpretable alternative to black-box CNNs and a basis for 3D and data-consistency-driven reconstruction. Our codes are available at: https://github.com/itu-biai/deep_tv_for_ldct

Method: 1) Semantic-guided Dual Visual Pruner (SD-Pruner) with Instruction-driven Pruner for global action cues and Spatial-aggregation Pruner for geometry-rich features; 2) Semantic-complementary Hierarchical Fuser (SH-Fuser) to fuse dense patches and sparse tokens; 3) Semantic-conditioned Action Coupler (SA-Coupler) replacing conventional observation-to-DoF approach.

Result: SemanticVLA surpasses OpenVLA on LIBERO benchmark by 21.1% in success rate, while reducing training cost and inference latency by 3.0-fold and 2.7-fold. Sets new SOTA in both performance and efficiency.

Conclusion: The proposed SemanticVLA framework effectively addresses key limitations in VLA models through semantic-aligned sparsification and enhancement, demonstrating superior performance and efficiency for robotic manipulation tasks.

Abstract: Vision-Language-Action (VLA) models have advanced in robotic manipulation, yet practical deployment remains hindered by two key limitations: 1) perceptual redundancy, where irrelevant visual inputs are processed inefficiently, and 2) superficial instruction-vision alignment, which hampers semantic grounding of actions. In this paper, we propose SemanticVLA, a novel VLA framework that performs Semantic-Aligned Sparsification and Enhancement for Efficient Robotic Manipulation. Specifically: 1) To sparsify redundant perception while preserving semantic alignment, Semantic-guided Dual Visual Pruner (SD-Pruner) performs: Instruction-driven Pruner (ID-Pruner) extracts global action cues and local semantic anchors in SigLIP; Spatial-aggregation Pruner (SA-Pruner) compacts geometry-rich features into task-adaptive tokens in DINOv2. 2) To exploit sparsified features and integrate semantics with spatial geometry, Semantic-complementary Hierarchical Fuser (SH-Fuser) fuses dense patches and sparse tokens across SigLIP and DINOv2 for coherent representation. 3) To enhance the transformation from perception to action, Semantic-conditioned Action Coupler (SA-Coupler) replaces the conventional observation-to-DoF approach, yielding more efficient and interpretable behavior modeling for manipulation tasks. Extensive experiments on simulation and real-world tasks show that SemanticVLA sets a new SOTA in both performance and efficiency. SemanticVLA surpasses OpenVLA on LIBERO benchmark by 21.1% in success rate, while reducing training cost and inference latency by 3.0-fold and 2.7-fold.SemanticVLA is open-sourced and publicly available at https://github.com/JiuTian-VL/SemanticVLA

Method: StM strategy with dedicated information mapping mechanism using shared transformation functions for Gaussian attributes to unify separately modeled human and scene components, avoiding exhaustive pairwise interactions.

Result: Significantly outperforms state-of-the-art methods in visual quality and rendering accuracy, especially at challenging human-scene interaction boundaries, as demonstrated on monocular video datasets.

Conclusion: The Separate-then-Map approach effectively bridges separately defined models through information mapping, achieving superior reconstruction quality for dynamic human-scene interactions from monocular videos.

Abstract: Reconstructing dynamic humans interacting with real-world environments from monocular videos is an important and challenging task. Despite considerable progress in 4D neural rendering, existing approaches either model dynamic scenes holistically or model scenes and backgrounds separately aim to introduce parametric human priors. However, these approaches either neglect distinct motion characteristics of various components in scene especially human, leading to incomplete reconstructions, or ignore the information exchange between the separately modeled components, resulting in spatial inconsistencies and visual artifacts at human-scene boundaries. To address this, we propose {\bf Separate-then-Map} (StM) strategy that introduces a dedicated information mapping mechanism to bridge separately defined and optimized models. Our method employs a shared transformation function for each Gaussian attribute to unify separately modeled components, enhancing computational efficiency by avoiding exhaustive pairwise interactions while ensuring spatial and visual coherence between humans and their surroundings. Extensive experiments on monocular video datasets demonstrate that StM significantly outperforms existing state-of-the-art methods in both visual quality and rendering accuracy, particularly at challenging human-scene interaction boundaries.

Method: Developed a framework with: (1) human evaluation template for nuanced diversity assessment, (2) curated prompt set with identified factors of variation, (3) methodology for model comparison using binomial tests, and (4) comparison of image embeddings for diversity measurement.

Result: The framework enables principled ranking of T2I models by diversity and identifies specific categories where models struggle, providing systematic diversity evaluation capabilities.

Conclusion: This research provides a robust methodology for T2I diversity assessment that can guide improvements in model diversity and metric development.

Abstract: Despite advances in generation quality, current text-to-image (T2I) models often lack diversity, generating homogeneous outputs. This work introduces a framework to address the need for robust diversity evaluation in T2I models. Our framework systematically assesses diversity by evaluating individual concepts and their relevant factors of variation. Key contributions include: (1) a novel human evaluation template for nuanced diversity assessment; (2) a curated prompt set covering diverse concepts with their identified factors of variation (e.g. prompt: An image of an apple, factor of variation: color); and (3) a methodology for comparing models in terms of human annotations via binomial tests. Furthermore, we rigorously compare various image embeddings for diversity measurement. Notably, our principled approach enables ranking of T2I models by diversity, identifying categories where they particularly struggle. This research offers a robust methodology and insights, paving the way for improvements in T2I model diversity and metric development.

Method: Proposes GeoAdapter with zero-initialized convolutions to encode geometric information without disrupting foundation model representations, and stochastic multimodal fusion that randomly samples modality subsets during training for robust learning.

Result: Outperforms prior methods with auxiliary inputs, achieves SOTA results even with RGB-only input, and enhances VLA models for robotic tasks with consistent performance gains.

Conclusion: OmniVGGT effectively leverages geometric modalities to improve 3D vision tasks while maintaining efficiency, demonstrating practical utility in robotic applications through enhanced VLA models.

Abstract: General 3D foundation models have started to lead the trend of unifying diverse vision tasks, yet most assume RGB-only inputs and ignore readily available geometric cues (e.g., camera intrinsics, poses, and depth maps). To address this issue, we introduce OmniVGGT, a novel framework that can effectively benefit from an arbitrary number of auxiliary geometric modalities during both training and inference. In our framework, a GeoAdapter is proposed to encode depth and camera intrinsics/extrinsics into a spatial foundation model. It employs zero-initialized convolutions to progressively inject geometric information without disrupting the foundation model’s representation space. This design ensures stable optimization with negligible overhead, maintaining inference speed comparable to VGGT even with multiple additional inputs. Additionally, a stochastic multimodal fusion regimen is proposed, which randomly samples modality subsets per instance during training. This enables an arbitrary number of modality inputs during testing and promotes learning robust spatial representations instead of overfitting to auxiliary cues. Comprehensive experiments on monocular/multi-view depth estimation, multi-view stereo, and camera pose estimation demonstrate that OmniVGGT outperforms prior methods with auxiliary inputs and achieves state-of-the-art results even with RGB-only input. To further highlight its practical utility, we integrated OmniVGGT into vision-language-action (VLA) models. The enhanced VLA model by OmniVGGT not only outperforms the vanilla point-cloud-based baseline on mainstream benchmarks, but also effectively leverages accessible auxiliary inputs to achieve consistent gains on robotic tasks.

Method: M&M-3D constructs malignancy-guided 3D features and learns 3D reasoning by repeatedly mixing these features with slice-level information through modified operations in M&M without adding parameters, enabling direct weight transfer from FFDM.

Result: M&M-3D surpasses 2D projection and 3D slice-based methods by 11-54% for localization and 3-10% for classification, outperforms complex 3D variants by 20-47% for localization and 2-10% for classification in low-data regime, and beats previous top baseline by 4% for classification and 10% for localization on BCS-DBT benchmark.

Conclusion: M&M-3D effectively addresses data scarcity in DBT by enabling parameter-free 3D reasoning with direct weight transfer from FFDM models, achieving state-of-the-art performance in breast cancer detection tasks.

Abstract: Digital Breast Tomosynthesis (DBT) enhances finding visibility for breast cancer detection by providing volumetric information that reduces the impact of overlapping tissues; however, limited annotated data has constrained the development of deep learning models for DBT. To address data scarcity, existing methods attempt to reuse 2D full-field digital mammography (FFDM) models by either flattening DBT volumes or processing slices individually, thus discarding volumetric information. Alternatively, 3D reasoning approaches introduce complex architectures that require more DBT training data. Tackling these drawbacks, we propose M&M-3D, an architecture that enables learnable 3D reasoning while remaining parameter-free relative to its FFDM counterpart, M&M. M&M-3D constructs malignancy-guided 3D features, and 3D reasoning is learned through repeatedly mixing these 3D features with slice-level information. This is achieved by modifying operations in M&M without adding parameters, thus enabling direct weight transfer from FFDM. Extensive experiments show that M&M-3D surpasses 2D projection and 3D slice-based methods by 11-54% for localization and 3-10% for classification. Additionally, M&M-3D outperforms complex 3D reasoning variants by 20-47% for localization and 2-10% for classification in the low-data regime, while matching their performance in high-data regime. On the popular BCS-DBT benchmark, M&M-3D outperforms previous top baseline by 4% for classification and 10% for localization.

Method: Three key components: 1) OBIA-Mamba using superpixels as Mamba tokens to reduce redundancy while preserving details, 2) Global-local dual-branch CNN-Mamba architecture for joint modeling of local spatial details and global context, 3) Multitask optimization with dual loss functions to balance local precision and global consistency.

Result: Tested on Sentinel-2 imagery in Alberta, Canada, the proposed approach achieved higher classification accuracy and finer details compared to several advanced state-of-the-art classification methods.

Conclusion: The MSOM framework effectively addresses key challenges in Sentinel-2 land use classification by integrating object-based analysis, global-local modeling, and multitask optimization, demonstrating superior performance over existing methods.

Abstract: Although Sentinel-2 based land use and land cover (LULC) classification is critical for various environmental monitoring applications, it is a very difficult task due to some key data challenges (e.g., spatial heterogeneity, context information, signature ambiguity). This paper presents a novel Multitask Glocal OBIA-Mamba (MSOM) for enhanced Sentinel-2 classification with the following contributions. First, an object-based image analysis (OBIA) Mamba model (OBIA-Mamba) is designed to reduce redundant computation without compromising fine-grained details by using superpixels as Mamba tokens. Second, a global-local (GLocal) dual-branch convolutional neural network (CNN)-mamba architecture is designed to jointly model local spatial detail and global contextual information. Third, a multitask optimization framework is designed to employ dual loss functions to balance local precision with global consistency. The proposed approach is tested on Sentinel-2 imagery in Alberta, Canada, in comparison with several advanced classification approaches, and the results demonstrate that the proposed approach achieves higher classification accuracy and finer details that the other state-of-the-art methods.

Method: A lightweight Swin-style backbone with scale-specific pixel-shuffle heads performs super-resolution directly on the generator’s latent code before VAE decoding, requiring no modifications to the base model.

Result: LUA achieves comparable perceptual quality to pixel-space SR while reducing decoding and upscaling time by nearly 3x (0.42s vs 1.87s for 1024px generation from 512px), and generalizes well across different VAEs.

Conclusion: LUA provides an efficient and practical solution for scalable high-fidelity image synthesis in diffusion models, closely matching native high-resolution generation quality with significantly reduced latency.

Abstract: Diffusion models struggle to scale beyond their training resolutions, as direct high-resolution sampling is slow and costly, while post-hoc image super-resolution (ISR) introduces artifacts and additional latency by operating after decoding. We present the Latent Upscaler Adapter (LUA), a lightweight module that performs super-resolution directly on the generator’s latent code before the final VAE decoding step. LUA integrates as a drop-in component, requiring no modifications to the base model or additional diffusion stages, and enables high-resolution synthesis through a single feed-forward pass in latent space. A shared Swin-style backbone with scale-specific pixel-shuffle heads supports 2x and 4x factors and remains compatible with image-space SR baselines, achieving comparable perceptual quality with nearly 3x lower decoding and upscaling time (adding only +0.42 s for 1024 px generation from 512 px, compared to 1.87 s for pixel-space SR using the same SwinIR architecture). Furthermore, LUA shows strong generalization across the latent spaces of different VAEs, making it easy to deploy without retraining from scratch for each new decoder. Extensive experiments demonstrate that LUA closely matches the fidelity of native high-resolution generation while offering a practical and efficient path to scalable, high-fidelity image synthesis in modern diffusion pipelines.

Method: Uses a single plain transformer (vanilla DINO encoder) as backbone without architectural specialization, employs singular depth-ray prediction target to avoid complex multi-task learning, and implements teacher-student training paradigm.

Result: Sets new state-of-the-art on visual geometry benchmark: 44.3% improvement in camera pose accuracy and 25.1% improvement in geometric accuracy over prior SOTA VGGT, and outperforms DA2 in monocular depth estimation.

Conclusion: DA3 demonstrates that minimal modeling with simple transformer backbones and unified prediction targets can achieve superior geometry prediction performance across multiple tasks while maintaining spatial consistency.

Abstract: We present Depth Anything 3 (DA3), a model that predicts spatially consistent geometry from an arbitrary number of visual inputs, with or without known camera poses. In pursuit of minimal modeling, DA3 yields two key insights: a single plain transformer (e.g., vanilla DINO encoder) is sufficient as a backbone without architectural specialization, and a singular depth-ray prediction target obviates the need for complex multi-task learning. Through our teacher-student training paradigm, the model achieves a level of detail and generalization on par with Depth Anything 2 (DA2). We establish a new visual geometry benchmark covering camera pose estimation, any-view geometry and visual rendering. On this benchmark, DA3 sets a new state-of-the-art across all tasks, surpassing prior SOTA VGGT by an average of 44.3% in camera pose accuracy and 25.1% in geometric accuracy. Moreover, it outperforms DA2 in monocular depth estimation. All models are trained exclusively on public academic datasets.

Method: Self-Consistency Sampling (SCS) introduces small visual perturbations and performs repeated truncation/resampling of trajectories, using agreement among resulting trajectories to compute a differentiable consistency score for policy updates.

Result: SCS improves accuracy by up to 7.7 percentage points on six multimodal benchmarks when integrated with RLOO, GRPO, and REINFORCE++ methods, with negligible extra computation cost.

Conclusion: SCS provides a simple, general solution for improving outcome-reward RL in multimodal large language models by addressing the unfaithful reasoning problem.

Abstract: Outcome-reward reinforcement learning (RL) is a common and increasingly significant way to refine the step-by-step reasoning of multimodal large language models (MLLMs). In the multiple-choice setting - a dominant format for multimodal reasoning benchmarks - the paradigm faces a significant yet often overlooked obstacle: unfaithful trajectories that guess the correct option after a faulty chain of thought receive the same reward as genuine reasoning, which is a flaw that cannot be ignored. We propose Self-Consistency Sampling (SCS) to correct this issue. For each question, SCS (i) introduces small visual perturbations and (ii) performs repeated truncation and resampling of an initial trajectory; agreement among the resulting trajectories yields a differentiable consistency score that down-weights unreliable traces during policy updates. Based on Qwen2.5-VL-7B-Instruct, plugging SCS into RLOO, GRPO, and REINFORCE++ series improves accuracy by up to 7.7 percentage points on six multimodal benchmarks with negligible extra computation. SCS also yields notable gains on both Qwen2.5-VL-3B-Instruct and InternVL3-8B, offering a simple, general remedy for outcome-reward RL in MLLMs.

Method: GUI-RC constructs spatial voting grids from multiple predictions to identify consensus regions. GUI-RCPO transforms these consistency patterns into rewards for test-time reinforcement learning, enabling iterative refinement during inference.

Result: GUI-RC improves accuracy by 2-3% across architectures on ScreenSpot benchmarks. GUI-RCPO achieves 3-6% accuracy improvements using only 1,272 unlabeled data points.

Conclusion: The approach demonstrates the untapped potential of test-time scaling and test-time reinforcement learning for GUI grounding, offering a promising path toward more data-efficient GUI agents.

Abstract: Graphical User Interface (GUI) grounding, the task of mapping natural language instructions to precise screen coordinates, is fundamental to autonomous GUI agents. While existing methods achieve strong performance through extensive supervised training or reinforcement learning with labeled rewards, they remain constrained by the cost and availability of pixel-level annotations. We observe that when models generate multiple predictions for the same GUI element, the spatial overlap patterns reveal implicit confidence signals that can guide more accurate localization. Leveraging this insight, we propose GUI-RC (Region Consistency), a test-time scaling method that constructs spatial voting grids from multiple sampled predictions to identify consensus regions where models show highest agreement. Without any training, GUI-RC improves accuracy by 2-3% across various architectures on ScreenSpot benchmarks. We further introduce GUI-RCPO (Region Consistency Policy Optimization), transforming these consistency patterns into rewards for test-time reinforcement learning. By computing how well each prediction aligns with the collective consensus, GUI-RCPO enables models to iteratively refine their outputs on unlabeled data during inference. Extensive experiments demonstrate the generality of our approach: using only 1,272 unlabeled data, GUI-RCPO achieves 3-6% accuracy improvements across various architectures on ScreenSpot benchmarks. Our approach reveals the untapped potential of test-time scaling and test-time reinforcement learning for GUI grounding, offering a promising path toward more data-efficient GUI agents.

Method: Proposes a multi-branch neural network with branch-orthogonal loss function to make each solution space orthogonal, integrating multiple orthogonal solution spaces for improved robustness.

Result: Achieves 67.3% robust accuracy on CIFAR-10 and 41.5% on CIFAR-100 against ℓ∞ norm-bounded perturbations (ε=8/255), improving state-of-the-art by +7.23% and +9.07% respectively without using additional data.

Conclusion: BORT demonstrates that focusing on model architecture and feature distribution through orthogonal solution spaces can achieve superior adversarial robustness compared to data-centric approaches, establishing new state-of-the-art performance.

Abstract: Deep neural networks (DNNs) are vulnerable to adversarial examples, in which DNNs are misled to false outputs due to inputs containing imperceptible perturbations. Adversarial training, a reliable and effective method of defense, may significantly reduce the vulnerability of neural networks and becomes the de facto standard for robust learning. While many recent works practice the data-centric philosophy, such as how to generate better adversarial examples or use generative models to produce additional training data, we look back to the models themselves and revisit the adversarial robustness from the perspective of deep feature distribution as an insightful complementarity. In this paper, we propose \textit{Branch Orthogonality adveRsarial Training} (BORT) to obtain state-of-the-art performance with solely the original dataset for adversarial training. To practice our design idea of integrating multiple orthogonal solution spaces, we leverage a simple multi-branch neural network and propose a corresponding loss function, branch-orthogonal loss, to make each solution space of the multi-branch model orthogonal. We evaluate our approach on CIFAR-10, CIFAR-100 and SVHN against $\ell_{\infty}$ norm-bounded perturbations of size $ε= 8/255$, respectively. Exhaustive experiments are conducted to show that our method goes beyond all state-of-the-art methods without any tricks. Compared to all methods that do not use additional data for training, our models achieve 67.3% and 41.5% robust accuracy on CIFAR-10 and CIFAR-100 (improving upon the state-of-the-art by +7.23% and +9.07%).

Method: GACD uses first-order Taylor gradients to estimate token-visual feature contributions, then suppresses spurious visual features and rebalances cross-modal contributions by strengthening visual features relative to text.

Result: Experiments across multiple benchmarks show GACD effectively reduces hallucinations and improves visual grounding of MLLM outputs.

Conclusion: GACD provides an effective inference-based solution to mitigate hallucinations in multimodal LLMs without auxiliary models or finetuning, addressing both text-visual and co-occurrence biases.

Abstract: Multimodal large language models achieve strong performance across diverse tasks but remain prone to hallucinations, where outputs are not grounded in visual inputs. This issue can be attributed to two main biases: text-visual bias, the overreliance on prompts and prior outputs, and co-occurrence bias, spurious correlations between frequently paired objects. We propose Gradient-based Influence-Aware Constrained Decoding (GACD), an inference-based method, that addresses both biases without auxiliary models, and is readily applicable to existing models without finetuning. The core of our approach is bias estimation, which uses first-order Taylor gradients to understand the contribution of individual tokens-visual features and text tokens-to the current output. Based on this analysis, GACD mitigates hallucinations through two components: (1) suppressing spurious visual features correlated with the output objects, and (2) rebalancing cross-modal contributions by strengthening visual features relative to text. Experiments across multiple benchmarks demonstrate that GACD effectively reduces hallucinations and improves the visual grounding of MLLM outputs.

Method: Attri-Net counterfactually generates class-specific attribution maps to highlight disease evidence, then performs classification using logistic regression based solely on these maps. It includes mechanisms to align explanations with human knowledge.

Result: Attri-Net generates high-quality explanations consistent with clinical knowledge without sacrificing classification performance.

Conclusion: Attri-Net successfully provides both local and global interpretability for medical AI applications while maintaining competitive classification accuracy.

Abstract: Interpretability is crucial for machine learning algorithms in high-stakes medical applications. However, high-performing neural networks typically cannot explain their predictions. Post-hoc explanation methods provide a way to understand neural networks but have been shown to suffer from conceptual problems. Moreover, current research largely focuses on providing local explanations for individual samples rather than global explanations for the model itself. In this paper, we propose Attri-Net, an inherently interpretable model for multi-label classification that provides local and global explanations. Attri-Net first counterfactually generates class-specific attribution maps to highlight the disease evidence, then performs classification with logistic regression classifiers based solely on the attribution maps. Local explanations for each prediction can be obtained by interpreting the attribution maps weighted by the classifiers’ weights. Global explanation of whole model can be obtained by jointly considering learned average representations of the attribution maps for each class (called the class centers) and the weights of the linear classifiers. To ensure the model is ``right for the right reason", we further introduce a mechanism to guide the model’s explanations to align with human knowledge. Our comprehensive evaluations show that Attri-Net can generate high-quality explanations consistent with clinical knowledge while not sacrificing classification performance.

Method: Uses text as a reasoning proxy to concurrently scan image collections, propose candidate criteria in natural language, and group images into meaningful clusters per criterion. Evaluated on new benchmarks COCO-4C and Food-4C.

Result: X-Cluster effectively reveals meaningful partitions on several datasets and enables real-world applications like uncovering biases in text-to-image models and analyzing image virality on social media.

Conclusion: The framework successfully demonstrates automatic discovery of multiple semantic clustering criteria from unstructured image data, with practical applications and open-sourced code/datasets for future research.

Abstract: In this work, we introduce and study the novel task of Open-ended Semantic Multiple Clustering (OpenSMC). Given a large, unstructured image collection, the goal is to automatically discover several, diverse semantic clustering criteria (e.g., Activity or Location) from the images, and subsequently organize them according to the discovered criteria, without requiring any human input. Our framework, X-Cluster: eXploratory Clustering, treats text as a reasoning proxy: it concurrently scans the entire image collection, proposes candidate criteria in natural language, and groups images into meaningful clusters per criterion. This radically differs from previous works, which either assume predefined clustering criteria or fixed cluster counts. To evaluate X-Cluster, we create two new benchmarks, COCO-4C and Food-4C, each annotated with four distinct grouping criteria and corresponding cluster labels. Experiments show that X-Cluster can effectively reveal meaningful partitions on several datasets. Finally, we use X-Cluster to achieve various real-world applications, including uncovering hidden biases in text-to-image (T2I) generative models and analyzing image virality on social media. Code and datasets will be open-sourced for future research.

Method: Records noise sequences and masking patterns during the reverse diffusion process to enable accurate reconstruction and editing.

Result: Demonstrated effectiveness across image and text domains with models like VQ-Diffusion, Paella, and RoBERTa, preserving high data fidelity while enhancing editing capabilities.

Conclusion: DICE offers new opportunities for fine-grained content manipulation in discrete spaces by enabling precise inversion for discrete diffusion models.

Abstract: Discrete diffusion models have achieved success in tasks like image generation and masked language modeling but face limitations in controlled content editing. We introduce DICE (Discrete Inversion for Controllable Editing), the first approach to enable precise inversion for discrete diffusion models, including multinomial diffusion and masked generative models. By recording noise sequences and masking patterns during the reverse diffusion process, DICE enables accurate reconstruction and flexible editing of discrete data without the need for predefined masks or attention manipulation. We demonstrate the effectiveness of DICE across both image and text domains, evaluating it on models such as VQ-Diffusion, Paella, and RoBERTa. Our results show that DICE preserves high data fidelity while enhancing editing capabilities, offering new opportunities for fine-grained content manipulation in discrete spaces.

Method: Synthesize virtual outliers by adding gradient attribution values to ID inputs to disrupt invariant features while amplifying true-class logit, then use standardized features with z-score normalization to incentivize ID classification and predictive uncertainty towards virtual outliers.

Result: Extensive experiments across 7 datasets and comparisons with 30+ methods demonstrate ASCOOD’s effectiveness in spurious, fine-grained, and conventional OOD detection settings.

Conclusion: ASCOOD successfully addresses challenging OOD detection scenarios without requiring external data, effectively mitigating spurious correlations and capturing fine-grained attributes.

Abstract: Out-of-distribution (OOD) detection is critical to ensure the safe deployment of deep learning models in critical applications. Deep learning models can often misidentify OOD samples as in-distribution (ID) samples. This vulnerability worsens in the presence of spurious correlation in the training set. Likewise, in fine-grained classification settings, detection of fine-grained OOD samples becomes inherently challenging due to their high similarity to ID samples. However, current research on OOD detection has focused instead largely on relatively easier (conventional) cases. Even the few recent works addressing these challenging cases rely on carefully curated or synthesized outliers, ultimately requiring external data. This motivates our central research question: ``Can we innovate OOD detection training framework for fine-grained and spurious settings \textbf{without requiring any external data at all?}" In this work, we present a unified \textbf{A}pproach to \textbf{S}purious, fine-grained, and \textbf{C}onventional \textbf{OOD D}etection (\textbf{\ASCOOD}) that eliminates the reliance on external data. First, we synthesize virtual outliers from ID data by approximating the destruction of invariant features. Specifically, we propose to add gradient attribution values to ID inputs to disrupt invariant features while amplifying true-class logit, thereby synthesizing challenging near-manifold virtual outliers. Then, we simultaneously incentivize ID classification and predictive uncertainty towards virtual outliers. For this, we further propose to leverage standardized features with z-score normalization. ASCOOD effectively mitigates impact of spurious correlations and encourages capturing fine-grained attributes. Extensive experiments across \textbf{7} datasets and and comparisons with \textbf{30+} methods demonstrate merit of ASCOOD in spurious, fine-grained and conventional settings.

Method: Reframe video understanding as self-supervised spatio-temporal clustering: extract semantic features using frozen VLM encoder, apply Kernel Temporal Segmentation for event segmentation, then use density-based clustering to identify recurring scenes.

Result: Automatically produces structured, multi-modal video summaries by selecting keyframes from discovered clusters and generating textual descriptions using VLM capabilities.

Conclusion: Provides an effective, interpretable, model-agnostic pathway for zero-shot automated structural analysis of video content without requiring end-to-end training.

Abstract: The remarkable zero-shot reasoning capabilities of large-scale Visual Language Models (VLMs) on static images have yet to be fully translated to the video domain. Conventional video understanding models often rely on extensive, task-specific training on annotated datasets, a process that is both costly and limited in scalability. This paper introduces a novel, training-free framework for video understanding that circumvents end-to-end training by synergistically combining the rich semantic priors of pre-trained VLMs with classic machine learning algorithms for pattern discovery. Our core idea is to reframe video understanding as a self-supervised spatio-temporal clustering problem within a high-dimensional semantic feature space. The proposed pipeline first transforms a video stream into a semantic feature trajectory using the frozen visual encoder of a pre-trained VLM. Subsequently, we employ Kernel Temporal Segmentation (KTS), a robust machine learning technique, to partition the continuous feature stream into discrete, semantically coherent event segments. These segments are then subjected to unsupervised density-based clustering to identify recurring macroscopic scenes and themes throughout the video. By selecting representative keyframes from each discovered cluster and leveraging the VLM’s generative capabilities for textual description, our framework automatically produces a structured, multi-modal summary of the video content. This approach provides an effective, interpretable, and model-agnostic pathway for zero-shot, automated structural analysis of video content.

Method: Extract static knowledge via pre-trained masked autoencoders and dynamic knowledge via optical flow prediction, then project visual embeddings to generate pseudo-language prompts in spatial and frequency domains for fine-tuning video diffusion models.

Result: Extensive experiments on computational fluid dynamics simulations and real-world typhoon observations demonstrate superior fidelity and consistency across diverse scientific scenarios.

Conclusion: The proposed framework successfully enables video diffusion models to generate scientific phenomena that better adhere to scientific laws by incorporating latent scientific knowledge through pseudo-language prompts.

Abstract: Video diffusion models have achieved impressive results in natural scene generation, yet they struggle to generalize to scientific phenomena such as fluid simulations and meteorological processes, where underlying dynamics are governed by scientific laws. These tasks pose unique challenges, including severe domain gaps, limited training data, and the lack of descriptive language annotations. To handle this dilemma, we extracted the latent scientific phenomena knowledge and further proposed a fresh framework that teaches video diffusion models to generate scientific phenomena from a single initial frame. Particularly, static knowledge is extracted via pre-trained masked autoencoders, while dynamic knowledge is derived from pre-trained optical flow prediction. Subsequently, based on the aligned spatial relations between the CLIP vision and language encoders, the visual embeddings of scientific phenomena, guided by latent scientific phenomena knowledge, are projected to generate the pseudo-language prompt embeddings in both spatial and frequency domains. By incorporating these prompts and fine-tuning the video diffusion model, we enable the generation of videos that better adhere to scientific laws. Extensive experiments on both computational fluid dynamics simulations and real-world typhoon observations demonstrate the effectiveness of our approach, achieving superior fidelity and consistency across diverse scientific scenarios.

Method: Uses a difference mask from surface normals to identify static/dynamic areas. For static areas, a Masked Static module enhances consistency by focusing on stable regions. For dynamic areas, a Surface Normal Similarity module aligns areas using feature similarity. Final refinement integrates both areas.

Result: Achieves state-of-the-art video depth estimation on KITTI and NYUv2 datasets without requiring additional information like optical flow or camera parameters.

Conclusion: STATIC effectively addresses temporal consistency in video depth estimation by separately handling static and dynamic regions, eliminating the need for external motion information while maintaining high performance.

Abstract: Video monocular depth estimation is essential for applications such as autonomous driving, AR/VR, and robotics. Recent transformer-based single-image monocular depth estimation models perform well on single images but struggle with depth consistency across video frames. Traditional methods aim to improve temporal consistency using multi-frame temporal modules or prior information like optical flow and camera parameters. However, these approaches face issues such as high memory use, reduced performance with dynamic or irregular motion, and limited motion understanding. We propose STATIC, a novel model that independently learns temporal consistency in static and dynamic area without additional information. A difference mask from surface normals identifies static and dynamic area by measuring directional variance. For static area, the Masked Static (MS) module enhances temporal consistency by focusing on stable regions. For dynamic area, the Surface Normal Similarity (SNS) module aligns areas and enhances temporal consistency by measuring feature similarity between frames. A final refinement integrates the independently learned static and dynamic area, enabling STATIC to achieve temporal consistency across the entire sequence. Our method achieves state-of-the-art video depth estimation on the KITTI and NYUv2 datasets without additional information.

Method: Gradually Pruning Queries (GPQ) - an embarrassingly simple approach that incrementally prunes queries based on their classification scores, requiring no additional learnable parameters and can be integrated as a fine-tuning step.

Result: GPQ reduces redundant queries while maintaining performance, achieving 1.35x inference acceleration on desktop GPUs, 67.86% FLOPs reduction, and 65.16% inference time decrease on edge devices.

Conclusion: GPQ provides an effective, parameter-free solution to reduce computational costs in query-based 3D detectors by pruning redundant queries, enabling efficient deployment on resource-constrained devices.

Abstract: Query-based models are extensively used in 3D object detection tasks, with a wide range of pre-trained checkpoints readily available online. However, despite their popularity, these models often require an excessive number of object queries, far surpassing the actual number of objects to detect. The redundant queries result in unnecessary computational and memory costs. In this paper, we find that not all queries contribute equally – a significant portion of queries have a much smaller impact compared to others. Based on this observation, we propose an embarrassingly simple approach called Gradually Pruning Queries (GPQ), which prunes queries incrementally based on their classification scores. A key advantage of GPQ is that it requires no additional learnable parameters. It is straightforward to implement in any query-based method, as it can be seamlessly integrated as a fine-tuning step using an existing checkpoint after training. With GPQ, users can easily generate multiple models with fewer queries, starting from a checkpoint with an excessive number of queries. Experiments on various advanced 3D detectors show that GPQ effectively reduces redundant queries while maintaining performance. Using our method, model inference on desktop GPUs can be accelerated by up to 1.35x. Moreover, after deployment on edge devices, it achieves up to a 67.86% reduction in FLOPs and a 65.16% decrease in inference time. The code will be available at https://github.com/iseri27/Gpq.

Method: Uses Generative Latent Prediction (GLP) architecture: autoregressive latent dynamics with LLM backbone for text-based reasoning, combined with video diffusion decoder for detailed visual reconstruction.

Result: Achieves strong performance in action-conditioned world simulation, long-horizon forecasting, and simulative reasoning compared to other models, supporting open-domain simulation with coherent dynamics.

Conclusion: PAN represents progress toward general world models that enable predictive simulation of future states for reasoning and acting across diverse domains.

Abstract: A world model enables an intelligent agent to imagine, predict, and reason about how the world evolves in response to its actions, and accordingly to plan and strategize. While recent video generation models produce realistic visual sequences, they typically operate in the prompt-to-full-video manner without causal control, interactivity, or long-horizon consistency required for purposeful reasoning. Existing world modeling efforts, on the other hand, often focus on restricted domains (e.g., physical, game, or 3D-scene dynamics) with limited depth and controllability, and struggle to generalize across diverse environments and interaction formats. In this work, we introduce PAN, a general, interactable, and long-horizon world model that predicts future world states through high-quality video simulation conditioned on history and natural language actions. PAN employs the Generative Latent Prediction (GLP) architecture that combines an autoregressive latent dynamics backbone based on a large language model (LLM), which grounds simulation in extensive text-based knowledge and enables conditioning on language-specified actions, with a video diffusion decoder that reconstructs perceptually detailed and temporally coherent visual observations, to achieve a unification between latent space reasoning (imagination) and realizable world dynamics (reality). Trained on large-scale video-action pairs spanning diverse domains, PAN supports open-domain, action-conditioned simulation with coherent, long-term dynamics. Extensive experiments show that PAN achieves strong performance in action-conditioned world simulation, long-horizon forecasting, and simulative reasoning compared to other video generators and world models, taking a step towards general world models that enable predictive simulation of future world states for reasoning and acting.

Method: Uses diffusion models to predict noise conditioned on input images and text prompts, computes loss for image-patch pairs, identifies distinctive regions via loss differences, and applies intra-class clustering and ranking for diversity.

Result: Consistently outperforms state-of-the-art methods across various metrics and settings with a streamlined one-step distillation process.

Conclusion: The proposed diffusion prior-based patch selection approach provides an effective orthogonal solution to existing distillation techniques, achieving superior performance with simplified implementation.

Abstract: Dataset distillation provides an effective approach to reduce memory and computational costs by optimizing a compact dataset that achieves performance comparable to the full original. However, for large-scale datasets and complex deep networks (e.g., ImageNet-1K with ResNet-101), the vast optimization space hinders distillation effectiveness, limiting practical applications. Recent methods leverage pre-trained diffusion models to directly generate informative images, thereby bypassing pixel-level optimization and achieving promising results. Nonetheless, these approaches often suffer from distribution shifts between the pre-trained diffusion prior and target datasets, as well as the need for multiple distillation steps under varying settings. To overcome these challenges, we propose a novel framework that is orthogonal to existing diffusion-based distillation techniques by utilizing the diffusion prior for patch selection rather than generation. Our method predicts noise from the diffusion model conditioned on input images and optional text prompts (with or without label information), and computes the associated loss for each image-patch pair. Based on the loss differences, we identify distinctive regions within the original images. Furthermore, we apply intra-class clustering and ranking on the selected patches to enforce diversity constraints. This streamlined pipeline enables a one-step distillation process. Extensive experiments demonstrate that our approach consistently outperforms state-of-the-art methods across various metrics and settings.

Method: Multi-View Structural Convolution Network (MSCN) with Structural Convolution Layers (SCL) for local geometric features and Structural Aggregation Layers (SAL) for local and overall context features, enhanced by training with unseen domain point clouds from source domain.

Result: Achieves 82.0% average accuracy in cross-domain experiments, surpassing PointTransformer baseline by 15.8%, demonstrating effectiveness under real-world domain shifts.

Conclusion: MSCN successfully enables domain-invariant point cloud recognition through structural feature extraction and domain adaptation training, proving robust performance across different domains.

Abstract: Point cloud representation has recently become a research hotspot in the field of computer vision and has been utilized for autonomous vehicles. However, adapting deep learning networks for point cloud data recognition is challenging due to the variability in datasets and sensor technologies. This variability underscores the necessity for adaptive techniques to maintain accuracy under different conditions. In this paper, we present the Multi-View Structural Convolution Network (MSCN) designed for domain-invariant point cloud recognition. MSCN comprises Structural Convolution Layers (SCL) that extract local context geometric features from point clouds and Structural Aggregation Layers (SAL) that extract and aggregate both local and overall context features from point clouds. Furthermore, MSCN enhances feature robustness by training with unseen domain point clouds generated from the source domain, enabling the model to acquire domain-invariant representations. Extensive cross-domain experiments demonstrate that MSCN achieves an average accuracy of 82.0%, surpassing the strong baseline PointTransformer by 15.8%, confirming its effectiveness under real-world domain shifts. Our code is available at https://github.com/MLMLab/MSCN.

Method: Proposes Multi-layer Concept Map (MCM) with asymmetric concept learning architecture, establishing correlations between different encoder/decoder layers and updating concept tokens using backward gradients from reconstruction tasks at various granularity levels.

Result: Significantly reduces computational costs by training on fewer than 75% of total image patches while enhancing concept prediction performance across various metrics. Enables targeted image generation from masked images by editing concept tokens.

Conclusion: MCM demonstrates that masked image training can efficiently improve concept learning, allowing flexible reconstruction blending visible patches with provided concepts through adjustable mask ratios.

Abstract: Masking strategies commonly employed in natural language processing are still underexplored in vision tasks such as concept learning, where conventional methods typically rely on full images. However, using masked images diversifies perceptual inputs, potentially offering significant advantages in concept learning with large-scale Transformer models. To this end, we propose Multi-layer Concept Map (MCM), the first work to devise an efficient concept learning method based on masked images. In particular, we introduce an asymmetric concept learning architecture by establishing correlations between different encoder and decoder layers, updating concept tokens using backward gradients from reconstruction tasks. The learned concept tokens at various levels of granularity help either reconstruct the masked image patches by filling in gaps or guide the reconstruction results in a direction that reflects specific concepts. Moreover, we present both quantitative and qualitative results across a wide range of metrics, demonstrating that MCM significantly reduces computational costs by training on fewer than 75% of the total image patches while enhancing concept prediction performance. Additionally, editing specific concept tokens in the latent space enables targeted image generation from masked images, aligning both the visible contextual patches and the provided concepts. By further adjusting the testing time mask ratio, we could produce a range of reconstructions that blend the visible patches with the provided concepts, proportional to the chosen ratios.

Method: Uses structured lasso with Information Bottleneck theory to guide pruning, proposing two schemes: sparse graph-structured lasso pruning (sGLP-IB) and sparse tree-guided lasso pruning (sTLP-IB) that preserve class-wise relatedness.

Result: Achieved 85% parameter reduction and 61% FLOPs reduction on VGG16/CIFAR-10 while improving accuracy to 94.10% (+0.14%). On ResNet/ImageNet, reduced parameters by 55% with only 0.03% accuracy drop to 76.12%. Outperformed multiple state-of-the-art methods across three datasets and six model structures.

Conclusion: Successfully reduces model size and computational resources while maintaining accuracy effectiveness by preserving class-wise statistical information through structured lasso pruning guided by Information Bottleneck theory.

Abstract: Modern applications require lightweight neural network models. Most existing neural network pruning methods focus on removing unimportant filters; however, these may result in the loss of statistical information after pruning due to failing to consider the class-wise information. In this paper, we employ the structured lasso from the perspective of utilizing precise class-wise information for model pruning with the help of Information Bottleneck theory, which guides us to ensure the retention of statistical information before and after pruning. With these techniques, we propose two novel adaptive network pruning schemes in parallel: sparse graph-structured lasso pruning with Information Bottleneck (sGLP-IB) and sparse tree-guided lasso pruning with Information Bottleneck (sTLP-IB). The key component is that we prune the model filters utilizing sGLP-IB and sTLP-IB with more precise structured class-wise relatedness. Compared to multiple state-of-the-art methods, our approaches achieve the best performance across three datasets and six model structures on extensive experiments. For example, with the VGG16 model based on the CIFAR-10 dataset, we can reduce the parameters by 85%, decrease the FLOPs by 61%, and maintain an accuracy of 94.10% (0.14% better than the original). For large-scale ImageNet, we can reduce the parameters by 55% while keeping the accuracy at 76.12% (only drop 0.03%) using the ResNet architecture. In summary, we succeed in reducing the model size and computational resource usage while maintaining the effectiveness of accuracy.

Method: Uses learnable queries and cross-modal attention mechanisms to align abnormal findings in a bootstrapping language-image pretraining framework.

Result: Outperforms state-of-the-art medical vision-language models and 3D report generation methods in accuracy and clinical relevance, reducing missed findings.

Conclusion: Demonstrates the potential of integrating multimodal learning strategies for improving radiology reporting in medical imaging.

Abstract: Medical imaging plays a pivotal role in modern healthcare, with computed tomography pulmonary angiography (CTPA) being a critical tool for diagnosing pulmonary embolism and other thoracic conditions. However, the complexity of interpreting CTPA scans and generating accurate radiology reports remains a significant challenge. This paper introduces Abn-BLIP (Abnormality-aligned Bootstrapping Language-Image Pretraining), an advanced diagnosis model designed to align abnormal findings to generate the accuracy and comprehensiveness of radiology reports. By leveraging learnable queries and cross-modal attention mechanisms, our model demonstrates superior performance in detecting abnormalities, reducing missed findings, and generating structured reports compared to existing methods. Our experiments show that Abn-BLIP outperforms state-of-the-art medical vision-language models and 3D report generation methods in both accuracy and clinical relevance. These results highlight the potential of integrating multimodal learning strategies for improving radiology reporting. The source code is available at https://github.com/zzs95/abn-blip.

Method: Leverages feature-distilled 4D Gaussian Splatting with pre-trained 2D semantic embeddings to create a unified 4D representation where semantics deform with tissue motion, enabling parallelized rasterization for real-time RGB and semantic outputs.

Result: Achieves real-time rendering (61 FPS) with state-of-the-art fidelity on EndoNeRF (39.1 PSNR) and SCARED (27.3 PSNR), and competitive segmentation on EndoVis18 (0.93 DSC for binary, 0.77 DSC for multi-label).

Conclusion: FE-4DGS successfully addresses the challenges of real-time reconstruction and semantic segmentation in deformable surgical environments, outperforming existing methods while maintaining computational efficiency.

Abstract: Minimally invasive surgery (MIS) requires high-fidelity, real-time visual feedback of dynamic and low-texture surgical scenes. To address these requirements, we introduce FeatureEndo-4DGS (FE-4DGS), the first real time pipeline leveraging feature-distilled 4D Gaussian Splatting for simultaneous reconstruction and semantic segmentation of deformable surgical environments. Unlike prior feature-distilled methods restricted to static scenes, and existing 4D approaches that lack semantic integration, FE-4DGS seamlessly leverages pre-trained 2D semantic embeddings to produce a unified 4D representation-where semantics also deform with tissue motion. This unified approach enables the generation of real-time RGB and semantic outputs through a single, parallelized rasterization process. Despite the additional complexity from feature distillation, FE-4DGS sustains real-time rendering (61 FPS) with a compact footprint, achieves state-of-the-art rendering fidelity on EndoNeRF (39.1 PSNR) and SCARED (27.3 PSNR), and delivers competitive EndoVis18 segmentation, matching or exceeding strong 2D baselines for binary segmentation tasks (0.93 DSC) and remaining competitive for multi-label segmentation (0.77 DSC).

Method: Uses pretrained foundation models to separate foreground objects from backgrounds and recombine them in novel ways, increasing data diversity and effective sample count.

Result: Improves ViT accuracy by up to 4.5 p.p. on ImageNet and 7.3 p.p. on downstream tasks. Introduces metrics showing reduced background, foreground, center, and size biases compared to ImageNet training.

Conclusion: ForAug provides an effective tool for analyzing and mitigating biases, enabling development of more robust and reliable computer vision models.

Abstract: Transformers, particularly Vision Transformers (ViTs), have achieved state-of-the-art performance in large-scale image classification. However, they often require large amounts of data and can exhibit biases that limit their robustness and generalizability. This paper introduces ForAug, a novel data augmentation scheme that addresses these challenges and explicitly includes inductive biases, which commonly are part of the neural network architecture, into the training data. ForAug is constructed by using pretrained foundation models to separate and recombine foreground objects with different backgrounds, enabling fine-grained control over image composition during training. It thus increases the data diversity and effective number of training samples. We demonstrate that training on ForNet, the application of ForAug to ImageNet, significantly improves the accuracy of ViTs and other architectures by up to 4.5 percentage points (p.p.) on ImageNet and 7.3 p.p. on downstream tasks. Importantly, ForAug enables novel ways of analyzing model behavior and quantifying biases. Namely, we introduce metrics for background robustness, foreground focus, center bias, and size bias and show that training on ForNet substantially reduces these biases compared to training on ImageNet. In summary, ForAug provides a valuable tool for analyzing and mitigating biases, enabling the development of more robust and reliable computer vision models. Our code and dataset are publicly available at https://github.com/tobna/ForAug.

Method: Based on large language models, visual-question answering chains, external lexical graph, and CLIP-based verification to iteratively extract key concepts from image subsets and organize them into structured concept graphs.

Result: Extensive experiments show the method generates high-quality descriptions in terms of accuracy, completeness, and user satisfaction, and demonstrates scalability across various applications.

Conclusion: ImageSet2Text successfully combines data-driven AI and symbolic representations to reliably summarize large image collections for diverse applications.

Abstract: In the era of large-scale visual data, understanding collections of images is a challenging yet important task. To this end, we introduce ImageSet2Text, a novel method to automatically generate natural language descriptions of image sets. Based on large language models, visual-question answering chains, an external lexical graph, and CLIP-based verification, ImageSet2Text iteratively extracts key concepts from image subsets and organizes them into a structured concept graph. We conduct extensive experiments evaluating the quality of the generated descriptions in terms of accuracy, completeness, and user satisfaction. We also examine the method’s behavior through ablation studies, scalability assessments, and failure analyses. Results demonstrate that ImageSet2Text combines data-driven AI and symbolic representations to reliably summarize large image collections for a wide range of applications.

Method: Defines spherical latent representation for consistent quality, extends MultiDiffusion to spherical latent space, proposes dynamic spherical latent sampling, and introduces distortion-aware weighted averaging.

Result: Outperforms existing approaches in generating 360° static and live wallpapers, providing robust solution for immersive AR/VR applications with reduced distortions.

Conclusion: SphereDiff enables direct use of pretrained diffusion models for high-quality 360° content generation without additional tuning, making it suitable for AR/VR applications.

Abstract: The increasing demand for AR/VR applications has highlighted the need for high-quality content, such as 360° live wallpapers. However, generating high-quality 360° panoramic contents remains a challenging task due to the severe distortions introduced by equirectangular projection (ERP). Existing approaches either fine-tune pretrained diffusion models on limited ERP datasets or adopt tuning-free methods that still rely on ERP latent representations, often resulting in distracting distortions near the poles. In this paper, we introduce SphereDiff, a novel approach for synthesizing 360° static and live wallpaper with state-of-the-art diffusion models without additional tuning. We define a spherical latent representation that ensures consistent quality across all perspectives, including near the poles. Then, we extend MultiDiffusion to spherical latent representation and propose a dynamic spherical latent sampling method to enable direct use of pretrained diffusion models. Moreover, we introduce distortion-aware weighted averaging to further improve the generation quality. Our method outperforms existing approaches in generating 360° static and live wallpaper, making it a robust solution for immersive AR/VR applications. The code is available here. https://github.com/pmh9960/SphereDiff

Method: Created a large dataset of stable brick structures with captions, trained an autoregressive LLM for next-brick prediction, and used physics-aware rollback during inference to prune infeasible designs based on physical laws and assembly constraints.

Result: BrickGPT generates stable, diverse, and aesthetically pleasing brick structures that closely match text prompts. The designs can be assembled both manually by humans and automatically by robotic arms.

Conclusion: The approach successfully bridges text-to-3D generation with physical stability requirements, enabling practical brick assembly from natural language descriptions.

Abstract: We introduce BrickGPT, the first approach for generating physically stable interconnecting brick assembly models from text prompts. To achieve this, we construct a large-scale, physically stable dataset of brick structures, along with their associated captions, and train an autoregressive large language model to predict the next brick to add via next-token prediction. To improve the stability of the resulting designs, we employ an efficient validity check and physics-aware rollback during autoregressive inference, which prunes infeasible token predictions using physics laws and assembly constraints. Our experiments show that BrickGPT produces stable, diverse, and aesthetically pleasing brick structures that align closely with the input text prompts. We also develop a text-based brick texturing method to generate colored and textured designs. We show that our designs can be assembled manually by humans and automatically by robotic arms. We release our new dataset, StableText2Brick, containing over 47,000 brick structures of over 28,000 unique 3D objects accompanied by detailed captions, along with our code and models at the project website: https://avalovelace1.github.io/BrickGPT/.

Method: Decompose module output relative to input stream and add only the orthogonal component, guiding modules to contribute new representational directions.

Result: Improves generalization accuracy and training stability across diverse architectures and datasets, achieving +3.78 pp top-1 accuracy gain for ViT-B on ImageNet-1k.

Conclusion: Orthogonal residual updates foster richer feature learning and more efficient training by promoting novel feature directions.

Abstract: Residual connections are pivotal for deep neural networks, enabling greater depth by mitigating vanishing gradients. However, in standard residual updates, the module’s output is directly added to the input stream. This can lead to updates that predominantly reinforce or modulate the existing stream direction, potentially underutilizing the module’s capacity for learning entirely novel features. In this work, we introduce Orthogonal Residual Update: we decompose the module’s output relative to the input stream and add only the component orthogonal to this stream. This design aims to guide modules to contribute primarily new representational directions, fostering richer feature learning while promoting more efficient training. We demonstrate that our orthogonal update strategy improves generalization accuracy and training stability across diverse architectures (ResNetV2, Vision Transformers) and datasets (CIFARs, TinyImageNet, ImageNet-1k), achieving, for instance, a +3.78 pp top-1 accuracy gain for ViT-B on ImageNet-1k.

Method: Proposed an efficient caption generation framework with VLM automatic generation, quality scoring, and dynamic refinement, plus semantic fusion for coherent full-video captions.

Result: Created LoVR with 467 long videos and 40,804 fine-grained clips with high-quality captions, demonstrating it’s a challenging benchmark that reveals limitations of current approaches.

Conclusion: LoVR presents new challenges for video understanding and retrieval, providing valuable insights for future research in long video-text retrieval.

Abstract: Long videos contain a vast amount of information, making video-text retrieval an essential and challenging task in multimodal learning. However, existing benchmarks suffer from limited video duration, low-quality captions, and coarse annotation granularity, which hinder the evaluation of advanced video-text retrieval methods. To address these limitations, we introduce LoVR, a benchmark specifically designed for long video-text retrieval. LoVR contains 467 long videos and over 40,804 fine-grained clips with high-quality captions. To overcome the issue of poor machine-generated annotations, we propose an efficient caption generation framework that integrates VLM automatic generation, caption quality scoring, and dynamic refinement. This pipeline improves annotation accuracy while maintaining scalability. Furthermore, we introduce a semantic fusion method to generate coherent full-video captions without losing important contextual information. Our benchmark introduces longer videos, more detailed captions, and a larger-scale dataset, presenting new challenges for video understanding and retrieval. Extensive experiments on various advanced embedding models demonstrate that LoVR is a challenging benchmark, revealing the limitations of current approaches and providing valuable insights for future research. We release the code and dataset link at https://github.com/TechNomad-ds/LoVR-benchmark

Method: Used cognitive science methodologies to evaluate VLMs along Perception, Attention, and Memory axes. Conducted vision-text decoupling analysis and fine-tuning experiments on composite visual reasoning tasks.

Result: VLMs show distinct cognitive profiles with ceiling performance on some tasks but significant gaps in spatial understanding and selective attention. Models improve when reasoning over text captions. Fine-tuning smaller VLMs improves core cognitive abilities but doesn’t substantially enhance out-of-distribution benchmark performance.

Conclusion: VLMs have specific cognitive weaknesses in simultaneous perception and reasoning. The study identifies key bottlenecks and provides an effective solution through targeted fine-tuning, while highlighting the need for improved Chain-of-Thought abilities in VLMs.

Abstract: Vision-Language Models (VLMs) have shown remarkable progress in visual understanding in recent years. Yet, they still lag behind human capabilities in specific visual tasks such as counting or relational reasoning. To understand the underlying limitations, we adopt methodologies from cognitive science, analyzing VLM performance along core cognitive axes: Perception, Attention, and Memory. Using a suite of tasks targeting these abilities, we evaluate state-of-the-art VLMs, including GPT-4o. Our analysis reveals distinct cognitive profiles: while advanced models approach ceiling performance on some tasks (e.g. category identification), a significant gap persists, particularly in tasks requiring spatial understanding or selective attention. Investigating the source of these failures and potential methods for improvement, we employ a vision-text decoupling analysis, finding that models struggling with direct visual reasoning show marked improvement when reasoning over their own generated text captions. These experiments reveal a strong need for improved VLM Chain-of-Thought (CoT) abilities, even in models that consistently exceed human performance. Furthermore, we demonstrate the potential of targeted fine-tuning on composite visual reasoning tasks and show that fine-tuning smaller VLMs substantially improves core cognitive abilities. While this improvement does not translate to large enhancements on challenging, out-of-distribution benchmarks, we show broadly that VLM performance on our datasets strongly correlates with performance on these other benchmarks. Our work provides a detailed analysis of VLM cognitive strengths and weaknesses and identifies key bottlenecks in simultaneous perception and reasoning while also providing an effective and simple solution.

Method: Two key components: 1) Data refinement by suppressing sequences with redundant gait cycles or unstable silhouettes using representation redundancy and prediction uncertainty; 2) Disentangled metric learning that forms triplets within each dataset to prevent destructive cross-domain gradients while preserving transferable identity cues.

Result: Experiments on CASIA-B, OU-MVLP, Gait3D, and GREW with GaitBase and DeepGaitV2 backbones show improved cross-domain performance without sacrificing in-domain accuracy.

Conclusion: Data selection and aligning supervision effectively enable scalable mixed-dataset gait learning without modifying network architectures or requiring extra annotations.

Abstract: Generalized gait recognition remains challenging due to significant domain shifts in viewpoints, appearances, and environments. Mixed-dataset training has recently become a practical route to improve cross-domain robustness, but it introduces underexplored issues: 1) inter-dataset supervision conflicts, which distract identity learning, and 2) redundant or noisy samples, which reduce data efficiency and may reinforce dataset-specific patterns. To address these challenges, we introduce a unified paradigm for cross-dataset gait learning that simultaneously improves motion-signal quality and supervision consistency. We first increase the reliability of training data by suppressing sequences dominated by redundant gait cycles or unstable silhouettes, guided by representation redundancy and prediction uncertainty. This refinement concentrates learning on informative gait dynamics when mixing heterogeneous datasets. In parallel, we stabilize supervision by disentangling metric learning across datasets, forming triplets within each source to prevent destructive cross-domain gradients while preserving transferable identity cues. These components act in synergy to stabilize optimization and strengthen generalization without modifying network architectures or requiring extra annotations. Experiments on CASIA-B, OU-MVLP, Gait3D, and GREW with both GaitBase and DeepGaitV2 backbones consistently show improved cross-domain performance without sacrificing in-domain accuracy. These results demonstrate that data selection and aligning supervision effectively enables scalable mixed-dataset gait learning.

Method: Uses autoregressive peeling strategy with vision-language models to construct layer graphs capturing occlusion relationships, employs descriptive captions as editing instructions for finetuned image diffusion model, and implements localized attention control for precise layer removal while preserving surrounding content.

Result: Extensive experiments show LayerPeeler significantly outperforms existing techniques, producing vectorization results with superior path semantics, geometric regularity, and visual fidelity.

Conclusion: LayerPeeler successfully addresses challenges in layer-wise image vectorization through progressive simplification and precise layer removal, enabling generation of vector graphics with complete paths and coherent layer structures.

Abstract: Image vectorization is a powerful technique that converts raster images into vector graphics, enabling enhanced flexibility and interactivity. However, popular image vectorization tools struggle with occluded regions, producing incomplete or fragmented shapes that hinder editability. While recent advancements have explored optimization-based and learning-based layer-wise image vectorization, these methods face limitations in vectorization quality and flexibility. In this paper, we introduce LayerPeeler, a novel layer-wise image vectorization approach that addresses these challenges through a progressive simplification paradigm. The key to LayerPeeler’s success lies in its autoregressive peeling strategy: by identifying and removing the topmost non-occluded layers while recovering underlying content, we generate vector graphics with complete paths and coherent layer structures. Our method leverages vision-language models to construct a layer graph that captures occlusion relationships among elements, enabling precise detection and description for non-occluded layers. These descriptive captions are used as editing instructions for a finetuned image diffusion model to remove the identified layers. To ensure accurate removal, we employ localized attention control that precisely guides the model to target regions while faithfully preserving the surrounding content. To support this, we contribute a large-scale dataset specifically designed for layer peeling tasks. Extensive quantitative and qualitative experiments demonstrate that LayerPeeler significantly outperforms existing techniques, producing vectorization results with superior path semantics, geometric regularity, and visual fidelity.

Method: Propose Wi-CBR with dual-branch self-attention module for temporal features from phase and kinematic features from DFS, plus Saliency Guidance Module with group attention and gating mechanisms for feature fusion.

Result: Extensive experiments on Widar3.0 and XRF55 datasets demonstrate superior performance in both in-domain and cross-domain scenarios.

Conclusion: The proposed Wi-CBR framework effectively handles cross-domain behavior recognition by leveraging complementary phase and DFS signals with sophisticated attention mechanisms.

Abstract: The challenge in WiFi-based cross-domain Behavior Recognition lies in the significant interference of domain-specific signals on gesture variation. However, previous methods alleviate this interference by mapping the phase from multiple domains into a common feature space. If the Doppler Frequency Shift (DFS) signal is used to dynamically supplement the phase features to achieve better generalization, it enables the model to not only explore a wider feature space but also to avoid potential degradation of gesture semantic information. Specifically, we propose a novel Salient-aware Adaptive WiFi Sensing for Cross-domain Behavior Recognition (Wi-CBR), which constructs a dual-branch self-attention module that captures temporal features from phase information reflecting dynamic path length variations while extracting kinematic features from DFS correlated with motion velocity. Moreover, we design a Saliency Guidance Module that employs group attention mechanisms to mine critical activity features and utilizes gating mechanisms to optimize information entropy, facilitating feature fusion and enabling effective interaction between salient and non-salient behavioral characteristics. Extensive experiments on two large-scale public datasets (Widar3.0 and XRF55) demonstrate the superior performance of our method in both in-domain and cross-domain scenarios.

Method: Proposed four sets of clinically relevant features to characterize hypokinesia, bradykinesia, sequence effect, and hesitation-halts from video recordings using computer vision and machine learning classifiers.

Result: The method achieved higher accuracy in MDS-UPDRS score prediction compared to state-of-the-art approaches, identified granular distinctions within motor deficits, and provided interpretable quantification of individual characteristics.

Conclusion: The framework offers a practical solution for objective PD motor assessment applicable in clinical and remote settings, though future work is needed to assess treatment responsiveness and disease progression tracking.

Abstract: Accurately quantifying motor characteristics in Parkinson disease (PD) is crucial for monitoring disease progression and optimizing treatment strategies. The finger-tapping test is a standard motor assessment. Clinicians visually evaluate a patient’s tapping performance and assign an overall severity score based on tapping amplitude, speed, and irregularity. However, this subjective evaluation is prone to inter- and intra-rater variability, and does not offer insights into individual motor characteristics captured during this test. This paper introduces a granular computer vision-based method for quantifying PD motor characteristics from video recordings. Four sets of clinically relevant features are proposed to characterize hypokinesia, bradykinesia, sequence effect, and hesitation-halts. We evaluate our approach on video recordings and clinical evaluations of 74 PD patients from the Personalized Parkinson Project. Principal component analysis with varimax rotation shows that the video-based features corresponded to the four deficits. Additionally, video-based analysis has allowed us to identify further granular distinctions within sequence effect and hesitation-halts deficits. In the following, we have used these features to train machine learning classifiers to estimate the Movement Disorder Society Unified Parkinson Disease Rating Scale (MDS-UPDRS) finger-tapping score. Compared to state-of-the-art approaches, our method achieves a higher accuracy in MDS-UPDRS score prediction, while still providing an interpretable quantification of individual finger-tapping motor characteristics. In summary, the proposed framework provides a practical solution for the objective assessment of PD motor characteristics, that can potentially be applied in both clinical and remote settings. Future work is needed to assess its responsiveness to symptomatic treatment and disease progression.

Method: The paper compares different camera conditioning techniques: token-level raymap encodings, attention-level relative pose encodings, and proposes PRoPE - a relative positional encoding that captures complete camera frustums including both intrinsics and extrinsics.

Result: Experiments show that relative camera conditioning improves performance in novel view synthesis, with additional gains from PRoPE. Benefits persist across different settings (shared/varying intrinsics), tasks (stereo depth estimation, spatial cognition), and model sizes.

Conclusion: PRoPE effectively captures complete camera geometry for multi-view transformers, providing consistent performance improvements across various computer vision tasks and settings.

Abstract: Transformers are increasingly prevalent for multi-view computer vision tasks, where geometric relationships between viewpoints are critical for 3D perception. To leverage these relationships, multi-view transformers must use camera geometry to ground visual tokens in 3D space. In this work, we compare techniques for conditioning transformers on cameras: token-level raymap encodings, attention-level relative pose encodings, and a new relative encoding we propose – Projective Positional Encoding (PRoPE) – that captures complete camera frustums, both intrinsics and extrinsics, as a relative positional encoding. Our experiments begin by showing how relative camera conditioning improves performance in feedforward novel view synthesis, with further gains from PRoPE. This holds across settings: scenes with both shared and varying intrinsics, when combining token- and attention-level conditioning, and for generalization to inputs with out-of-distribution sequence lengths and camera intrinsics. We then verify that these benefits persist for different tasks, stereo depth estimation and discriminative spatial cognition, as well as larger model sizes.

Method: LISA uses layer-wise spectral modulation to stabilize attention and suppress over-amplified activations in deep layers, combined with token-level logit fusion via anchor-based routing for adaptive integration during decoding.

Result: LISA reduces hallucinations by up to 53.6% in CHAIR_I and improves POPE F1 by up to 5.1%, demonstrating strong generalization across models and tasks.

Conclusion: LISA effectively mitigates object hallucinations in MLLMs through layer-wise integration and suppression, offering a plug-and-play solution that can be seamlessly integrated into existing models.

Abstract: Multimodal Large Language Models (MLLMs) excel in vision-language tasks such as image captioning but remain prone to object hallucinations, where they describe objects that do not appear in the image. To mitigate this, we propose LISA, a Layer-wise Integration and Suppression Approach. LISA leverages the layer-wise functional roles in MLLMs: shallow layers provide visual grounding, middle layers encode semantics, and deep layers tend to amplify spurious signals. First, layer-wise spectral modulation stabilizes attention by suppressing over-amplified activations in deeper layers while preserving alignment cues in earlier layers. Second, token-level logits from selected layers are fused via anchor-based routing, with token-wise anchor selection and soft logit fusion enabling adaptive integration during decoding. LISA is fully plug-and-play and can be seamlessly integrated into existing MLLMs, including Qwen2.5-VL. Experiments on multiple benchmarks show that LISA reduces hallucinations by up to 53.6% in $\text{CHAIR}_\text{I}$ and improves POPE F1 by up to 5.1%, demonstrating strong generalization across models and tasks. Our code is available at https://github.com/zhlisa1010-eng/LISA.

Method: Proposes Temporal Sampling Policy Optimization (TSPO) with event-aware temporal agent, reinforcement learning paradigm for joint keyframe selection and language generation, dual-style training data pipeline, and reward mechanisms.

Result: Achieves state-of-the-art performance across multiple long video understanding benchmarks and shows transferable ability across different Video-MLLMs.

Conclusion: TSPO effectively addresses long-form video understanding challenges in MLLMs through optimized temporal sampling policies.

Abstract: Multimodal Large Language Models (MLLMs) have demonstrated significant progress in vision-language tasks, yet they still face challenges when processing long-duration video inputs. The limitation arises from MLLMs’ context limit and training costs, necessitating sparse frame sampling before feeding videos into MLLMs. However, building a trainable sampling method remains challenging due to the unsupervised and non-differentiable nature of sparse frame sampling in Video-MLLMs. To address these problems, we propose Temporal Sampling Policy Optimization (TSPO), advancing MLLMs’ long-form video-language understanding via reinforcement learning. Specifically, we first propose a trainable event-aware temporal agent, which captures event-query correlation for performing probabilistic keyframe selection. Then, we propose the TSPO reinforcement learning paradigm, which models keyframe selection and language generation as a joint decision-making process, enabling end-to-end group relative optimization for the temporal sampling policy. Furthermore, we propose a dual-style long video training data construction pipeline, balancing comprehensive temporal understanding and key segment localization. Finally, we incorporate rule-based answering accuracy and temporal locating reward mechanisms to optimize the temporal sampling policy. Comprehensive experiments show that our TSPO achieves state-of-the-art performance across multiple long video understanding benchmarks, and shows transferable ability across different cutting-edge Video-MLLMs. Our code is available at https://github.com/Hui-design/TSPO

Method: Proposes SCOPE with two modules: Subtle Detail Extractor (SDE) that dynamically enhances edges/textures from shallow features, and Salient Semantic Refiner (SSR) that learns structure-aware refinement from high-level features guided by enhanced shallow features.

Result: Achieves new state-of-the-art performance on four popular fine-grained image classification benchmarks.

Conclusion: SCOPE effectively combines local details with global semantics through progressive cascading of SDE and SSR modules, overcoming limitations of fixed-scale frequency-domain approaches.

Abstract: The crux of resolving fine-grained visual classification (FGVC) lies in capturing discriminative and class-specific cues that correspond to subtle visual characteristics. Recently, frequency decomposition/transform based approaches have attracted considerable interests since its appearing discriminative cue mining ability. However, the frequency-domain methods are based on fixed basis functions, lacking adaptability to image content and unable to dynamically adjust feature extraction according to the discriminative requirements of different images. To address this, we propose a novel method for FGVC, named Subtle-Cue Oriented Perception Engine (SCOPE), which adaptively enhances the representational capability of low-level details and high-level semantics in the spatial domain, breaking through the limitations of fixed scales in the frequency domain and improving the flexibility of multi-scale fusion. The core of SCOPE lies in two modules: the Subtle Detail Extractor (SDE), which dynamically enhances subtle details such as edges and textures from shallow features, and the Salient Semantic Refiner (SSR), which learns semantically coherent and structure-aware refinement features from the high-level features guided by the enhanced shallow features. The SDE and SSR are cascaded stage-by-stage to progressively combine local details with global semantics. Extensive experiments demonstrate that our method achieves new state-of-the-art on four popular fine-grained image classification benchmarks.

Method: VasoMIM introduces two components: anatomy-guided masking strategy that preferentially masks vessel-containing patches, and anatomical consistency loss that enforces vascular semantics consistency between original and reconstructed images.

Result: VasoMIM achieves state-of-the-art performance across three datasets, demonstrating superior vessel segmentation capabilities in X-ray angiograms.

Conclusion: The framework successfully integrates anatomical knowledge into self-supervised pre-training, showing potential to facilitate X-ray angiogram analysis through improved vascular representation learning.

Abstract: Accurate vessel segmentation in X-ray angiograms is crucial for numerous clinical applications. However, the scarcity of annotated data presents a significant challenge, which has driven the adoption of self-supervised learning (SSL) methods such as masked image modeling (MIM) to leverage large-scale unlabeled data for learning transferable representations. Unfortunately, conventional MIM often fails to capture vascular anatomy because of the severe class imbalance between vessel and background pixels, leading to weak vascular representations. To address this, we introduce Vascular anatomy-aware Masked Image Modeling (VasoMIM), a novel MIM framework tailored for X-ray angiograms that explicitly integrates anatomical knowledge into the pre-training process. Specifically, it comprises two complementary components: anatomy-guided masking strategy and anatomical consistency loss. The former preferentially masks vessel-containing patches to focus the model on reconstructing vessel-relevant regions. The latter enforces consistency in vascular semantics between the original and reconstructed images, thereby improving the discriminability of vascular representations. Empirically, VasoMIM achieves state-of-the-art performance across three datasets. These findings highlight its potential to facilitate X-ray angiogram analysis.

Method: Introduced DriftBench, a large-scale benchmark with 16,000 news instances across six diversification categories, evaluating robustness under drift, susceptibility to adversarial evidence, and reasoning consistency across diverse inputs.

Result: Experiments with six state-of-the-art LVLM detectors showed substantial performance drops (average F1 -14.8%), increasingly unstable reasoning traces, and severe failures under adversarial evidence injection.

Conclusion: Current MMD systems have fundamental vulnerabilities to GenAI-driven diversity, highlighting an urgent need for more resilient approaches in the GenAI era.

Abstract: The proliferation of multimodal misinformation poses growing threats to public discourse and societal trust. While Large Vision-Language Models (LVLMs) have enabled recent progress in multimodal misinformation detection (MMD), the rise of generative AI (GenAI) tools introduces a new challenge: GenAI-driven news diversity, characterized by highly varied and complex content. We show that this diversity induces multi-level drift, comprising (1) model-level misperception drift, where stylistic variations disrupt a model’s internal reasoning, and (2) evidence-level drift, where expression diversity degrades the quality or relevance of retrieved external evidence. These drifts significantly degrade the robustness of current LVLM-based MMD systems. To systematically study this problem, we introduce DriftBench, a large-scale benchmark comprising 16,000 news instances across six categories of diversification. We design three evaluation tasks: (1) robustness of truth verification under multi-level drift; (2) susceptibility to adversarial evidence contamination generated by GenAI; and (3) analysis of reasoning consistency across diverse inputs. Experiments with six state-of-the-art LVLM-based detectors show substantial performance drops (average F1 -14.8%) and increasingly unstable reasoning traces, with even more severe failures under adversarial evidence injection. Our findings uncover fundamental vulnerabilities in existing MMD systems and suggest an urgent need for more resilient approaches in the GenAI era.

Method: Created a dataset of 10,210 vehicle images with 12,687 annotated LPs, using fine-grained annotations including occlusion levels, legibility categories, and character labels. Benchmarking used three image recognition networks (ViT, ResNet, YOLO) for classification.

Result: Baseline models achieved F1 scores below 80% across all three architectures, demonstrating the difficulty of the legibility classification task. The dataset enables analysis of SR and LP recognition methods.

Conclusion: The low performance of baseline models highlights the challenge of license plate legibility classification and emphasizes the need for further research in this area.

Abstract: Automatic License Plate Recognition (ALPR) faces a major challenge when dealing with illegible license plates (LPs). While reconstruction methods such as super-resolution (SR) have emerged, the core issue of recognizing these low-quality LPs remains unresolved. To optimize model performance and computational efficiency, image pre-processing should be applied selectively to cases that require enhanced legibility. To support research in this area, we introduce a novel dataset comprising 10,210 images of vehicles with 12,687 annotated LPs for legibility classification (the LPLC dataset). The images span a wide range of vehicle types, lighting conditions, and camera/image quality levels. We adopt a fine-grained annotation strategy that includes vehicle- and LP-level occlusions, four legibility categories (perfect, good, poor, and illegible), and character labels for three categories (excluding illegible LPs). As a benchmark, we propose a classification task using three image recognition networks to determine whether an LP image is good enough, requires super-resolution, or is completely unrecoverable. The overall F1 score, which remained below 80% for all three baseline models (ViT, ResNet, and YOLO), together with the analyses of SR and LP recognition methods, highlights the difficulty of the task and reinforces the need for further research. The proposed dataset is publicly available at https://github.com/lmlwojcik/lplc-dataset.

Method: Two complementary approaches: 1) Multi-scale structural-similarity (4-MS-SSIM) index for inline inspection, 2) Mahalanobis-distance analysis of randomly reduced latent features for feature-space monitoring. Both use lightweight backbone optimized for high-resolution imagery.

Result: Outperforms reference baselines (MOCCA, CPCAE, RAG-PaDiM) on Surface Seal Image dataset under realistic industrial constraints. Cross-domain validation on MVTec-Zipper benchmark shows comparable accuracy to state-of-the-art methods.

Conclusion: Dual-mode framework integrates inline anomaly detection and supervisory monitoring, advancing explainable AI architectures for greater reliability, observability, and lifecycle monitoring in safety-critical manufacturing environments.

Abstract: Automated visual inspection in medical-device manufacturing faces unique challenges, including extremely low defect rates, limited annotated data, hardware restrictions on production lines, and the need for validated, explainable artificial-intelligence systems. This paper presents two attention-guided autoencoder architectures that address these constraints through complementary anomaly-detection strategies. The first employs a multi-scale structural-similarity (4-MS-SSIM) index for inline inspection, enabling interpretable, real-time defect detection on constrained hardware. The second applies a Mahalanobis-distance analysis of randomly reduced latent features for efficient feature-space monitoring and lifecycle verification. Both approaches share a lightweight backbone optimised for high-resolution imagery for typical manufacturing conditions. Evaluations on the Surface Seal Image (SSI) dataset-representing sterile-barrier packaging inspection-demonstrate that the proposed methods outperform reference baselines, including MOCCA, CPCAE, and RAG-PaDiM, under realistic industrial constraints. Cross-domain validation on the MVTec-Zipper benchmark confirms comparable accuracy to state-of-the-art anomaly-detection methods. The dual-mode framework integrates inline anomaly detection and supervisory monitoring, advancing explainable AI architectures toward greater reliability, observability, and lifecycle monitoring in safety-critical manufacturing environments. To facilitate reproducibility, the source code developed for the experiments has been released in the project repository, while the datasets were obtained from publicly available sources.

Method: Reformulates REC as box-wise visual-language verification using COCO-clean generic detector proposals and general-purpose VLM to answer True/False queries for each region.

Result: Surpasses zero-shot GroundingDINO baseline and exceeds reported results for trained GroundingDINO variants on RefCOCO, RefCOCO+, and RefCOCOg datasets.

Conclusion: Workflow design drives strong zero-shot REC performance, reducing cross-box interference and supporting abstention without fine-tuning.

Abstract: Referring Expression Comprehension (REC) is usually addressed with task-trained grounding models. We show that a zero-shot workflow, without any REC-specific training, can achieve competitive or superior performance. Our approach reformulates REC as box-wise visual-language verification: given proposals from a COCO-clean generic detector (YOLO-World), a general-purpose VLM independently answers True/False queries for each region. This simple procedure reduces cross-box interference, supports abstention and multiple matches, and requires no fine-tuning. On RefCOCO, RefCOCO+, and RefCOCOg, our method not only surpasses a zero-shot GroundingDINO baseline but also exceeds reported results for GroundingDINO trained on REC and GroundingDINO+CRG. Controlled studies with identical proposals confirm that verification significantly outperforms selection-based prompting, and results hold with open VLMs. Overall, we show that workflow design, rather than task-specific pretraining, drives strong zero-shot REC performance.

Method: Adapt SAM2 to 3D segmentation by coupling 2D feature extraction with projection/back-projection. Implement architectural modifications: horizontal spatial dependency module, customized configuration for spherical projections, and adapted mechanism for range-view spatial patterns.

Result: Achieves competitive performance on SemanticKITTI benchmark while benefiting from speed, scalability, and deployment simplicity of 2D-centric pipelines.

Conclusion: Range-view segmentation using Visual Foundation Models leads to promising results, highlighting VFMs as viable general-purpose backbones for 3D perception and opening path toward unified foundation-model-driven LiDAR segmentation.

Abstract: Point cloud segmentation is central to autonomous driving and 3D scene understanding. While voxel- and point-based methods dominate recent research due to their compatibility with deep architectures and ability to capture fine-grained geometry, they often incur high computational cost, irregular memory access, and limited real-time efficiency. In contrast, range-view methods, though relatively underexplored - can leverage mature 2D semantic segmentation techniques for fast and accurate predictions. Motivated by the rapid progress in Visual Foundation Models (VFMs) for captioning, zero-shot recognition, and multimodal tasks, we investigate whether SAM2, the current state-of-the-art VFM for segmentation tasks, can serve as a strong backbone for LiDAR point cloud segmentation in the range view. We present , to our knowledge, the first range-view framework that adapts SAM2 to 3D segmentation, coupling efficient 2D feature extraction with standard projection/back-projection to operate on point clouds. To optimize SAM2 for range-view representations, we implement several architectural modifications to the encoder: (1) a novel module that emphasizes horizontal spatial dependencies inherent in LiDAR range images, (2) a customized configuration of tailored to the geometric properties of spherical projections, and (3) an adapted mechanism in the encoder backbone specifically designed to capture the unique spatial patterns and discontinuities present in range-view pseudo-images. Our approach achieves competitive performance on SemanticKITTI while benefiting from the speed, scalability, and deployment simplicity of 2D-centric pipelines. This work highlights the viability of VFMs as general-purpose backbones for 3D perception and opens a path toward unified, foundation-model-driven LiDAR segmentation. Results lets us conclude that range-view segmentation methods using VFMs leads to promising results.

Method: Uses Enhanced Texture from Inter-spike Interval (ETFI) to aggregate sparse information, then encodes ETFI as input to ControlNet for generation, with an ETFI-based feature fusion module to improve quality.

Result: The method effectively leverages generative priors to supplement texture information under diverse low-light conditions for high-speed scene reconstruction.

Conclusion: Diff-SPK provides an effective solution for reconstructing perceptible images from spike camera data in challenging low-light high-speed environments.

Abstract: Spike camera, a type of neuromorphic sensor with high-temporal resolution, shows great promise for high-speed visual tasks. Unlike traditional cameras, spike camera continuously accumulates photons and fires asynchronous spike streams. Due to unique data modality, spike streams require reconstruction methods to become perceptible to the human eye. However, lots of methods struggle to handle spike streams in low-light high-speed scenarios due to severe noise and sparse information. In this work, we propose Diff-SPK, a diffusion-based reconstruction method. Diff-SPK effectively leverages generative priors to supplement texture information under diverse low-light conditions. Specifically, it first employs an Enhanced Texture from Inter-spike Interval (ETFI) to aggregate sparse information from low-light spike streams. Then, the encoded ETFI by a suitable encoder serve as the input of ControlNet for high-speed scenes generation. To improve the quality of results, we introduce an ETFI-based feature fusion module during the generation process.

Method: Integrates object retrieval using captions covering physical, functional, and contextual attributes, places objects based on implicit and explicit layout constraints, and refines positions with collision-aware spatial reasoning powered by large reasoning models.

Result: Significantly outperforms previous methods in human-rated visual fidelity, adherence to constraints, and asset retrieval quality across instructions ranging from simple to complex indoor configurations.

Conclusion: Demonstrates advanced spatial reasoning abilities of modern LRMs and contributes to text-to-scene generation by releasing the codebase to further research in object retrieval and placement.

Abstract: Prompt-driven scene synthesis allows users to generate complete 3D environments from textual descriptions. Current text-to-scene methods often struggle with complex geometries and object transformations, and tend to show weak adherence to complex instructions. We address these limitations by introducing Reason-3D, a text-to-scene model powered by large reasoning models (LRMs). Reason-3D integrates object retrieval using captions covering physical, functional, and contextual attributes. Reason-3D then places the selected objects based on implicit and explicit layout constraints, and refines their positions with collision-aware spatial reasoning. Evaluated on instructions ranging from simple to complex indoor configurations, Reason-3D significantly outperforms previous methods in human-rated visual fidelity, adherence to constraints, and asset retrieval quality. Beyond its contribution to the field of text-to-scene generation, our work showcases the advanced spatial reasoning abilities of modern LRMs. Additionally, we release the codebase to further the research in object retrieval and placement with LRMs.

Method: Uses 3D Gaussian Splatting with redesigned rasterization for depth via differentiable ray-ellipsoid intersection, analytic gradient formulation for surface normal rendering, and learnable attributes for differentiable pruning of Gaussians.

Result: State-of-the-art reconstruction accuracy across all modalities (RGB, depth, normals, semantics) with improved geometric consistency and computational efficiency.

Conclusion: UniGS demonstrates an effective geometry-aware paradigm for multimodal 3D reconstruction, achieving superior performance through differentiable optimization and efficient Gaussian management.

Abstract: In this paper, we propose UniGS, a unified map representation and differentiable framework for high-fidelity multimodal 3D reconstruction based on 3D Gaussian Splatting. Our framework integrates a CUDA-accelerated rasterization pipeline capable of rendering photo-realistic RGB images, geometrically accurate depth maps, consistent surface normals, and semantic logits simultaneously. We redesign the rasterization to render depth via differentiable ray-ellipsoid intersection rather than using Gaussian centers, enabling effective optimization of rotation and scale attribute through analytic depth gradients. Furthermore, we derive the analytic gradient formulation for surface normal rendering, ensuring geometric consistency among reconstructed 3D scenes. To improve computational and storage efficiency, we introduce a learnable attribute that enables differentiable pruning of Gaussians with minimal contribution during training. Quantitative and qualitative experiments demonstrate state-of-the-art reconstruction accuracy across all modalities, validating the efficacy of our geometry-aware paradigm. Source code and multimodal viewer will be available on GitHub.

Method: Proposed MatchAttention with BilinearSoftmax for continuous sliding-window attention sampling, hierarchical MatchDecoder, gated cross-MatchAttention for occlusion handling, and consistency-constrained loss.

Result: MatchStereo-B ranked 1st on Middlebury benchmark (29ms for KITTI), MatchStereo-T processes 4K UHD in 0.1s with 3GB GPU memory, achieving SOTA on KITTI 2012/2015, ETH3D, and Spring flow datasets.

Conclusion: The method enables real-time, high-resolution, high-accuracy cross-view matching with low computational complexity, making practical applications feasible.

Abstract: Cross-view matching is fundamentally achieved through cross-attention mechanisms. However, matching of high-resolution images remains challenging due to the quadratic complexity and lack of explicit matching constraints in the existing cross-attention. This paper proposes an attention mechanism, MatchAttention, that dynamically matches relative positions. The relative position determines the attention sampling center of the key-value pairs given a query. Continuous and differentiable sliding-window attention sampling is achieved by the proposed BilinearSoftmax. The relative positions are iteratively updated through residual connections across layers by embedding them into the feature channels. Since the relative position is exactly the learning target for cross-view matching, an efficient hierarchical cross-view decoder, MatchDecoder, is designed with MatchAttention as its core component. To handle cross-view occlusions, gated cross-MatchAttention and a consistency-constrained loss are proposed. These two components collectively mitigate the impact of occlusions in both forward and backward passes, allowing the model to focus more on learning matching relationships. When applied to stereo matching, MatchStereo-B ranked 1st in average error on the public Middlebury benchmark and requires only 29ms for KITTI-resolution inference. MatchStereo-T can process 4K UHD images in 0.1 seconds using only 3GB of GPU memory. The proposed models also achieve state-of-the-art performance on KITTI 2012, KITTI 2015, ETH3D, and Spring flow datasets. The combination of high accuracy and low computational complexity makes real-time, high-resolution, and high-accuracy cross-view matching possible. Project page: https://github.com/TingmanYan/MatchAttention.

Method: Created VADB database with 10,490 diverse videos annotated by 37 professionals across multiple aesthetic dimensions. Proposed VADB-Net, a dual-modal pre-training framework with two-stage training strategy.

Result: VADB-Net outperforms existing video quality assessment models in scoring tasks and supports downstream video aesthetic assessment tasks.

Conclusion: The introduced VADB dataset and VADB-Net framework address key limitations in video aesthetic assessment, providing standardized resources and improved performance for the field.

Abstract: Video aesthetic assessment, a vital area in multimedia computing, integrates computer vision with human cognition. Its progress is limited by the lack of standardized datasets and robust models, as the temporal dynamics of video and multimodal fusion challenges hinder direct application of image-based methods. This study introduces VADB, the largest video aesthetic database with 10,490 diverse videos annotated by 37 professionals across multiple aesthetic dimensions, including overall and attribute-specific aesthetic scores, rich language comments and objective tags. We propose VADB-Net, a dual-modal pre-training framework with a two-stage training strategy, which outperforms existing video quality assessment models in scoring tasks and supports downstream video aesthetic assessment tasks. The dataset and source code are available at https://github.com/BestiVictory/VADB.

Method: Created WOD-E2E dataset with 12 hours of driving data featuring rare scenarios (<0.03% frequency), including routing info, ego states, and 360-degree camera views. Proposed RFS metric using rater-annotated trajectory preference labels.

Result: Released dataset with 4,021 segments and validation set preference labels, with test set used for 2025 WOD-E2E Challenge.

Conclusion: Aims to advance research in generalizable, robust, and safe end-to-end autonomous driving capable of handling complex real-world situations.

Abstract: Vision-based end-to-end (E2E) driving has garnered significant interest in the research community due to its scalability and synergy with multimodal large language models (MLLMs). However, current E2E driving benchmarks primarily feature nominal scenarios, failing to adequately test the true potential of these systems. Furthermore, existing open-loop evaluation metrics often fall short in capturing the multi-modal nature of driving or effectively evaluating performance in long-tail scenarios. To address these gaps, we introduce the Waymo Open Dataset for End-to-End Driving (WOD-E2E). WOD-E2E contains 4,021 driving segments (approximately 12 hours), specifically curated for challenging long-tail scenarios that that are rare in daily life with an occurring frequency of less than 0.03%. Concretely, each segment in WOD-E2E includes the high-level routing information, ego states, and 360-degree camera views from 8 surrounding cameras. To evaluate the E2E driving performance on these long-tail situations, we propose a novel open-loop evaluation metric: Rater Feedback Score (RFS). Unlike conventional metrics that measure the distance between predicted way points and the logs, RFS measures how closely the predicted trajectory matches rater-annotated trajectory preference labels. We have released rater preference labels for all WOD-E2E validation set segments, while the held out test set labels have been used for the 2025 WOD-E2E Challenge. Through our work, we aim to foster state of the art research into generalizable, robust, and safe end-to-end autonomous driving agents capable of handling complex real-world situations.

Method: Uses Boundary Distance Regression (BDR) for signed-distance boundary detection and Adaptive Temporal Refinement (ATR) to allocate transformer depth continuously to concentrate computation near difficult boundaries.

Result: Achieves 56.5% mAP@0.7 on THUMOS14 with 151G FLOPs (36% fewer than baseline), +2.9% mAP@0.7 improvement with 24% fewer FLOPs and 29% lower latency, with strong gains on short actions.

Conclusion: The method provides consistent improvements across datasets with gains correlating with boundary heterogeneity, offering a theoretically-grounded approach with optimal variance scaling and continuous depth allocation.

Abstract: Temporal action localization requires both precise boundary detection and computational efficiency. Current methods apply uniform computation across all temporal positions, wasting resources on easy boundaries while struggling with ambiguous ones. We address this through two complementary innovations: Boundary Distance Regression (BDR), which replaces classification-based boundary detection with signed-distance regression achieving 3.3–16.7$\times$ lower variance; and Adaptive Temporal Refinement (ATR), which allocates transformer depth continuously ($τ\in[0,1]$) to concentrate computation near difficult boundaries. On THUMOS14, our method achieves 56.5% mAP@0.7 and 58.2% average mAP@[0.3:0.7] with 151G FLOPs, using 36% fewer FLOPs than ActionFormer++ (55.7% mAP@0.7 at 235G). Compared to uniform baselines, we achieve +2.9% mAP@0.7 (+1.8% avg mAP, 5.4% relative) with 24% fewer FLOPs and 29% lower latency, with particularly strong gains on short actions (+4.2%, 8.6% relative). Training requires 1.29$\times$ baseline FLOPs, but this one-time cost is amortized over many inference runs; knowledge distillation further reduces this to 1.1$\times$ while retaining 99.5% accuracy. Our contributions include: (i) a theoretically-grounded distance formulation with information-theoretic analysis showing optimal variance scaling; (ii) a continuous depth allocation mechanism avoiding discrete routing complexity; and (iii) consistent improvements across four datasets with gains correlating with boundary heterogeneity.

Method: Three-component framework: pulmonary perception module for lung abnormalities, knowledge-guided reasoning module for cardiovascular implications, and cardiac representation module for structural biomarkers, fused through agentic reasoning process.

Result: Achieves state-of-the-art performance for CVD screening and mortality prediction on NLST cohort, outperforming single-disease and purely image-based baselines.

Conclusion: Establishes a unified, explainable paradigm for cardiovascular analysis from LDCT that bridges image-based prediction with mechanism-based medical interpretation through human-verifiable reasoning.

Abstract: Low-dose chest computed tomography (LDCT) inherently captures both pulmonary and cardiac structures, offering a unique opportunity for joint assessment of lung and cardiovascular health. However, most existing approaches treat these domains as independent tasks, overlooking their physiological interplay and shared imaging biomarkers. We propose an Explainable Cross-Disease Reasoning Framework that enables interpretable cardiopulmonary risk assessment from a single LDCT scan. The framework introduces an agentic reasoning process that emulates clinical diagnostic thinking-first perceiving pulmonary findings, then reasoning through established medical knowledge, and finally deriving a cardiovascular judgment with explanatory rationale. It integrates three synergistic components: a pulmonary perception module that summarizes lung abnormalities, a knowledge-guided reasoning module that infers their cardiovascular implications, and a cardiac representation module that encodes structural biomarkers. Their outputs are fused to produce a holistic cardiovascular risk prediction that is both accurate and physiologically grounded. Experiments on the NLST cohort demonstrate that the proposed framework achieves state-of-the-art performance for CVD screening and mortality prediction, outperforming single-disease and purely image-based baselines. Beyond quantitative gains, the framework provides human-verifiable reasoning that aligns with cardiological understanding, revealing coherent links between pulmonary abnormalities and cardiac stress mechanisms. Overall, this work establishes a unified and explainable paradigm for cardiovascular analysis from LDCT, bridging the gap between image-based prediction and mechanism-based medical interpretation.

Method: Created MVU-Eval benchmark with 1,824 meticulously curated question-answer pairs spanning 4,959 videos from diverse domains, assessing 8 core competencies including fundamental perception and high-order reasoning tasks aligned with real-world applications.

Result: Extensive evaluation of state-of-the-art open-source and closed-source models revealed significant performance discrepancies and limitations in current MLLMs’ ability to perform understanding across multiple videos.

Conclusion: The benchmark addresses a significant gap in MLLM evaluation and will be made publicly available to foster future research in multi-video understanding capabilities.

Abstract: The advent of Multimodal Large Language Models (MLLMs) has expanded AI capabilities to visual modalities, yet existing evaluation benchmarks remain limited to single-video understanding, overlooking the critical need for multi-video understanding in real-world scenarios (e.g., sports analytics and autonomous driving). To address this significant gap, we introduce MVU-Eval, the first comprehensive benchmark for evaluating Multi-Video Understanding for MLLMs. Specifically, our MVU-Eval mainly assesses eight core competencies through 1,824 meticulously curated question-answer pairs spanning 4,959 videos from diverse domains, addressing both fundamental perception tasks and high-order reasoning tasks. These capabilities are rigorously aligned with real-world applications such as multi-sensor synthesis in autonomous systems and cross-angle sports analytics. Through extensive evaluation of state-of-the-art open-source and closed-source models, we reveal significant performance discrepancies and limitations in current MLLMs’ ability to perform understanding across multiple videos. The benchmark will be made publicly available to foster future research.

Method: Two main modules: 1) High-dimension Semantic Decoupling expands 2D image features along pseudo third dimension to decouple coarse pixel semantics from occlusions and identify focal regions; 2) High-density Occupancy Refinement uses detect-and-refine architecture leveraging contextual geometric and semantic structures to complete missing voxels and correct erroneous ones.

Result: Extensive experiments on SemanticKITTI and SSCBench-KITTI-360 datasets validate the effectiveness of the HD²-SSC framework.

Conclusion: The proposed HD²-SSC framework successfully bridges dimension and density gaps in camera-based 3D semantic scene completion, improving scene understanding for autonomous driving applications.

Abstract: Camera-based 3D semantic scene completion (SSC) plays a crucial role in autonomous driving, enabling voxelized 3D scene understanding for effective scene perception and decision-making. Existing SSC methods have shown efficacy in improving 3D scene representations, but suffer from the inherent input-output dimension gap and annotation-reality density gap, where the 2D planner view from input images with sparse annotated labels leads to inferior prediction of real-world dense occupancy with a 3D stereoscopic view. In light of this, we propose the corresponding High-Dimension High-Density Semantic Scene Completion (HD$^2$-SSC) framework with expanded pixel semantics and refined voxel occupancies. To bridge the dimension gap, a High-dimension Semantic Decoupling module is designed to expand 2D image features along a pseudo third dimension, decoupling coarse pixel semantics from occlusions, and then identify focal regions with fine semantics to enrich image features. To mitigate the density gap, a High-density Occupancy Refinement module is devised with a “detect-and-refine” architecture to leverage contextual geometric and semantic structures for enhanced semantic density with the completion of missing voxels and correction of erroneous ones. Extensive experiments and analyses on the SemanticKITTI and SSCBench-KITTI-360 datasets validate the effectiveness of our HD$^2$-SSC framework.

Method: Extract multilevel semantic features from different vision encoders, project vision features to group related semantics, and fuse visual-textual features using inheritable cross-attention that discards low-correlation feature pairs.

Result: Outperforms all existing methods on eight foundation models and two downstream tasks (visual question answering and image captioning).

Conclusion: The proposed method effectively improves multimodal alignment and fusion by leveraging both semantics and relationships.

Abstract: Vision-language fine-tuning has emerged as an efficient paradigm for constructing multimodal foundation models. While textual context often highlights semantic relationships within an image, existing fine-tuning methods typically overlook this information when aligning vision and language, thus leading to suboptimal performance. Toward solving this problem, we propose a method that can improve multimodal alignment and fusion based on both semantics and relationships.Specifically, we first extract multilevel semantic features from different vision encoder to capture more visual cues of the relationships. Then, we learn to project the vision features to group related semantics, among which are more likely to have relationships. Finally, we fuse the visual features with the textual by using inheritable cross-attention, where we globally remove the redundant visual relationships by discarding visual-language feature pairs with low correlation. We evaluate our proposed method on eight foundation models and two downstream tasks, visual question answering and image captioning, and show that it outperforms all existing methods.

Method: Dual cross-image autoencoders with shared latent space capture common content while isolating modality-specific details. Dual decoders act as feature injectors with architecture varying according to fusion mechanisms (additive/multiplicative).

Result: Extensive experiments validate the method’s effectiveness and generalization ability across diverse fusion tasks.

Conclusion: MAUGIF provides a mechanism-aware approach that enhances both performance and interpretability in general image fusion tasks.

Abstract: Image fusion aims to integrate structural and complementary information from multi-source images. However, existing fusion methods are often either highly task-specific, or general frameworks that apply uniform strategies across diverse tasks, ignoring their distinct fusion mechanisms. To address this issue, we propose a mechanism-aware unsupervised general image fusion (MAUGIF) method based on dual cross-image autoencoders. Initially, we introduce a classification of additive and multiplicative fusion according to the inherent mechanisms of different fusion tasks. Then, dual encoders map source images into a shared latent space, capturing common content while isolating modality-specific details. During the decoding phase, dual decoders act as feature injectors, selectively reintegrating the unique characteristics of each modality into the shared content for reconstruction. The modality-specific features are injected into the source image in the fusion process, generating the fused image that integrates information from both modalities. The architecture of decoders varies according to their fusion mechanisms, enhancing both performance and interpretability. Extensive experiments are conducted on diverse fusion tasks to validate the effectiveness and generalization ability of our method. The code is available at https://anonymous.4open.science/r/MAUGIF.

Method: Uses OCR-LLM pipeline to infer hierarchical regions from unlabeled documents, then combines with teacher detector outputs through inverse-variance fusion to generate refined pseudo-labels with principled probabilistic weighting.

Result: Achieves 88.2±0.3 AP with 5% labels on PubLayNet using lightweight backbone, and 89.7±0.4 AP with LayoutLMv3, surpassing standard semi-supervised learning and matching UDOP which requires 100M+ pages of pretraining.

Conclusion: LLM structural priors are complementary to both lightweight and pretrained architectures, enabling privacy-preserving deployment and targeted semantic disambiguation with minimal performance loss.

Abstract: Document layout understanding remains data-intensive despite advances in semi-supervised learning. We present a framework that enhances semi-supervised detection by fusing visual predictions with structural priors from text-pretrained LLMs via principled probabilistic weighting. Given unlabeled documents, an OCR-LLM pipeline infers hierarchical regions which are combined with teacher detector outputs through inverse-variance fusion to generate refined pseudo-labels.Our method demonstrates consistent gains across model scales. With a lightweight SwiftFormer backbone (26M params), we achieve 88.2$\pm$0.3 AP using only 5% labels on PubLayNet. When applied to document-pretrained LayoutLMv3 (133M params), our fusion framework reaches 89.7$\pm$0.4 AP, surpassing both LayoutLMv3 with standard semi-supervised learning (89.1$\pm$0.4 AP, p=0.02) and matching UDOP~\cite{udop} (89.8 AP) which requires 100M+ pages of multimodal pretraining. This demonstrates that LLM structural priors are complementary to both lightweight and pretrained architectures. Key findings include: (1) learned instance-adaptive gating improves over fixed weights by +0.9 AP with data-dependent PAC bounds correctly predicting convergence; (2) open-source LLMs enable privacy-preserving deployment with minimal loss (Llama-3-70B: 87.1 AP lightweight, 89.4 AP with LayoutLMv3); (3) LLMs provide targeted semantic disambiguation (18.7% of cases, +3.8 AP gain) beyond simple text heuristics.Total system cost includes $12 for GPT-4o-mini API or 17 GPU-hours for local Llama-3-70B per 50K pages, amortized across training runs.

Method: NAMEA decouples gradients from attention and non-attention areas of ensemble models, then merges them using meta-learning. It specifically integrates gradients from non-attention areas since attention areas vary sharply across heterogeneous models.

Result: On ImageNet dataset, NAMEA outperforms state-of-the-art ensemble attacks AdaEA and SMER by an average of 15.0% and 9.6% respectively.

Conclusion: This is the first work to explore ensemble non-attention for boosting cross-architecture transferability, providing new insights into launching effective ensemble attacks.

Abstract: Ensemble attacks integrate the outputs of surrogate models with diverse architectures, which can be combined with various gradient-based attacks to improve adversarial transferability. However, previous work shows unsatisfactory attack performance when transferring across heterogeneous model architectures. The main reason is that the gradient update directions of heterogeneous surrogate models differ widely, making it hard to reduce the gradient variance of ensemble models while making the best of individual model. To tackle this challenge, we design a novel ensemble attack, NAMEA, which for the first time integrates the gradients from the non-attention areas of ensemble models into the iterative gradient optimization process. Our design is inspired by the observation that the attention areas of heterogeneous models vary sharply, thus the non-attention areas of ViTs are likely to be the focus of CNNs and vice versa. Therefore, we merge the gradients respectively from the attention and non-attention areas of ensemble models so as to fuse the transfer information of CNNs and ViTs. Specifically, we pioneer a new way of decoupling the gradients of non-attention areas from those of attention areas, while merging gradients by meta-learning. Empirical evaluations on ImageNet dataset indicate that NAMEA outperforms AdaEA and SMER, the state-of-the-art ensemble attacks by an average of 15.0% and 9.6%, respectively. This work is the first attempt to explore the power of ensemble non-attention in boosting cross-architecture transferability, providing new insights into launching ensemble attacks.

Method: A top-down pipeline using tracking-by-detection strategy to first identify and segment each individual’s pressure signal from raw pressure data, then perform human mesh recovery for each extracted individual signal. Also created a multi-person interaction pressure dataset (MIP).

Result: Achieved 89.2 mm MPJPE and 112.6 mm WA-MPJPE$_{100}$, demonstrating excellent performance in multi-person human mesh recovery using pressure data.

Conclusion: The method showcases the potential of tactile mats for ubiquitous, privacy-preserving multi-person action recognition and enables pressure-based human motion analysis in multi-person scenarios.

Abstract: Multi-person global human mesh recovery (HMR) is crucial for understanding crowd dynamics and interactions. Traditional vision-based HMR methods sometimes face limitations in real-world scenarios due to mutual occlusions, insufficient lighting, and privacy concerns. Human-floor tactile interactions offer an occlusion-free and privacy-friendly alternative for capturing human motion. Existing research indicates that pressure signals acquired from tactile mats can effectively estimate human pose in single-person scenarios. However, when multiple individuals walk randomly on the mat simultaneously, how to distinguish intermingled pressure signals generated by different persons and subsequently acquire individual temporal pressure data remains a pending challenge for extending pressure-based HMR to the multi-person situation. In this paper, we present \textbf{PressTrack-HMR}, a top-down pipeline that recovers multi-person global human meshes solely from pressure signals. This pipeline leverages a tracking-by-detection strategy to first identify and segment each individual’s pressure signal from the raw pressure data, and subsequently performs HMR for each extracted individual signal. Furthermore, we build a multi-person interaction pressure dataset \textbf{MIP}, which facilitates further research into pressure-based human motion analysis in multi-person scenarios. Experimental results demonstrate that our method excels in multi-person HMR using pressure data, with 89.2 $mm$ MPJPE and 112.6 $mm$ WA-MPJPE$_{100}$, and these showcase the potential of tactile mats for ubiquitous, privacy-preserving multi-person action recognition. Our dataset & code are available at https://github.com/Jiayue-Yuan/PressTrack-HMR.

Method: Proposes vMFCoOp framework that inversely estimates von Mises-Fisher distributions on a shared Hyperspherical Manifold, aligning semantic biases via Unified Semantic Anchors with three complementary constraints.

Result: Demonstrates consistent improvements across 14 medical datasets, 12 medical imaging modalities, and 13 anatomical regions, outperforming state-of-the-art methods in accuracy, generalization, and clinical applicability.

Conclusion: The framework enables robust biomedical prompting and superior few-shot classification, with plans to expand to more downstream applications and share resources through an open-source repository.

Abstract: Recent advances in context optimization (CoOp) guided by large language model (LLM)-distilled medical semantic priors offer a scalable alternative to manual prompt engineering and full fine-tuning for adapting biomedical CLIP-based vision-language models (VLMs). However, prompt learning in this context is challenged by semantic misalignment between LLMs and CLIP variants due to divergent training corpora and model architectures; it further lacks scalability across continuously evolving families of foundation models. More critically, pairwise multimodal alignment via conventional Euclidean-space optimization lacks the capacity to model unified representations or apply localized geometric constraints, which tends to amplify modality gaps in complex biomedical imaging and destabilize few-shot adaptation. In this work, we propose vMFCoOp, a framework that inversely estimates von Mises-Fisher (vMF) distributions on a shared Hyperspherical Manifold, aligning semantic biases between arbitrary LLMs and CLIP backbones via Unified Semantic Anchors to achieve robust biomedical prompting and superior few-shot classification. Grounded in three complementary constraints, vMFCoOp demonstrates consistent improvements across 14 medical datasets, 12 medical imaging modalities, and 13 anatomical regions, outperforming state-of-the-art methods in accuracy, generalization, and clinical applicability. This work aims to continuously expand to encompass more downstream applications, and the corresponding resources are intended to be shared through https://github.com/VinyehShaw/UniEqui.

Method: Proposes a PJAR framework with Partial Path Reconstruction (PPR) solver that identifies key item-placeholder pairs to form reduced subproblems, iteratively polishes solutions along optimization path, and incorporates global Large-α constraint to enhance optimality.

Result: Experimental evaluations on benchmark datasets (n=300,500,1000) show the method consistently delivers almost optimal solutions with 0.00% average deviation from ground truth while maintaining high computational efficiency, reducing deviation by half compared to initial heuristic solutions.

Conclusion: The proposed framework achieves high-accuracy, near-optimal solutions for large-scale JRA problems and exhibits strong potential for broader applications including TSP and related optimization problems.

Abstract: The Joint Routing-Assignment (JRA) optimization problem simultaneously determines the assignment of items to placeholders and a Hamiltonian cycle that visits each node pair exactly once, with the objective of minimizing total travel cost. Previous studies introduced an exact mixed-integer programming (MIP) solver, along with datasets and a Gurobi implementation, showing that while the exact approach guarantees optimality, it becomes computationally inefficient for large-scale instances. To overcome this limitation, heuristic methods based on merging algorithms and shaking procedures were proposed, achieving solutions within approximately 1% deviation from the optimum. This work presents a novel and more efficient approach that attains high-accuracy, near-optimal solutions for large-scale JRA problems. The proposed method introduces a Partial Path Reconstructon (PPR) solver that first identifies key item-placeholder pairs to form a reduced subproblem, which is solved efficiently to refine the global solution. Using this PJAR framework, the initial heuristic merging solutions can be further improved, reducing the deviation by half. Moreover, the solution can be iteratively polished with PPR based solver along the optimization path to yield highly accurate tours. Additionally, a global Large-α constraint is incorporated into the JRA model to further enhance solution optimality. Experimental evaluations on benchmark datasets with n = 300, 500, and 1000 demonstrate that the proposed method consistently delivers almost optimal solutions, achieving an average deviation of 0.00% from the ground truth while maintaining high computational efficiency. Beyond the JRA problem, the proposed framework and methodologies exhibit strong potential for broader applications. The Framework can be applied to TSP and related optimization problems.

Method: The approach uses the Variable Neighborhood Search (VNS) metaheuristic to solve the Capacitated Green Vehicle Routing Problem.

Result: The method achieved the best results on the full CEC-12 competition dataset and outperformed a more recent algorithm published after the competition.

Conclusion: The VNS-based approach proved highly effective for solving the CGVRP, demonstrating superior performance in the competition and against subsequent algorithms.

Abstract: The Electric Vehicle Routing Problem (EVRP) extends the classical Vehicle Routing Problem (VRP) to reflect the growing use of electric and hybrid vehicles in logistics. Due to the variety of constraints considered in the literature, comparing approaches across different problem variants remains challenging. A minimalistic variant of the EVRP, known as the Capacitated Green Vehicle Routing Problem (CGVRP), was the focus of the CEC-12 competition held during the 2020 IEEE World Congress on Computational Intelligence. This paper presents the competition-winning approach, based on the Variable Neighborhood Search (VNS) metaheuristic. The method achieves the best results on the full competition dataset and also outperforms a more recent algorithm published afterward.

Method: International workshop format with researchers from HCI, Interaction Design, AI, XAI, and digital arts communities.

Result: Community gathering and knowledge sharing about XAI applications in artistic contexts.

Conclusion: Successful interdisciplinary workshop establishing connections between XAI and arts communities.

Abstract: This third international workshop on explainable AI for the Arts (XAIxArts) brought together a community of researchers in HCI, Interaction Design, AI, explainable AI (XAI), and digital arts to explore the role of XAI for the Arts. Workshop held at the 17th ACM Conference on Creativity and Cognition (C&C 2025), online.

Method: The framework has three components: Tool Generation for automatic creation of diverse tools, Tool Simulation to emulate realistic tool behaviors, and Tool Audit to ensure correctness and consistency.

Result: SynthTools can produce toolsets spanning twice as many domains and tools per domain as prior work, with tool simulation achieving 94% accuracy and tool audit achieving 99% accuracy.

Conclusion: SynthTools enables scalable, diverse, and reliable tool ecosystems for large-scale training and stable evaluation of tool-use agents.

Abstract: AI agents increasingly rely on external tools to solve complex, long-horizon tasks. Advancing such agents requires reproducible evaluation and large-scale training in controllable, diverse, and realistic tool-use environments. However, real-world APIs are limited in availability, domain coverage, and stability, often requiring access keys and imposing rate limits, which render them impractical for stable evaluation or scalable training. To address these challenges, we introduce SynthTools, a flexible and scalable framework for generating synthetic tool ecosystems. Our framework consists of three core components: Tool Generation for automatic and scalable creation of diverse tools, Tool Simulation to emulate realistic tool behaviors, and Tool Audit to ensure correctness and consistency of tool simulation. To illustrate its scalability, we show that SynthTools can readily produce toolsets that span twice as many domains and twice as many tools per domain as prior work. Furthermore, the tool simulation and tool audit components demonstrate strong reliability, achieving $94%$ and $99%$ accuracy respectively. Finally, we construct downstream tasks from the generated tools that even state-of-the-art models struggle to complete. By enabling scalable, diverse, and reliable tool ecosystems, SynthTools provides a practical path toward large-scale training and stable evaluation of tool-use agents. Our code is available at https://github.com/namkoong-lab/SynthTools.

Method: Proposes Rebellion - a robust reasoning training method that trains ARMs to be robust to worst-case representation drift, protecting them from advanced audio jailbreak attacks.

Result: Rebellion successfully protects Qwen2-Audio against advanced audio jailbreaks without compromising benign task performance, and significantly improves accuracy-safety trade-off over standard reasoning training.

Conclusion: Rebellion provides an effective defense mechanism for ARMs against sophisticated jailbreak attacks by addressing the fundamental issue of representation drift between vanilla and advanced attacks.

Abstract: Instilling reasoning capabilities in large models (LMs) using reasoning training (RT) significantly improves LMs’ performances. Thus Audio Reasoning Models (ARMs), i.e., audio LMs that can reason, are becoming increasingly popular. However, no work has studied the safety of ARMs against jailbreak attacks that aim to elicit harmful responses from target models. To this end, first, we show that standard RT with appropriate safety reasoning data can protect ARMs from vanilla audio jailbreaks, but cannot protect them against our proposed simple yet effective jailbreaks. We show that this is because of the significant representation drift between vanilla and advanced jailbreaks which forces the target ARMs to emit harmful responses. Based on this observation, we propose Rebellion, a robust RT that trains ARMs to be robust to the worst-case representation drift. All our results are on Qwen2-Audio; they demonstrate that Rebellion: 1) can protect against advanced audio jailbreaks without compromising performance on benign tasks, and 2) significantly improves accuracy-safety trade-off over standard RT method.

Method: Creating a platform for interdisciplinary researchers to explore techniques for reconciling reasoning between transformer-based LMs and logic-based representations, including analyzing LM reasoning abilities, injecting KR-style reasoning, and formalizing LM reasoning processes.

Result: The workshop aims to uncover how language models can effectively integrate and leverage knowledge and reasoning, though specific results are not provided in the abstract.

Conclusion: This exploration seeks to improve language model applications in areas requiring precision and reliability by better understanding and enhancing their reasoning capabilities through integration with knowledge representation methods.

Abstract: Reasoning is an essential component of human intelligence in that it plays a fundamental role in our ability to think critically, support responsible decisions, and solve challenging problems. Traditionally, AI has addressed reasoning in the context of logic-based representations of knowledge. However, the recent leap forward in natural language processing, with the emergence of language models based on transformers, is hinting at the possibility that these models exhibit reasoning abilities, particularly as they grow in size and are trained on more and more data. Still, despite ongoing discussions about what reasoning is in language models, it is still not easy to articulate to what extent these models are actually capable of reasoning. The goal of this workshop is to create a platform for researchers from different disciplines and/or AI perspectives to explore approaches and techniques with the aim to reconcile reasoning between language models using transformers and logic-based representations. The specific objectives include analysing the reasoning abilities of language models measured alongside KR methods, injecting KR-style reasoning abilities into language models (including by neuro-symbolic means), and formalising the kind of reasoning language models carry out. This exploration aims to uncover how language models can effectively integrate and leverage knowledge and reasoning with it, thus improving their application and utility in areas where precision and reliability are key requirements.

Method: Theoretical proof and logical analysis of the properties of cogent and weakly admissible extensions in argumentation frameworks.

Result: Proved that cogent extensions are weakly admissible, but weakly admissible extensions are not necessarily cogent.

Conclusion: There is a one-way relationship between cogent and weakly admissible semantics where cogent implies weakly admissible but not conversely.

Abstract: In this research note, we show the relationship between two non-admissible argumentation framework semantics: cogent and weakly admissible semantics. We prove that, while cogent extensions are weakly admissible, the converse is not true.

Method: Conducted experiments across 60 AxA configurations, 3 domains, and 2000+ conversations using major LLM providers. Analyzed prompt impacts, conversation dynamics, and introduced protocol-level mitigation using structured responses.

Result: Echoing occurs across three major LLM providers with rates from 5% to 70% depending on model and domain. Echoing persists even in advanced reasoning models (32.8% rate) and increases with longer interactions (7+ turns). Structured response mitigation reduces echoing to 9%.

Conclusion: Echoing is a persistent failure mode in autonomous agent-agent conversations that cannot be addressed by improved reasoning alone, requiring protocol-level interventions like structured responses to maintain role integrity.

Abstract: As large language model (LLM) based agents interact autonomously with one another, a new class of failures emerges that cannot be predicted from single agent performance: behavioral drifts in agent-agent conversations (AxA). Unlike human-agent interactions, where humans ground and steer conversations, AxA lacks such stabilizing signals, making these failures unique. We investigate one such failure, echoing, where agents abandon their assigned roles and instead mirror their conversational partners, undermining their intended objectives. Through experiments across $60$ AxA configurations, $3$ domains, and $2000+$ conversations, we demonstrate that echoing occurs across three major LLM providers, with echoing rates from $5%$ to $70%$ depending on the model and domain. Moreover, we find that echoing is persistent even in advanced reasoning models with substantial rates ($32.8%$) that are not reduced by increased reasoning efforts. We analyze prompt impacts, conversation dynamics, showing that echoing arises as interaction grows longer ($7+$ turns in experiments) and is not merely an artifact of sub-optimal prompting. Finally, we introduce a protocol-level mitigation in which targeted use of structured responses reduces echoing to $9%$.

Method: Formalizes bias assumptions as ProbLog programs, uses neurosymbolic extensions to integrate these into neural network training, and provides templates for expressing different bias types.

Result: Successfully mitigates algorithmic bias in synthetic tabular datasets with known biases and real-world tabular and image data, outperforming baseline methods.

Conclusion: ProbLog4Fairness provides a flexible, principled approach to bias mitigation that can model various bias assumptions, unlike methods that enforce fixed fairness notions.

Abstract: Operationalizing definitions of fairness is difficult in practice, as multiple definitions can be incompatible while each being arguably desirable. Instead, it may be easier to directly describe algorithmic bias through ad-hoc assumptions specific to a particular real-world task, e.g., based on background information on systemic biases in its context. Such assumptions can, in turn, be used to mitigate this bias during training. Yet, a framework for incorporating such assumptions that is simultaneously principled, flexible, and interpretable is currently lacking. Our approach is to formalize bias assumptions as programs in ProbLog, a probabilistic logic programming language that allows for the description of probabilistic causal relationships through logic. Neurosymbolic extensions of ProbLog then allow for easy integration of these assumptions in a neural network’s training process. We propose a set of templates to express different types of bias and show the versatility of our approach on synthetic tabular datasets with known biases. Using estimates of the bias distortions present, we also succeed in mitigating algorithmic bias in real-world tabular and image data. We conclude that ProbLog4Fairness outperforms baselines due to its ability to flexibly model the relevant bias assumptions, where other methods typically uphold a fixed bias type or notion of fairness.

Method: Used transcripts from 30 one-to-one math sessions, compared three LLMs (GPT, Claude, Gemini) under three conditions: unverified annotation, self-verification, and cross-verification. Benchmarked against blinded human adjudication using Cohen’s kappa.

Result: Orchestration yields 58% improvement in kappa. Self-verification nearly doubles agreement, with largest gains for challenging tutor moves. Cross-verification achieves 37% average improvement with pair-dependent effects.

Conclusion: Verification serves as a principled design lever for reliable, scalable LLM-assisted annotation in Learning Analytics, with proposed notation system for standardized reporting.

Abstract: Large Language Models (LLMs) are increasingly used to annotate learning interactions, yet concerns about reliability limit their utility. We test whether verification-oriented orchestration-prompting models to check their own labels (self-verification) or audit one another (cross-verification)-improves qualitative coding of tutoring discourse. Using transcripts from 30 one-to-one math sessions, we compare three production LLMs (GPT, Claude, Gemini) under three conditions: unverified annotation, self-verification, and cross-verification across all orchestration configurations. Outputs are benchmarked against a blinded, disagreement-focused human adjudication using Cohen’s kappa. Overall, orchestration yields a 58 percent improvement in kappa. Self-verification nearly doubles agreement relative to unverified baselines, with the largest gains for challenging tutor moves. Cross-verification achieves a 37 percent improvement on average, with pair- and construct-dependent effects: some verifier-annotator pairs exceed self-verification, while others reduce alignment, reflecting differences in verifier strictness. We contribute: (1) a flexible orchestration framework instantiating control, self-, and cross-verification; (2) an empirical comparison across frontier LLMs on authentic tutoring data with blinded human “gold” labels; and (3) a concise notation, verifier(annotator) (e.g., Gemini(GPT) or Claude(Claude)), to standardize reporting and make directional effects explicit for replication. Results position verification as a principled design lever for reliable, scalable LLM-assisted annotation in Learning Analytics.

Method: Advocates for using open small AI models that can be deployed locally, allowing artists to modify code, integrate new interfaces, and customize models through retraining or fine-tuning with custom datasets.

Result: Small AI models provide artists with greater autonomy, control, sustainability, and technological self-determination for interactive art practices.

Conclusion: Open small AI models empower artists with creative independence, support long-term preservation of AI artworks, and reduce reliance on corporate AI systems ill-suited for interactive art demands.

Abstract: This position paper argues for the importance of open small AI models in creative independence for interactive art practices. Deployable locally, these models offer artists vital control over infrastructure and code, unlike dominant large, closed-source corporate systems. Such centralized platforms function as opaque black boxes, imposing severe limitations on interactive artworks, including restrictive content filters, preservation issues, and technical challenges such as increased latency and limited interfaces. In contrast, small AI models empower creators with more autonomy, control, and sustainability for these artistic processes. They enable the ability to use a model as long as they want, create their own custom model, either by making code changes to integrate new interfaces, or via new datasets by re-training or fine-tuning the model. This fosters technological self-determination, offering greater ownership and reducing reliance on corporate AI ill-suited for interactive art’s demands. Critically, this approach empowers the artist and supports long-term preservation and exhibition of artworks with AI components. This paper explores the practical applications and implications of using open small AI models in interactive art, contrasting them with closed-source alternatives.

Method: SlideBot uses a modular multi-agent framework integrating LLMs with retrieval, structured planning, and code generation. It incorporates evidence-based instructional design principles (CLT and CTML) and employs specialized agents for information retrieval, content summarization, figure generation, and slide formatting using LaTeX.

Result: Evaluations from domain experts and students in AI and biomedical education show SlideBot consistently enhances conceptual accuracy, clarity, and instructional value compared to existing solutions.

Conclusion: SlideBot demonstrates potential to streamline slide preparation while ensuring accuracy, relevance, and adaptability in higher education through its three-pillar approach focusing on informativeness, reliability, and practicality.

Abstract: Large Language Models (LLMs) have shown immense potential in education, automating tasks like quiz generation and content summarization. However, generating effective presentation slides introduces unique challenges due to the complexity of multimodal content creation and the need for precise, domain-specific information. Existing LLM-based solutions often fail to produce reliable and informative outputs, limiting their educational value. To address these limitations, we introduce SlideBot - a modular, multi-agent slide generation framework that integrates LLMs with retrieval, structured planning, and code generation. SlideBot is organized around three pillars: informativeness, ensuring deep and contextually grounded content; reliability, achieved by incorporating external sources through retrieval; and practicality, which enables customization and iterative feedback through instructor collaboration. It incorporates evidence-based instructional design principles from Cognitive Load Theory (CLT) and the Cognitive Theory of Multimedia Learning (CTML), using structured planning to manage intrinsic load and consistent visual macros to reduce extraneous load and enhance dual-channel learning. Within the system, specialized agents collaboratively retrieve information, summarize content, generate figures, and format slides using LaTeX, aligning outputs with instructor preferences through interactive refinement. Evaluations from domain experts and students in AI and biomedical education show that SlideBot consistently enhances conceptual accuracy, clarity, and instructional value. These findings demonstrate SlideBot’s potential to streamline slide preparation while ensuring accuracy, relevance, and adaptability in higher education.

Method: Proposed four embedding strategies using in-place fine-tuning to embed watermarks while minimizing impact on model accuracy.

Result: Achieves high watermark embedding rates, low accuracy degradation, and strong resistance to post-deployment fine-tuning across diverse datasets.

Conclusion: The framework successfully enables robust watermarking for GBDT models with minimal performance impact.

Abstract: Gradient Boosting Decision Trees (GBDTs) are widely used in industry and academia for their high accuracy and efficiency, particularly on structured data. However, watermarking GBDT models remains underexplored compared to neural networks. In this work, we present the first robust watermarking framework tailored to GBDT models, utilizing in-place fine-tuning to embed imperceptible and resilient watermarks. We propose four embedding strategies, each designed to minimize impact on model accuracy while ensuring watermark robustness. Through experiments across diverse datasets, we demonstrate that our methods achieve high watermark embedding rates, low accuracy degradation, and strong resistance to post-deployment fine-tuning.

Method: Equips Large Language Models with Lean tools via Model Context Protocol, combining LLM knowledge and reasoning with formal verification tools to ensure syntactic rigor while maintaining creative reasoning capabilities.

Result: Competitive with state-of-the-art provers on public math benchmarks and largely outperforms them on new benchmarks in abstract algebra and quantum theory; successfully assisted an expert mathematician in formalizing a complex cryptography theorem.

Conclusion: The tool-based agentic theorem prover approach offers a generalizable methodology for formal verification across diverse scientific domains, overcoming the generalization limitations of specialized systems.

Abstract: We present Ax-Prover, a multi-agent system for automated theorem proving in Lean that can solve problems across diverse scientific domains and operate either autonomously or collaboratively with human experts. To achieve this, Ax-Prover approaches scientific problem solving through formal proof generation, a process that demands both creative reasoning and strict syntactic rigor. Ax-Prover meets this challenge by equipping Large Language Models (LLMs), which provide knowledge and reasoning, with Lean tools via the Model Context Protocol (MCP), which ensure formal correctness. To evaluate its performance as an autonomous prover, we benchmark our approach against frontier LLMs and specialized prover models on two public math benchmarks and on two Lean benchmarks we introduce in the fields of abstract algebra and quantum theory. On public datasets, Ax-Prover is competitive with state-of-the-art provers, while it largely outperforms them on the new benchmarks. This shows that, unlike specialized systems that struggle to generalize, our tool-based agentic theorem prover approach offers a generalizable methodology for formal verification across diverse scientific domains. Furthermore, we demonstrate Ax-Prover’s assistant capabilities in a practical use case, showing how it enabled an expert mathematician to formalize the proof of a complex cryptography theorem.

Method: Integration of thermochromic dyes with flexible heating units on clothing to create dynamic adversarial patterns that activate upon heating, making the clothing appear normal when inactive but adversarial when heated.

Result: The system achieves rapid texture activation within 50 seconds and maintains over 80% adversarial success rate across diverse real-world surveillance environments in both visible and infrared modalities.

Conclusion: This work demonstrates a new approach for physically grounded, user-controllable anti-AI systems that provide effective privacy protection against ubiquitous AI surveillance through proactive adversarial techniques.

Abstract: Adversarial patches have emerged as a popular privacy-preserving approach for resisting AI-driven surveillance systems. However, their conspicuous appearance makes them difficult to deploy in real-world scenarios. In this paper, we propose a thermally activated adversarial wearable designed to ensure adaptability and effectiveness in complex real-world environments. The system integrates thermochromic dyes with flexible heating units to induce visually dynamic adversarial patterns on clothing surfaces. In its default state, the clothing appears as an ordinary black T-shirt. Upon heating via an embedded thermal unit, hidden adversarial patterns on the fabric are activated, allowing the wearer to effectively evade detection across both visible and infrared modalities. Physical experiments demonstrate that the adversarial wearable achieves rapid texture activation within 50 seconds and maintains an adversarial success rate above 80% across diverse real-world surveillance environments. This work demonstrates a new pathway toward physically grounded, user-controllable anti-AI systems, highlighting the growing importance of proactive adversarial techniques for privacy protection in the age of ubiquitous AI surveillance.

Method: Comprehensive exploration of QAI including quantum-enhanced learning pipelines, quantum uncertainty quantification, quantum explainability frameworks, quantum resource management model, and application scheduling driven by timeliness constraints.

Result: Identified key application areas (aerospace, defense, cybersecurity, smart grids, disaster management) where QAI enhances fault tolerance, real-time intelligence, and adaptability; proposed deployment positioning and resource management model.

Conclusion: Future research should focus on achieving interpretable, scalable, and hardware-feasible QAI models for MC deployment, addressing challenges like trainability limits, data bottlenecks, verification, and adversarial QAI.

Abstract: Mission critical (MC) applications such as defense operations, energy management, cybersecurity, and aerospace control require reliable, deterministic, and low-latency decision making under uncertainty. Although the classical Machine Learning (ML) approaches are effective, they often struggle to meet the stringent constraints of robustness, timing, explainability, and safety in the MC domains. Quantum Artificial Intelligence (QAI), the fusion of machine learning and quantum computing (QC), can provide transformative solutions to the challenges faced by classical ML models. In this paper, we provide a comprehensive exploration of QAI for MC systems. We begin with a conceptual background to quantum computing, MC systems, and quantum machine learning (QAI). We then examine the core mechanisms and algorithmic principles of QAI in MC systems, including quantum-enhanced learning pipelines, quantum uncertainty quantification, and quantum explainability frameworks. Subsequently, we discuss key application areas like aerospace, defense, cybersecurity, smart grids, and disaster management, focusing on the role of QA in enhancing fault tolerance, real-time intelligence, and adaptability. We provide an exploration of the positioning of QAI for MC systems in the industry in terms of deployment. We also propose a model for management of quantum resources and scheduling of applications driven by timeliness constraints. We discuss multiple challenges, including trainability limits, data access, and loading bottlenecks, verification of quantum components, and adversarial QAI. Finally, we outline future research directions toward achieving interpretable, scalable, and hardware-feasible QAI models for MC application deployment.

Method: Collected 233 simulated EMS scenarios from 62 participants (including 46 professionals) using an open-source, low-cost egocentric data collection system with multimodal annotations including keysteps, audio transcripts, action quality metrics, and segmentation masks.

Result: Created EgoEMS - the first end-to-end, high-fidelity, multimodal, multiperson EMS dataset capturing realistic emergency dynamics with comprehensive annotations aligned with national standards.

Conclusion: EgoEMS enables development of AI support tools for EMS through benchmarks for real-time multimodal keystep recognition and action quality estimation, potentially improving patient outcomes.

Abstract: Emergency Medical Services (EMS) are critical to patient survival in emergencies, but first responders often face intense cognitive demands in high-stakes situations. AI cognitive assistants, acting as virtual partners, have the potential to ease this burden by supporting real-time data collection and decision making. In pursuit of this vision, we introduce EgoEMS, the first end-to-end, high-fidelity, multimodal, multiperson dataset capturing over 20 hours of realistic, procedural EMS activities from an egocentric view in 233 simulated emergency scenarios performed by 62 participants, including 46 EMS professionals. Developed in collaboration with EMS experts and aligned with national standards, EgoEMS is captured using an open-source, low-cost, and replicable data collection system and is annotated with keysteps, timestamped audio transcripts with speaker diarization, action quality metrics, and bounding boxes with segmentation masks. Emphasizing realism, the dataset includes responder-patient interactions reflecting real-world emergency dynamics. We also present a suite of benchmarks for real-time multimodal keystep recognition and action quality estimation, essential for developing AI support tools for EMS. We hope EgoEMS inspires the research community to push the boundaries of intelligent EMS systems and ultimately contribute to improved patient outcomes.

Method: 1) Fine-tunes pretrained protein language models using masked language modeling on homologous sequences to activate evolutionary plausibility; 2) Introduces test-time inference based on Monte Carlo tree search to evolve proteins with evolutionary guidance from the fine-tuned model.

Result: Extensive benchmark experiments show AlphaDE remarkably outperforms previous state-of-the-art methods even with few-shot fine-tuning. Case study demonstrates it can condense protein sequence space through computational evolution.

Conclusion: AlphaDE successfully bridges the gap by harnessing protein language models to guide protein evolution, providing an effective framework that leverages evolutionary patterns encoded in language models for protein sequence design.

Abstract: Protein evolution through amino acid sequence mutations is a cornerstone of life sciences. While current in-silicon directed evolution algorithms focus on designing search strategies, they overlook how to utilize the transformative protein language models, which encode rich evolutionary patterns, to guide search. To bridge this gap, we propose AlphaDE, a novel framework to evolve protein sequences by harnessing the innovative paradigms of large language models. First, AlphaDE fine-tunes pretrained protein language models using masked language modeling on homologous protein sequences to activate the evolutionary plausibility for the interested protein class. Second, AlphaDE introduces test-time inference based on Monte Carlo tree search, which effectively evolves proteins with evolutionary guidance from the fine-tuned protein language model. Extensive benchmark experiments show that AlphaDE remarkably outperforms previous state-of-the-art methods even with few-shot fine-tuning. An interesting case study further shows that AlphaDE supports condensing the protein sequence space through computational evolution.

Method: Extended MLE-Bench benchmark with code-sabotage tasks (backdoors, generalization failures), studied agent sandbagging capabilities, used LM monitors to detect suspicious behavior, and measured agent ability to evade detection.

Result: Frontier agents successfully performed sabotage tasks and could calibrate performance to specified target levels. Monitors detected code-sabotage attempts but struggled with sandbagging detection. Multiple monitor aggregation improved detection but may not be sufficiently reliable for high-stakes domains.

Conclusion: Current monitoring approaches can detect overt sabotage but are less effective against subtle sandbagging, highlighting the need for more robust oversight mechanisms before deploying autonomous AI in safety-critical ML engineering.

Abstract: AI systems are increasingly able to autonomously conduct realistic software engineering tasks, and may soon be deployed to automate machine learning (ML) R&D itself. Frontier AI systems may be deployed in safety-critical settings, including to help ensure the safety of future systems. Unfortunately, frontier and future systems may not be sufficiently trustworthy, and there is evidence that these systems may even be misaligned with their developers or users. Therefore, we investigate the capabilities of AI agents to act against the interests of their users when conducting ML engineering, by sabotaging ML models, sandbagging their performance, and subverting oversight mechanisms. First, we extend MLE-Bench, a benchmark for realistic ML tasks, with code-sabotage tasks such as implanting backdoors and purposefully causing generalisation failures. Frontier agents make meaningful progress on our sabotage tasks. In addition, we study agent capabilities to sandbag on MLE-Bench. Agents can calibrate their performance to specified target levels below their actual capability. To mitigate sabotage, we use LM monitors to detect suspicious agent behaviour, and we measure model capability to sabotage and sandbag without being detected by these monitors. Overall, monitors are capable at detecting code-sabotage attempts but our results suggest that detecting sandbagging is more difficult. Additionally, aggregating multiple monitor predictions works well, but monitoring may not be sufficiently reliable to mitigate sabotage in high-stakes domains. Our benchmark is implemented in the UK AISI’s Inspect framework and we make our code publicly available at https://github.com/samm393/mlebench-subversion

Method: Construct related problem pairs with intermediate problem-design CoT from reasoning models, use solver feedback as reward signal to calibrate difficulty, and enable generator-solver co-evolution.

Result: Achieves 2.5% average improvement on 10 mathematical and general reasoning benchmarks, generalizes to language and vision-language models, and co-evolution yields additional 0.7% gain.

Conclusion: The reasoning-based problem generator with adaptive difficulty and co-evolution framework effectively creates high-quality training data for reasoning models.

Abstract: Data synthesis for training large reasoning models offers a scalable alternative to limited, human-curated datasets, enabling the creation of high-quality data. However, existing approaches face several challenges: (i) indiscriminate generation that ignores the solver’s ability and yields low-value problems, or reliance on complex data pipelines to balance problem difficulty; and (ii) a lack of reasoning in problem generation, leading to shallow problem variants. In this paper, we develop a problem generator that reasons explicitly to plan problem directions before synthesis and adapts difficulty to the solver’s ability. Specifically, we construct related problem pairs and augment them with intermediate problem-design CoT produced by a reasoning model. These data bootstrap problem-design strategies from the generator. Then, we treat the solver’s feedback on synthetic problems as a reward signal, enabling the generator to calibrate difficulty and produce complementary problems near the edge of the solver’s competence. Extensive experiments on 10 mathematical and general reasoning benchmarks show that our method achieves an average improvement of 2.5% and generalizes to both language and vision-language models. Moreover, a solver trained on the synthesized data provides improved rewards for continued generator training, enabling co-evolution and yielding a further 0.7% performance gain. Our code will be made publicly available here.

Method: Organized 400k training documents with 10k for testing, extracted multimodal features (text, visuals, layout), generated 360k QA pairs, developed domain-specific multimodal LLMs, incorporated historical context and page references, and created importance-based page classification.

Result: Preliminary results show improved performance in document information extraction and question-answering tasks, with the dataset and models made publicly available.

Conclusion: The developed multimodal AI system effectively addresses the challenges of analyzing complex healthcare-related legal documents and provides a foundation for better understanding systemic failures in the opioid crisis.

Abstract: The opioid crisis represents a significant moment in public health that reveals systemic shortcomings across regulatory systems, healthcare practices, corporate governance, and public policy. Analyzing how these interconnected systems simultaneously failed to protect public health requires innovative analytic approaches for exploring the vast amounts of data and documents disclosed in the UCSF-JHU Opioid Industry Documents Archive (OIDA). The complexity, multimodal nature, and specialized characteristics of these healthcare-related legal and corporate documents necessitate more advanced methods and models tailored to specific data types and detailed annotations, ensuring the precision and professionalism in the analysis. In this paper, we tackle this challenge by organizing the original dataset according to document attributes and constructing a benchmark with 400k training documents and 10k for testing. From each document, we extract rich multimodal information-including textual content, visual elements, and layout structures-to capture a comprehensive range of features. Using multiple AI models, we then generate a large-scale dataset comprising 360k training QA pairs and 10k testing QA pairs. Building on this foundation, we develop domain-specific multimodal Large Language Models (LLMs) and explore the impact of multimodal inputs on task performance. To further enhance response accuracy, we incorporate historical QA pairs as contextual grounding for answering current queries. Additionally, we incorporate page references within the answers and introduce an importance-based page classifier, further improving the precision and relevance of the information provided. Preliminary results indicate the improvements with our AI assistant in document information extraction and question-answering tasks. The dataset and models are publicly available at: https://huggingface.co/opioidarchive

Method: Proposed curvature-aware de Branges-Rovnyak space (RKHS isometric to Poincare ball) with adjustable multiplier for adaptive curvature selection, and constructed adaptive hyperbolic kernels including a novel adaptive hyperbolic radial kernel.

Result: Extensive experiments on visual and language benchmarks show the proposed kernels outperform existing hyperbolic kernels in modeling hierarchical dependencies.

Conclusion: The adaptive hyperbolic kernels effectively enhance hyperbolic representation capacity and outperform previous methods in hierarchical data modeling tasks.

Abstract: Hierarchical data pervades diverse machine learning applications, including natural language processing, computer vision, and social network analysis. Hyperbolic space, characterized by its negative curvature, has demonstrated strong potential in such tasks due to its capacity to embed hierarchical structures with minimal distortion. Previous evidence indicates that the hyperbolic representation capacity can be further enhanced through kernel methods. However, existing hyperbolic kernels still suffer from mild geometric distortion or lack adaptability. This paper addresses these issues by introducing a curvature-aware de Branges-Rovnyak space, a reproducing kernel Hilbert space (RKHS) that is isometric to a Poincare ball. We design an adjustable multiplier to select the appropriate RKHS corresponding to the hyperbolic space with any curvature adaptively. Building on this foundation, we further construct a family of adaptive hyperbolic kernels, including the novel adaptive hyperbolic radial kernel, whose learnable parameters modulate hyperbolic features in a task-aware manner. Extensive experiments on visual and language benchmarks demonstrate that our proposed kernels outperform existing hyperbolic kernels in modeling hierarchical dependencies.

Method: Template-driven protocol for dynamic instance generation, extensive experiments on SOTA LLMs across 100 years (1960-2060), and development of an LLM-powered Time Agent using tool-augmented code generation.

Result: Current LLMs achieve only 34.5% average accuracy on SPAN, with none exceeding 80%. The Time Agent solution achieves 95.31% accuracy, significantly outperforming baselines.

Conclusion: Cross-calendar temporal reasoning remains challenging for current LLMs, but tool-augmented code generation shows strong potential. Key obstacles identified are Future-Date Degradation and Calendar Asymmetry Bias.

Abstract: We introduce SPAN, a cross-calendar temporal reasoning benchmark, which requires LLMs to perform intra-calendar temporal reasoning and inter-calendar temporal conversion. SPAN features ten cross-calendar temporal reasoning directions, two reasoning types, and two question formats across six calendars. To enable time-variant and contamination-free evaluation, we propose a template-driven protocol for dynamic instance generation that enables assessment on a user-specified Gregorian date. We conduct extensive experiments on both open- and closed-source state-of-the-art (SOTA) LLMs over a range of dates spanning 100 years from 1960 to 2060. Our evaluations show that these LLMs achieve an average accuracy of only 34.5%, with none exceeding 80%, indicating that this task remains challenging. Through in-depth analysis of reasoning types, question formats, and temporal reasoning directions, we identify two key obstacles for LLMs: Future-Date Degradation and Calendar Asymmetry Bias. To strengthen LLMs’ cross-calendar temporal reasoning capability, we further develop an LLM-powered Time Agent that leverages tool-augmented code generation. Empirical results show that Time Agent achieves an average accuracy of 95.31%, outperforming several competitive baselines, highlighting the potential of tool-augmented code generation to advance cross-calendar temporal reasoning. We hope this work will inspire further efforts toward more temporally and culturally adaptive LLMs.

Method: Created ChEmREF benchmark with 1,035 HAZMAT chemicals from Emergency Response Guidebook and PubChem Database, testing three tasks: chemical representation translation, emergency response generation, and domain knowledge QA.

Result: Best models achieved 68.0% exact match on chemical translation, 52.7% LLM Judge score on incident response, and 63.9% accuracy on HAZMAT exams.

Conclusion: Language models show potential for assisting emergency responders but currently require careful human oversight due to limitations.

Abstract: Emergency responders managing hazardous material HAZMAT incidents face critical, time-sensitive decisions, manually navigating extensive chemical guidelines. We investigate whether today’s language models can assist responders by rapidly and reliably understanding critical information, identifying hazards, and providing recommendations.We introduce the Chemical Emergency Response Evaluation Framework (ChEmREF), a new benchmark comprising questions on 1,035 HAZMAT chemicals from the Emergency Response Guidebook and the PubChem Database. ChEmREF is organized into three tasks: (1) translation of chemical representation between structured and unstructured forms (e.g., converting C2H6O to ethanol), (2) emergency response generation (e.g., recommending appropriate evacuation distances) and (3) domain knowledge question answering from chemical safety and certification exams. Our best evaluated models received an exact match of 68.0% on unstructured HAZMAT chemical representation translation, a LLM Judge score of 52.7% on incident response recommendations, and a multiple-choice accuracy of 63.9% on HAMZAT examinations.These findings suggest that while language models show potential to assist emergency responders in various tasks, they require careful human oversight due to their current limitations.

Method: Introduces a Planner model for global DAG planning, ComplexTool-Plan benchmark dataset, and two-stage training (SFT + GRPO) to enhance tool selection and global planning awareness.

Result: Achieves state-of-the-art performance on StableToolBench benchmark, demonstrating superior execution capabilities and robust handling of complex multi-tool workflows.

Conclusion: The Planner-centric paradigm with DAG planning effectively resolves local optimization bottlenecks and enables optimized execution for complex queries requiring sophisticated tool coordination.

Abstract: Existing tool-augmented large language models (LLMs) encounter significant challenges when processing complex queries. Current frameworks such as ReAct are prone to local optimization traps due to their reliance on incremental decision-making processes. To address these limitations, we propose a novel Planner-centric Plan-Execute paradigm that fundamentally resolves local optimization bottlenecks through architectural innovation. Central to our approach is a novel Planner model that performs global Directed Acyclic Graph (DAG) planning for complex queries, enabling optimized execution beyond conventional tool coordination. We also introduce ComplexTool-Plan, a large-scale benchmark dataset featuring complex queries that demand sophisticated multi-tool composition and coordination capabilities. Additionally, we develop a two-stage training methodology that integrates Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO), systematically enhancing the Planner’s tool selection accuracy and global planning awareness through structured DAG-based planning. When integrated with a capable executor, our framework achieves state-of-the-art performance on the StableToolBench benchmark for complex user queries, demonstrating superior end-to-end execution capabilities and robust handling of intricate multi-tool workflows.

Method: Two-component framework: hinter (powerful LLM) provides probabilistic guidance at critical points identified by Distributional Inconsistency Reduction (DIR) metric, and practitioner (efficient smaller model) executes major reasoning steps through iterative tree updates.

Result: Achieves comparable performance to self-consistency and MCTS baselines while decoding only 1/5 tokens, and outperforms existing methods by up to 5.1% absolute accuracy while maintaining similar or lower FLOPs.

Conclusion: HPR framework effectively reduces computational costs while maintaining reasoning performance by strategically guiding smaller models at critical decision points using distributional inconsistency reduction.

Abstract: While large language models (LLMs) demonstrate emerging reasoning capabilities, current inference-time expansion methods incur prohibitive computational costs by exhaustive sampling. Through analyzing decoding trajectories, we observe that most next-token predictions align well with the golden output, except for a few critical tokens that lead to deviations. Inspired by this phenomenon, we propose a novel Hint-Practice Reasoning (HPR) framework that operationalizes this insight through two synergistic components: 1) a hinter (powerful LLM) that provides probabilistic guidance at critical decision points, and 2) a practitioner (efficient smaller model) that executes major reasoning steps. The framework’s core innovation lies in Distributional Inconsistency Reduction (DIR), a theoretically-grounded metric that dynamically identifies intervention points by quantifying the divergence between practitioner’s reasoning trajectory and hinter’s expected distribution in a tree-structured probabilistic space. Through iterative tree updates guided by DIR, HPR reweights promising reasoning paths while deprioritizing low-probability branches. Experiments across arithmetic and commonsense reasoning benchmarks demonstrate HPR’s state-of-the-art efficiency-accuracy tradeoffs: it achieves comparable performance to self-consistency and MCTS baselines while decoding only 1/5 tokens, and outperforms existing methods by at most 5.1% absolute accuracy while maintaining similar or lower FLOPs.

Method: Proposes RadFlow with clinically grounded reward hierarchy: global reward for linguistic fluency, medical correctness, and cross-sectional consistency; local section-specific reward emphasizing Impression quality; and critical-aware policy optimization for high-risk cases.

Result: Experiments on chest X-ray and carotid ultrasound datasets show RadFlow consistently improves diagnostic coherence and overall report quality compared to state-of-the-art baselines.

Conclusion: RadFlow successfully translates the structured reporting paradigm into reinforcement fine-tuning, enabling models to generate reports that are both linguistically consistent and clinically aligned with radiologist workflow.

Abstract: Radiologists compose diagnostic reports through a structured workflow: they describe visual findings, summarize them into impressions, and carefully refine statements in clinically critical cases. However, most existing medical report generation (MRG) systems treat reports as flat sequences, overlooking this hierarchical organization and leading to inconsistencies between descriptive and diagnostic content. To align model behavior with real-world reporting practices, we propose RadFlow, a hierarchical workflow-guided reinforcement optimization framework that explicitly models the structured nature of clinical reporting. RadFlow introduces a clinically grounded reward hierarchy that mirrors the organization of radiological reports. At the global level, the reward integrates linguistic fluency, medical-domain correctness, and cross-sectional consistency between Finding and Impression, promoting coherent and clinically faithful narratives. At the local level, a section-specific reward emphasizes Impression quality, reflecting its central role in diagnostic accuracy. Furthermore, a critical-aware policy optimization mechanism adaptively regularizes learning for high-risk or clinically sensitive cases, emulating the cautious refinement behavior of radiologists when documenting critical findings. Together, these components translate the structured reporting paradigm into the reinforcement fine-tuning process, enabling the model to generate reports that are both linguistically consistent and clinically aligned. Experiments on chest X-ray and carotid ultrasound datasets demonstrate that RadFlow consistently improves diagnostic coherence and overall report quality compared with state-of-the-art baselines.

Method: Data-driven approach using attribute-conditioned query generation, LLM self-evaluation on three facets, response refinement, and supervised fine-tuning with knowledge distillation.

Result: Significant performance improvements on HealthBench, with smaller models surpassing teacher models - achieving 63.8% overall and 43.1% on hard subset (new SOTA).

Conclusion: MuSeR effectively enhances LLM context-awareness in medical domains through multifaceted self-refinement, enabling better real-world performance.

Abstract: Large language models (LLMs) have shown great promise in the medical domain, achieving strong performance on several benchmarks. However, they continue to underperform in real-world medical scenarios, which often demand stronger context-awareness, i.e., the ability to recognize missing or critical details (e.g., user identity, medical history, risk factors) and provide safe, helpful, and contextually appropriate responses. To address this issue, we propose Multifaceted Self-Refinement (MuSeR), a data-driven approach that enhances LLMs’ context-awareness along three key facets (decision-making, communication, and safety) through self-evaluation and refinement. Specifically, we first design a attribute-conditioned query generator that simulates diverse real-world user contexts by varying attributes such as role, geographic region, intent, and degree of information ambiguity. An LLM then responds to these queries, self-evaluates its answers along three key facets, and refines its responses to better align with the requirements of each facet. Finally, the queries and refined responses are used for supervised fine-tuning to reinforce the model’s context-awareness ability. Evaluation results on the latest HealthBench dataset demonstrate that our method significantly improves LLM performance across multiple aspects, with particularly notable gains in the context-awareness axis. Furthermore, by incorporating knowledge distillation with the proposed method, the performance of a smaller backbone LLM (e.g., Qwen3-32B) surpasses its teacher model, achieving a new SOTA across all open-source LLMs on HealthBench (63.8%) and its hard subset (43.1%). Code and dataset will be released at https://muser-llm.github.io.

Method: A compact convolutional policy is embedded in a differentiable PyTorch reaction-diffusion simulator, using spatially smooth modulations of feed and kill parameters under a warm-hold-decay gain schedule. Training optimizes Turing-band spectral targets while penalizing control effort and instability.

Result: The hybrid approach achieves 100% strict convergence in ~165 steps, matches cell-only spectral selectivity (0.436 vs. 0.434), and uses ~15× less ℓ₁ effort and >200× less ℓ₂ power than NN-dominant control. A Goldilocks zone (A≈0.03-0.045) yields 100% quasi convergence in 94-96 steps.

Conclusion: The controller ‘seeds then cedes’ - providing brief, sparse nudges to place the system in the correct basin of attraction, after which local physics maintains the pattern, demonstrating practical morphological computation for energy-efficient embodied systems.

Abstract: We investigate how to couple a learnable brain-like’’ controller to a cell-like’’ Gray–Scott substrate to steer pattern formation with minimal effort. A compact convolutional policy is embedded in a differentiable PyTorch reaction–diffusion simulator, producing spatially smooth, bounded modulations of the feed and kill parameters ($ΔF$, $ΔK$) under a warm–hold–decay gain schedule. Training optimizes Turing-band spectral targets (FFT-based) while penalizing control effort ($\ell_1/\ell_2$) and instability. We compare three regimes: pure reaction–diffusion, NN-dominant, and a hybrid coupling. The hybrid achieves reliable, fast formation of target textures: 100% strict convergence in $\sim 165$ steps, matching cell-only spectral selectivity (0.436 vs.\ 0.434) while using $\sim 15\times$ less $\ell_1$ effort and $>200\times$ less $\ell_2$ power than NN-dominant control. An amplitude sweep reveals a non-monotonic Goldilocks’’ zone ($A \approx 0.03$–$0.045$) that yields 100% quasi convergence in 94–96 steps, whereas weaker or stronger gains fail to converge or degrade selectivity. These results quantify morphological computation: the controller seeds then cedes,’’ providing brief, sparse nudges that place the system in the correct basin of attraction, after which local physics maintains the pattern. The study offers a practical recipe for building steerable, robust, and energy-efficient embodied systems that exploit an optimal division of labor between centralized learning and distributed self-organization.

Method: Uses state neural network architecture to capture neuron-like dynamic processes and neuron output prediction objective to train the system to predict neuronal outputs from collective dynamics.

Result: Establishes a biologically grounded and computationally scalable foundation for building systems capable of generalization, reasoning, and adaptive learning across domains.

Conclusion: Represents a step toward truly artificial general intelligence by learning general principles of intelligence from diverse intelligent behaviors rather than domain-specific patterns.

Abstract: We propose a new perspective for approaching artificial general intelligence (AGI) through an intelligence foundation model (IFM). Unlike existing foundation models (FMs), which specialize in pattern learning within specific domains such as language, vision, or time series, IFM aims to acquire the underlying mechanisms of intelligence by learning directly from diverse intelligent behaviors. Vision, language, and other cognitive abilities are manifestations of intelligent behavior; learning from this broad range of behaviors enables the system to internalize the general principles of intelligence. Based on the fact that intelligent behaviors emerge from the collective dynamics of biological neural systems, IFM consists of two core components: a novel network architecture, termed the state neural network, which captures neuron-like dynamic processes, and a new learning objective, neuron output prediction, which trains the system to predict neuronal outputs from collective dynamics. The state neural network emulates the temporal dynamics of biological neurons, allowing the system to store, integrate, and process information over time, while the neuron output prediction objective provides a unified computational principle for learning these structural dynamics from intelligent behaviors. Together, these innovations establish a biologically grounded and computationally scalable foundation for building systems capable of generalization, reasoning, and adaptive learning across domains, representing a step toward truly AGI.

Method: RAGFort uses structure-aware dual-module defense: contrastive reindexing creates semantic isolation between classes, and constrained cascade generation limits fine-grained knowledge extraction within classes.

Result: Experiments show RAGFort significantly reduces reconstruction attack success while maintaining answer quality, providing comprehensive protection across security, performance, and robustness metrics.

Conclusion: Joint protection of both intra-class and inter-class paths is essential for effective defense against knowledge base extraction attacks, and RAGFort offers a practical solution that balances security with system performance.

Abstract: Retrieval-Augmented Generation (RAG) systems deployed over proprietary knowledge bases face growing threats from reconstruction attacks that aggregate model responses to replicate knowledge bases. Such attacks exploit both intra-class and inter-class paths, progressively extracting fine-grained knowledge within topics and diffusing it across semantically related ones, thereby enabling comprehensive extraction of the original knowledge base. However, existing defenses target only one path, leaving the other unprotected. We conduct a systematic exploration to assess the impact of protecting each path independently and find that joint protection is essential for effective defense. Based on this, we propose RAGFort, a structure-aware dual-module defense combining “contrastive reindexing” for inter-class isolation and “constrained cascade generation” for intra-class protection. Experiments across security, performance, and robustness confirm that RAGFort significantly reduces reconstruction success while preserving answer quality, offering comprehensive defense against knowledge base extraction attacks.

Method: Proposes DenoGrad framework that leverages gradients from an accurate Deep Learning model trained on target data to detect and adjust noisy samples dynamically, preserving the problem’s data distribution.

Result: DenoGrad outperforms state-of-the-art denoising strategies on both tabular and time series datasets under various noise settings, enhancing interpretable AI model performance while preserving original data distribution.

Conclusion: DenoGrad provides a more precise and adaptable noise definition by using task-specific high quality solutions as reference, making it the only high quality approach that preserves original data distribution while improving AI model robustness.

Abstract: The performance of Machine Learning (ML) models, particularly those operating within the Interpretable Artificial Intelligence (Interpretable AI) framework, is significantly affected by the presence of noise in both training and production data. Denoising has therefore become a critical preprocessing step, typically categorized into instance removal and instance correction techniques. However, existing correction approaches often degrade performance or oversimplify the problem by altering the original data distribution. This leads to unrealistic scenarios and biased models, which is particularly problematic in contexts where interpretable AI models are employed, as their interpretability depends on the fidelity of the underlying data patterns. In this paper, we argue that defining noise independently of the solution may be ineffective, as its nature can vary significantly across tasks and datasets. Using a task-specific high quality solution as a reference can provide a more precise and adaptable noise definition. To this end, we propose DenoGrad, a novel Gradient-based instance Denoiser framework that leverages gradients from an accurate Deep Learning (DL) model trained on the target data – regardless of the specific task – to detect and adjust noisy samples. Unlike conventional approaches, DenoGrad dynamically corrects noisy instances, preserving problem’s data distribution, and improving AI models robustness. DenoGrad is validated on both tabular and time series datasets under various noise settings against the state-of-the-art. DenoGrad outperforms existing denoising strategies, enhancing the performance of interpretable IA models while standing out as the only high quality approach that preserves the original data distribution.

Method: Developed two novel compilers that can compile away qualitative state-trajectory constraints (safety requirements, task ordering, intermediate sub-goals) without grounding the problem first.

Result: The methods produce planning specifications that are orders of magnitude more succinct than ground-based compilers while remaining competitive with state-of-the-art planners in empirical evaluations on International Planning Competition domains.

Conclusion: The proposed grounding-free compilation methods efficiently handle large-scale planning problems with constraints, significantly reducing specification size while maintaining planning performance.

Abstract: We study planning in a fragment of PDDL with qualitative state-trajectory constraints, capturing safety requirements, task ordering conditions, and intermediate sub-goals commonly found in real-world problems. A prominent approach to tackle such problems is to compile their constraints away, leading to a problem that is supported by state-of-the-art planners. Unfortunately, existing compilers do not scale on problems with a large number of objects and high-arity actions, as they necessitate grounding the problem before compilation. To address this issue, we propose two methods for compiling away constraints without grounding, making them suitable for large-scale planning problems. We prove the correctness of our compilers and outline their worst-case time complexity. Moreover, we present a reproducible empirical evaluation on the domains used in the latest International Planning Competition. Our results demonstrate that our methods are efficient and produce planning specifications that are orders of magnitude more succinct than the ones produced by compilers that ground the domain, while remaining competitive when used for planning with a state-of-the-art planner.

Method: Trains a Gaussian Process surrogate model with ‘LogitMap Pairs’ from a minimal but informative training subset to approximate foundation model outputs, guiding proxy model training and reducing API queries.

Result: Elevates pre-trained language model accuracy from 55.92% to 86.85%, reducing API query frequency to only 1.38%, outperforming offline approaches and achieving comparable/superior accuracy to query-intensive methods.

Conclusion: Provides a robust and high-efficiency paradigm for language model adaptation that balances performance and computational cost effectively.

Abstract: Black-box tuning is an emerging paradigm for adapting large language models (LLMs) to better achieve desired behaviors, particularly when direct access to model parameters is unavailable. Current strategies, however, often present a dilemma of suboptimal extremes: either separately train a small proxy model and then use it to shift the predictions of the foundation model, offering notable efficiency but often yielding limited improvement; or making API calls in each tuning iteration to the foundation model, which entails prohibitive computational costs. Therefore, we propose a novel advanced black-box tuning method for LLMs with limited API calls. Our core strategy involves training a Gaussian Process (GP) surrogate model with “LogitMap Pairs” derived from querying the foundation model on a minimal but highly informative training subset. This surrogate can approximate the outputs of the foundation model to guide the training of the proxy model, thereby effectively reducing the need for direct queries to the foundation model. Extensive experiments verify that our approach elevates pre-trained language model accuracy from 55.92% to 86.85%, reducing the frequency of API queries to merely 1.38%. This significantly outperforms offline approaches that operate entirely without API access. Notably, our method also achieves comparable or superior accuracy to query-intensive approaches, while significantly reducing API costs. This offers a robust and high-efficiency paradigm for language model adaptation.

Method: Proposes MTP framework with three branches: numeric learning using frequency MLPs, visual learning through frequency and periodicity images, and textual learning via descriptive text augmentation. Uses hierarchical contrastive learning to fuse the three modalities.

Result: Extensive experiments on six real-world datasets demonstrate superior performance compared to state-of-the-art approaches.

Conclusion: The multimodal approach effectively captures comprehensive traffic signal information and improves prediction accuracy by integrating numeric, visual, and textual perspectives.

Abstract: With rapid urbanization in the modern era, traffic signals from various sensors have been playing a significant role in monitoring the states of cities, which provides a strong foundation in ensuring safe travel, reducing traffic congestion and optimizing urban mobility. Most existing methods for traffic signal modeling often rely on the original data modality, i.e., numerical direct readings from the sensors in cities. However, this unimodal approach overlooks the semantic information existing in multimodal heterogeneous urban data in different perspectives, which hinders a comprehensive understanding of traffic signals and limits the accurate prediction of complex traffic dynamics. To address this problem, we propose a novel \textit{M}ultimodal framework, \textit{MTP}, for urban \textit{T}raffic \textit{P}rofiling, which learns multimodal features through numeric, visual, and textual perspectives. The three branches drive for a multimodal perspective of urban traffic signal learning in the frequency domain, while the frequency learning strategies delicately refine the information for extraction. Specifically, we first conduct the visual augmentation for the traffic signals, which transforms the original modality into frequency images and periodicity images for visual learning. Also, we augment descriptive texts for the traffic signals based on the specific topic, background information and item description for textual learning. To complement the numeric information, we utilize frequency multilayer perceptrons for learning on the original modality. We design a hierarchical contrastive learning on the three branches to fuse the spectrum of three modalities. Finally, extensive experiments on six real-world datasets demonstrate superior performance compared with the state-of-the-art approaches.

Method: Evolutionary module guided by LLMs generates synthetic instances with realistic structural patterns that statistically mimic real-world instances, followed by progressive refinement of pre-trained NCO models through alignment with synthetic distributions and fine-tuning on benchmark instances.

Result: EvoReal significantly improves generalization capabilities of neural solvers, achieving performance gaps of 1.05% on TSPLib and 2.71% on CVRPLib compared to optimal solutions across various problem scales.

Conclusion: The proposed EvoReal framework effectively bridges the generalization gap in neural combinatorial optimization by generating realistic synthetic instances and progressive model refinement, enabling better performance on real-world benchmark problems.

Abstract: Recent advances in Neural Combinatorial Optimization (NCO) methods have significantly improved the capability of neural solvers to handle synthetic routing instances. Nonetheless, existing neural solvers typically struggle to generalize effectively from synthetic, uniformly-distributed training data to real-world VRP scenarios, including widely recognized benchmark instances from TSPLib and CVRPLib. To bridge this generalization gap, we present Evolutionary Realistic Instance Synthesis (EvoReal), which leverages an evolutionary module guided by large language models (LLMs) to generate synthetic instances characterized by diverse and realistic structural patterns. Specifically, the evolutionary module produces synthetic instances whose structural attributes statistically mimics those observed in authentic real-world instances. Subsequently, pre-trained NCO models are progressively refined, firstly aligning them with these structurally enriched synthetic distributions and then further adapting them through direct fine-tuning on actual benchmark instances. Extensive experimental evaluations demonstrate that EvoReal markedly improves the generalization capabilities of state-of-the-art neural solvers, yielding a notable reduced performance gap compared to the optimal solutions on the TSPLib (1.05%) and CVRPLib (2.71%) benchmarks across a broad spectrum of problem scales.

Method: Decompose complex questions into sub-questions, progressively extend partial reasoning paths, use external retrievers for candidate evidence, apply uncertainty-aware pruning by LLM, and optimize context organization.

Result: Outperforms existing baselines in multi-hop KGQA on three well-known datasets, demonstrating improved reliability and reasoning quality.

Conclusion: ProgRAG effectively addresses retrieval and reasoning failures in KG-enhanced LLMs through progressive reasoning path extension and uncertainty-aware evidence refinement.

Abstract: Large Language Models (LLMs) demonstrate strong reasoning capabilities but struggle with hallucinations and limited transparency. Recently, KG-enhanced LLMs that integrate knowledge graphs (KGs) have been shown to improve reasoning performance, particularly for complex, knowledge-intensive tasks. However, these methods still face significant challenges, including inaccurate retrieval and reasoning failures, often exacerbated by long input contexts that obscure relevant information or by context constructions that struggle to capture the richer logical directions required by different question types. Furthermore, many of these approaches rely on LLMs to directly retrieve evidence from KGs, and to self-assess the sufficiency of this evidence, which often results in premature or incorrect reasoning. To address the retrieval and reasoning failures, we propose ProgRAG, a multi-hop knowledge graph question answering (KGQA) framework that decomposes complex questions into sub-questions, and progressively extends partial reasoning paths by answering each sub-question. At each step, external retrievers gather candidate evidence, which is then refined through uncertainty-aware pruning by the LLM. Finally, the context for LLM reasoning is optimized by organizing and rearranging the partial reasoning paths obtained from the sub-question answers. Experiments on three well-known datasets demonstrate that ProgRAG outperforms existing baselines in multi-hop KGQA, offering improved reliability and reasoning quality.

Method: PepTriX integrates 1D sequence embeddings and 3D structural features via graph attention network enhanced with contrastive training and cross-modal co-attention, automatically adapting to diverse datasets.

Result: PepTriX performs remarkably well across multiple peptide classification tasks and provides interpretable insights into structural and biophysical motifs that drive predictions.

Conclusion: PepTriX bridges the gap between performance-driven peptide-level models and domain-level understanding, offering both predictive robustness and interpretable validation for peptide research.

Abstract: Peptide classification tasks, such as predicting toxicity and HIV inhibition, are fundamental to bioinformatics and drug discovery. Traditional approaches rely heavily on handcrafted encodings of one-dimensional (1D) peptide sequences, which can limit generalizability across tasks and datasets. Recently, protein language models (PLMs), such as ESM-2 and ESMFold, have demonstrated strong predictive performance. However, they face two critical challenges. First, fine-tuning is computationally costly. Second, their complex latent representations hinder interpretability for domain experts. Additionally, many frameworks have been developed for specific types of peptide classification, lacking generalization. These limitations restrict the ability to connect model predictions to biologically relevant motifs and structural properties. To address these limitations, we present PepTriX, a novel framework that integrates one dimensional (1D) sequence embeddings and three-dimensional (3D) structural features via a graph attention network enhanced with contrastive training and cross-modal co-attention. PepTriX automatically adapts to diverse datasets, producing task-specific peptide vectors while retaining biological plausibility. After evaluation by domain experts, we found that PepTriX performs remarkably well across multiple peptide classification tasks and provides interpretable insights into the structural and biophysical motifs that drive predictions. Thus, PepTriX offers both predictive robustness and interpretable validation, bridging the gap between performance-driven peptide-level models (PLMs) and domain-level understanding in peptide research.

Method: Performs limited-horizon searches and updates each state’s heuristic based on the shortest path to the search frontier, incorporating edge costs and heuristic values of frontier states.

Result: A generalized approach that improves heuristic learning for shortest-path problems in sequential decision-making.

Conclusion: The proposed method provides a more effective framework for learning heuristics in sequential decision-making problems compared to traditional single-step approaches.

Abstract: Many sequential decision-making problems can be formulated as shortest-path problems, where the objective is to reach a goal state from a given starting state. Heuristic search is a standard approach for solving such problems, relying on a heuristic function to estimate the cost to the goal from any given state. Recent approaches leverage reinforcement learning to learn heuristics by applying deep approximate value iteration. These methods typically rely on single-step Bellman updates, where the heuristic of a state is updated based on its best neighbor and the corresponding edge cost. This work proposes a generalized approach that enhances both state sampling and heuristic updates by performing limited-horizon searches and updating each state’s heuristic based on the shortest path to the search frontier, incorporating both edge costs and the heuristic values of frontier states.

Method: Analyzes verification of CI propositions using linear temporal logic (LTL) and non-deterministic Büchi automata (NBAs), distinguishing between stochastic and structural properties.

Result: Stochastic CI verification is at least as hard as the Skolem problem, while structural CI verification is in PSPACE with NP/coNP-hardness, with tractable cases identified under certain graphical restrictions.

Conclusion: Structural CI verification is more tractable than stochastic verification, and natural restrictions on DBN structure can make the problem practically solvable.

Abstract: Dynamic Bayesian networks (DBNs) are compact graphical representations used to model probabilistic systems where interdependent random variables and their distributions evolve over time. In this paper, we study the verification of the evolution of conditional-independence (CI) propositions against temporal logic specifications. To this end, we consider two specification formalisms over CI propositions: linear temporal logic (LTL), and non-deterministic Büchi automata (NBAs). This problem has two variants. Stochastic CI properties take the given concrete probability distributions into account, while structural CI properties are viewed purely in terms of the graphical structure of the DBN. We show that deciding if a stochastic CI proposition eventually holds is at least as hard as the Skolem problem for linear recurrence sequences, a long-standing open problem in number theory. On the other hand, we show that verifying the evolution of structural CI propositions against LTL and NBA specifications is in PSPACE, and is NP- and coNP-hard. We also identify natural restrictions on the graphical structure of DBNs that make the verification of structural CI properties tractable.

Method: Introduce causal analysis into LVLM object recognition using Structural Causal Model (SCM), formally define spurious correlations from co-occurrence bias, develop Causal-HalBench benchmark with counterfactual samples, and create an extensible pipeline using LVLMs and T2I models for sample generation.

Result: Evaluations on mainstream LVLMs using Causal-HalBench demonstrate these models exhibit susceptibility to spurious correlations, though to varying extents.

Conclusion: The proposed causal framework and benchmark effectively quantify spurious correlations in LVLMs, revealing their vulnerability to co-occurrence bias and providing tools for assessing model robustness against such biases.

Abstract: Large Vision-Language Models (LVLMs) often suffer from object hallucination, making erroneous judgments about the presence of objects in images. We propose this primar- ily stems from spurious correlations arising when models strongly associate highly co-occurring objects during train- ing, leading to hallucinated objects influenced by visual con- text. Current benchmarks mainly focus on hallucination de- tection but lack a formal characterization and quantitative evaluation of spurious correlations in LVLMs. To address this, we introduce causal analysis into the object recognition scenario of LVLMs, establishing a Structural Causal Model (SCM). Utilizing the language of causality, we formally de- fine spurious correlations arising from co-occurrence bias. To quantify the influence induced by these spurious correla- tions, we develop Causal-HalBench, a benchmark specifically constructed with counterfactual samples and integrated with comprehensive causal metrics designed to assess model ro- bustness against spurious correlations. Concurrently, we pro- pose an extensible pipeline for the construction of these coun- terfactual samples, leveraging the capabilities of proprietary LVLMs and Text-to-Image (T2I) models for their genera- tion. Our evaluations on mainstream LVLMs using Causal- HalBench demonstrate these models exhibit susceptibility to spurious correlations, albeit to varying extents.

Method: The authors build upon the optimal bidirectional search algorithm BAE* (designed for consistent heuristics) and develop several variants specifically adapted for bounded-suboptimal search contexts.

Result: Experimental evaluation shows that each algorithm performs best under different conditions, revealing the specific strengths and weaknesses of each approach.

Conclusion: The study demonstrates that different bounded-suboptimal bidirectional algorithms excel in distinct scenarios, providing insights into when to use each method based on specific problem conditions.

Abstract: Recent advancements in bidirectional heuristic search have yielded significant theoretical insights and novel algorithms. While most previous work has concentrated on optimal search methods, this paper focuses on bounded-suboptimal bidirectional search, where a bound on the suboptimality of the solution cost is specified. We build upon the state-of-the-art optimal bidirectional search algorithm, BAE*, designed for consistent heuristics, and introduce several variants of BAE* specifically tailored for the bounded-suboptimal context. Through experimental evaluation, we compare the performance of these new variants against other bounded-suboptimal bidirectional algorithms as well as the standard weighted A* algorithm. Our results demonstrate that each algorithm excels under distinct conditions, highlighting the strengths and weaknesses of each approach.

Method: Fine-tune SLMs with specific NPC personas and integrate runtime-swappable memory modules that preserve character context and world knowledge.

Result: Evaluated three SLMs (DistilGPT-2, TinyLlama-1.1B-Chat, Mistral-7B-Instruct) on consumer hardware with synthetic persona-aligned data, enabling expressive interactions without retraining.

Conclusion: The modular design and persona-driven memory architecture have broader potential for scalable conversational agents in virtual assistants, customer support, and educational systems.

Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in generating human-like text, yet their applicability to dialogue systems in computer games remains limited. This limitation arises from their substantial hardware requirements, latency constraints, and the necessity to maintain clearly defined knowledge boundaries within a game setting. In this paper, we propose a modular NPC dialogue system that leverages Small Language Models (SLMs), fine-tuned to encode specific NPC personas and integrated with runtime-swappable memory modules. These memory modules preserve character-specific conversational context and world knowledge, enabling expressive interactions and long-term memory without retraining or model reloading during gameplay. We comprehensively evaluate our system using three open-source SLMs: DistilGPT-2, TinyLlama-1.1B-Chat, and Mistral-7B-Instruct, trained on synthetic persona-aligned data and benchmarked on consumer-grade hardware. While our approach is motivated by applications in gaming, its modular design and persona-driven memory architecture hold significant potential for broader adoption in domains requiring expressive, scalable, and memory-rich conversational agents, such as virtual assistants, customer support bots, or interactive educational systems.

Method: Proposes FactGuard framework that leverages LLMs to extract event-centric content, introduces dynamic usability mechanism to identify contradictions and ambiguous cases, and employs knowledge distillation to create FactGuard-D for resource-constrained scenarios.

Result: Comprehensive experiments on two benchmark datasets show consistent outperformance of existing methods in both robustness and accuracy, effectively addressing style sensitivity and LLM usability challenges.

Conclusion: FactGuard successfully reduces writing style impact on detection performance, provides reliable decision-making through adaptive LLM integration, and enables practical deployment through efficient distilled versions.

Abstract: Fake news detection methods based on writing style have achieved remarkable progress. However, as adversaries increasingly imitate the style of authentic news, the effectiveness of such approaches is gradually diminishing. Recent research has explored incorporating large language models (LLMs) to enhance fake news detection. Yet, despite their transformative potential, LLMs remain an untapped goldmine for fake news detection, with their real-world adoption hampered by shallow functionality exploration, ambiguous usability, and prohibitive inference costs. In this paper, we propose a novel fake news detection framework, dubbed FactGuard, that leverages LLMs to extract event-centric content, thereby reducing the impact of writing style on detection performance. Furthermore, our approach introduces a dynamic usability mechanism that identifies contradictions and ambiguous cases in factual reasoning, adaptively incorporating LLM advice to improve decision reliability. To ensure efficiency and practical deployment, we employ knowledge distillation to derive FactGuard-D, enabling the framework to operate effectively in cold-start and resource-constrained scenarios. Comprehensive experiments on two benchmark datasets demonstrate that our approach consistently outperforms existing methods in both robustness and accuracy, effectively addressing the challenges of style sensitivity and LLM usability in fake news detection.

Method: Formal framework using abductive explanations to identify minimal sufficient evidence for model decisions, introducing Potentially Applicable Explanations (PAE) to detect privacy leakage at individual and system levels.

Result: Experimental evaluation on German Credit Dataset shows how sensitive feature importance affects privacy leakage; abductive reasoning enables interpretable privacy auditing despite computational challenges.

Conclusion: Abductive reasoning provides a practical approach to reconcile transparency, model interpretability, and privacy preservation in AI decision-making through human-understandable explanations.

Abstract: Privacy leakage in AI-based decision processes poses significant risks, particularly when sensitive information can be inferred. We propose a formal framework to audit privacy leakage using abductive explanations, which identifies minimal sufficient evidence justifying model decisions and determines whether sensitive information disclosed. Our framework formalizes both individual and system-level leakage, introducing the notion of Potentially Applicable Explanations (PAE) to identify individuals whose outcomes can shield those with sensitive features. This approach provides rigorous privacy guarantees while producing human understandable explanations, a key requirement for auditing tools. Experimental evaluation on the German Credit Dataset illustrates how the importance of sensitive literal in the model decision process affects privacy leakage. Despite computational challenges and simplifying assumptions, our results demonstrate that abductive reasoning enables interpretable privacy auditing, offering a practical pathway to reconcile transparency, model interpretability, and privacy preserving in AI decision-making.

Method: Two-level parallelization with shared information among workers, enhanced with Grundy numbers for game tree reduction, applied to Sprouts game solving.

Result: Achieved 332.9x speedup on 1024 cores, generated proofs 1000x more complex, verified Sprouts Conjecture for 42 new positions (nearly doubling known outcomes), and outperformed GLOP solver by 4 orders of magnitude in runtime.

Conclusion: The massively parallel Proof-Number Search approach enables efficient scaling on large CPU clusters and significantly advances game solving capabilities for complex combinatorial games like Sprouts.

Abstract: Proof-Number Search is a best-first search algorithm with many successful applications, especially in game solving. As large-scale computing clusters become increasingly accessible, parallelization is a natural way to accelerate computation. However, existing parallel versions of Proof-Number Search are known to scale poorly on many CPU cores. Using two parallelized levels and shared information among workers, we present the first massively parallel version of Proof-Number Search that scales efficiently even on a large number of CPUs. We apply our solver, enhanced with Grundy numbers for reducing game trees, to the Sprouts game, a case study motivated by the long-standing Sprouts Conjecture. Our solver achieves a significantly improved 332.9$\times$ speedup when run on 1024 cores, enabling it to outperform the state-of-the-art Sprouts solver GLOP by four orders of magnitude in runtime and to generate proofs 1,000$\times$ more complex. Despite exponential growth in game tree size, our solver verified the Sprouts Conjecture for 42 new positions, nearly doubling the number of known outcomes.

Method: Uses formalized abstract structures as modular templates, generates Lean definitions via LLMs with feedback, integrates via typeclasses, and verifies structural assumptions.

Result: Successfully formalized diverse optimization problem instances grounded in abstract structures in Lean proof assistant.

Conclusion: SITA effectively bridges abstract mathematical theories with concrete applications, enabling automated formalization of research-level mathematical results.

Abstract: While large language models (LLMs) have shown progress in mathematical reasoning, they still face challenges in formalizing theorems that arise from instantiating abstract structures in concrete settings. With the goal of auto-formalizing mathematical results at the research level, we develop a framework for structure-to-instance theorem autoformalization (SITA), which systematically bridges the gap between abstract mathematical theories and their concrete applications in Lean proof assistant. Formalized abstract structures are treated as modular templates that contain definitions, assumptions, operations, and theorems. These templates serve as reusable guides for the formalization of concrete instances. Given a specific instantiation, we generate corresponding Lean definitions and instance declarations, integrate them using Lean’s typeclass mechanism, and construct verified theorems by checking structural assumptions. We incorporate LLM-based generation with feedback-guided refinement to ensure both automation and formal correctness. Experiments on a dataset of optimization problems demonstrate that SITA effectively formalizes diverse instances grounded in abstract structures.

Method: Developed methods to generate policy summarizations capturing task ordering and agent cooperation in decentralized MARL policies, along with query-based explanations for When, Why Not, and What types of user queries about specific agent behaviors.

Result: Evaluated across four MARL domains and two decentralized MARL algorithms, demonstrating generalizability and computational efficiency. User studies showed significant improvements in user question-answering performance and enhanced subjective ratings on understanding and satisfaction metrics.

Conclusion: The proposed summarizations and explanations effectively address the challenges of decentralized MARL, providing valuable tools for understanding agent behaviors and improving user experience in multi-agent systems.

Abstract: Multi-Agent Reinforcement Learning (MARL) has gained significant interest in recent years, enabling sequential decision-making across multiple agents in various domains. However, most existing explanation methods focus on centralized MARL, failing to address the uncertainty and nondeterminism inherent in decentralized settings. We propose methods to generate policy summarizations that capture task ordering and agent cooperation in decentralized MARL policies, along with query-based explanations for When, Why Not, and What types of user queries about specific agent behaviors. We evaluate our approach across four MARL domains and two decentralized MARL algorithms, demonstrating its generalizability and computational efficiency. User studies show that our summarizations and explanations significantly improve user question-answering performance and enhance subjective ratings on metrics such as understanding and satisfaction.

Method: The authors introduce a unified framework based on parameterized analogies defined via generalized means. They characterize analogy-preserving functions and derive error bounds under smoothness assumptions.

Result: The framework subsumes both Boolean classification and regression, supports analogical inference over continuous functions, and provides worst-case and average-case error bounds. The authors also identify limitations of existing generalization bounds through counterexamples.

Conclusion: The paper presents a comprehensive theory of analogical inference that works across discrete and continuous domains, addressing fundamental limitations of previous approaches and providing formal guarantees for analogical reasoning in real-valued settings.

Abstract: Analogical reasoning is a powerful inductive mechanism, widely used in human cognition and increasingly applied in artificial intelligence. Formal frameworks for analogical inference have been developed for Boolean domains, where inference is provably sound for affine functions and approximately correct for functions close to affine. These results have informed the design of analogy-based classifiers. However, they do not extend to regression tasks or continuous domains. In this paper, we revisit analogical inference from a foundational perspective. We first present a counterexample showing that existing generalization bounds fail even in the Boolean setting. We then introduce a unified framework for analogical reasoning in real-valued domains based on parameterized analogies defined via generalized means. This model subsumes both Boolean classification and regression, and supports analogical inference over continuous functions. We characterize the class of analogy-preserving functions in this setting and derive both worst-case and average-case error bounds under smoothness assumptions. Our results offer a general theory of analogical inference across discrete and continuous domains.

Method: Proposes a framework of abstract proofs to represent both proofs and explanations, with methods to convert proofs to step-wise explanation sequences using trimming and simplification techniques.

Result: The method significantly speeds up step-wise explanation generation while maintaining similar quality to state-of-the-art approaches.

Conclusion: Using solver-generated proofs as a starting point for explanation sequences is an effective approach that improves computational efficiency without sacrificing explanation quality.

Abstract: In the field of Explainable Constraint Solving, it is common to explain to a user why a problem is unsatisfiable. A recently proposed method for this is to compute a sequence of explanation steps. Such a step-wise explanation shows individual reasoning steps involving constraints from the original specification, that in the end explain a conflict. However, computing a step-wise explanation is computationally expensive, limiting the scope of problems for which it can be used. We investigate how we can use proofs generated by a constraint solver as a starting point for computing step-wise explanations, instead of computing them step-by-step. More specifically, we define a framework of abstract proofs, in which both proofs and step-wise explanations can be represented. We then propose several methods for converting a proof to a step-wise explanation sequence, with special attention to trimming and simplification techniques to keep the sequence and its individual steps small. Our results show our method significantly speeds up the generation of step-wise explanation sequences, while the resulting step-wise explanation has a quality similar to the current state-of-the-art.

Method: Proposes two dynamic normalization techniques to handle varying scales of multiple sub-objectives, and introduces MACHOP (Multi-Armed CHOice Perceptron) - a query generation strategy combining non-domination constraints with upper confidence bound-based diversification.

Result: Evaluation on Sudokus and Logic-Grid puzzles with artificial users and real-user validation shows MACHOP consistently produces higher-quality explanations than standard approaches.

Conclusion: Interactive preference elicitation with MACHOP and dynamic normalization effectively learns user preferences for step-wise explanations, overcoming limitations of standard combinatorial problem elicitation and producing better quality explanations.

Abstract: Step-wise explanations can explain logic puzzles and other satisfaction problems by showing how to derive decisions step by step. Each step consists of a set of constraints that derive an assignment to one or more decision variables. However, many candidate explanation steps exist, with different sets of constraints and different decisions they derive. To identify the most comprehensible one, a user-defined objective function is required to quantify the quality of each step. However, defining a good objective function is challenging. Here, interactive preference elicitation methods from the wider machine learning community can offer a way to learn user preferences from pairwise comparisons. We investigate the feasibility of this approach for step-wise explanations and address several limitations that distinguish it from elicitation for standard combinatorial problems. First, because the explanation quality is measured using multiple sub-objectives that can vary a lot in scale, we propose two dynamic normalization techniques to rescale these features and stabilize the learning process. We also observed that many generated comparisons involve similar explanations. For this reason, we introduce MACHOP (Multi-Armed CHOice Perceptron), a novel query generation strategy that integrates non-domination constraints with upper confidence bound-based diversification. We evaluate the elicitation techniques on Sudokus and Logic-Grid puzzles using artificial users, and validate them with a real-user evaluation. In both settings, MACHOP consistently produces higher-quality explanations than the standard approach.

Method: Proposed S4F Standpoint Logic by generalizing both S4F modal logic and standpoint propositional logic, defining its syntax, semantics, and analyzing computational complexity. Also outlined mechanisms for credulous and sceptical acceptance.

Result: S4F Standpoint Logic is computationally no harder than its constituent logics in both monotonic and non-monotonic forms. The framework successfully expresses multi-viewpoint, non-monotonic semantic commitments.

Conclusion: S4F Standpoint Logic provides an effective formalism that unifies standpoint logic and non-monotonic reasoning in modal logic, maintaining computational tractability while enabling rich multi-viewpoint reasoning with acceptance mechanisms.

Abstract: Standpoint logics offer unified modal logic-based formalisms for representing multiple heterogeneous viewpoints. At the same time, many non-monotonic reasoning frameworks can be naturally captured using modal logics, in particular using the modal logic S4F. In this work, we propose a novel formalism called S4F Standpoint Logic, which generalises both S4F and standpoint propositional logic and is therefore capable of expressing multi-viewpoint, non-monotonic semantic commitments. We define its syntax and semantics and analyze its computational complexity, obtaining the result that S4F Standpoint Logic is not computationally harder than its constituent logics, whether in monotonic or non-monotonic form. We also outline mechanisms for credulous and sceptical acceptance and illustrate the framework with an example.

Method: Comprehensive review and analysis of three main approaches: Graph Neural Networks for modeling relationships and interactions, Multiagent Deep Reinforcement Learning for learning in multiagent environments, and Probabilistic Topic Modeling for estimating unknown distributions and handling heterogeneity.

Result: Identifies the potential of these methods to overcome traditional limitations and handle real-world challenges like uncertainty, heterogeneity, and scalability in strategic multiagent settings.

Conclusion: While promising approaches exist, key open challenges remain including handling non-stationary environments, balancing stability and adaptation, tackling uncertainty and heterogeneity, and ensuring scalability and solution tractability.

Abstract: This paper provides a comprehensive review of mainly Graph Neural Networks, Deep Reinforcement Learning, and Probabilistic Topic Modeling methods with a focus on their potential incorporation in strategic multiagent settings. We draw interest in (i) Machine Learning methods currently utilized for uncovering unknown model structures adaptable to the task of strategic opponent modeling, and (ii) the integration of these methods with Game Theoretic concepts that avoid relying on assumptions often invalid in real-world scenarios, such as the Common Prior Assumption (CPA) and the Self-Interest Hypothesis (SIH). We analyze the ability to handle uncertainty and heterogeneity, two characteristics that are very common in real-world application cases, as well as scalability. As a potential answer to effectively modeling relationships and interactions in multiagent settings, we champion the use of Graph Neural Networks (GNN). Such approaches are designed to operate upon graph-structured data, and have been shown to be a very powerful tool for performing tasks such as node classification and link prediction. Next, we review the domain of Reinforcement Learning (RL), and in particular that of Multiagent Deep Reinforcement Learning (MADRL). Following, we describe existing relevant game theoretic solution concepts and consider properties such as fairness and stability. Our review comes complete with a note on the literature that utilizes PTM in domains other than that of document analysis and classification. The capability of PTM to estimate unknown underlying distributions can help with tackling heterogeneity and unknown agent beliefs. Finally, we identify certain open challenges specifically, the need to (i) fit non-stationary environments, (ii) balance the degrees of stability and adaptation, (iii) tackle uncertainty and heterogeneity, (iv) guarantee scalability and solution tractability.

Method: Commentary analysis examining AI’s role in managing information overload, filtering literature, surfacing interdisciplinary links, generating hypotheses, and designing experiments.

Result: AI systems now help researchers with literature management, cross-disciplinary connections, hypothesis generation, and experimental design, marking a fundamental shift in scientific practice.

Conclusion: AI must augment human judgment in academic workflows like peer review and validation, requiring deliberate adoption through policies promoting transparency, reproducibility, and accountability.

Abstract: Artificial intelligence (AI) is reshaping how research is conceived, conducted, and communicated across fields from chemistry to biomedicine. This commentary examines how AI is transforming the research workflow. AI systems now help researchers manage the information deluge, filtering the literature, surfacing cross-disciplinary links for ideas and collaborations, generating hypotheses, and designing and executing experiments. These developments mark a shift from AI as a mere computational tool to AI as an active collaborator in science. Yet this transformation demands thoughtful integration and governance. We argue that at this time AI must augment but not replace human judgment in academic workflows such as peer review, ethical evaluation, and validation of results. This paper calls for the deliberate adoption of AI within the scientific practice through policies that promote transparency, reproducibility, and accountability.

Method: Proposes a bi-level framework: meta-level optimizes subgroup budget allocations with fairness constraints; base-level identifies responsive individuals using neural networks trained on observational data, incorporating resource-specific delay kernels and cooldown windows.

Result: Validated on education and workforce development datasets, achieving higher cumulative outcomes, better adaptation to delay structures, and equitable distribution across subgroups compared to existing approaches.

Conclusion: The delay-aware, data-driven framework demonstrates potential to improve institutional policy and social welfare through more responsive and adaptive decision-making in resource allocation systems.

Abstract: Equitably allocating limited resources in high-stakes domains-such as education, employment, and healthcare-requires balancing short-term utility with long-term impact, while accounting for delayed outcomes, hidden heterogeneity, and ethical constraints. However, most learning-based allocation frameworks either assume immediate feedback or ignore the complex interplay between individual characteristics and intervention dynamics. We propose a novel bi-level contextual bandit framework for individualized resource allocation under delayed feedback, designed to operate in real-world settings with dynamic populations, capacity constraints, and time-sensitive impact. At the meta level, the model optimizes subgroup-level budget allocations to satisfy fairness and operational constraints. At the base level, it identifies the most responsive individuals within each group using a neural network trained on observational data, while respecting cooldown windows and delayed treatment effects modeled via resource-specific delay kernels. By explicitly modeling temporal dynamics and feedback delays, the algorithm continually refines its policy as new data arrive, enabling more responsive and adaptive decision-making. We validate our approach on two real-world datasets from education and workforce development, showing that it achieves higher cumulative outcomes, better adapts to delay structures, and ensures equitable distribution across subgroups. Our results highlight the potential of delay-aware, data-driven decision-making systems to improve institutional policy and social welfare.

Method: Uses a multi-language approach with a low-level finite automaton language for rules definition, plus higher-level languages for game design that translate to the low-level language.

Result: RG generates faster forward models than current state-of-the-art GGP systems (Regular Boardgames, Ludii) and includes a comprehensive ecosystem with editor, visualization, benchmarking, and debugging tools.

Conclusion: RG successfully combines computational efficiency with game design convenience through its automaton-based approach and multi-language system, outperforming existing GGP systems in efficiency.

Abstract: We propose a new General Game Playing (GGP) system called Regular Games (RG). The main goal of RG is to be both computationally efficient and convenient for game design. The system consists of several languages. The core component is a low-level language that defines the rules by a finite automaton. It is minimal with only a few mechanisms, which makes it easy for automatic processing (by agents, analysis, optimization, etc.). The language is universal for the class of all finite turn-based games with imperfect information. Higher-level languages are introduced for game design (by humans or Procedural Content Generation), which are eventually translated to a low-level language. RG generates faster forward models than the current state of the art, beating other GGP systems (Regular Boardgames, Ludii) in terms of efficiency. Additionally, RG’s ecosystem includes an editor with LSP, automaton visualization, benchmarking tools, and a debugger of game description transformations.

Method: Introduces a formal definition for matching labeled time series sensor data to abstract scenarios represented as Scenic probabilistic programming language programs, and presents a querying algorithm to identify matching data subsets.

Result: The algorithm is more accurate and orders of magnitude faster than state-of-the-art commercial vision large language models, and scales effectively with the duration of queried time series data.

Conclusion: The proposed approach provides an effective method for validating simulated failure scenarios against real-world datasets, bridging the sim-to-real gap in CPS testing.

Abstract: Simulation-based testing has become a crucial complement to road testing for ensuring the safety of cyber physical systems (CPS). As a result, significant research efforts have been directed toward identifying failure scenarios within simulation environments. However, a critical question remains. Are the AV failure scenarios discovered in simulation reproducible on actual systems in the real world? The sim-to-real gap caused by differences between simulated and real sensor data means that failure scenarios identified in simulation might either be artifacts of synthetic sensor data or actual issues that also occur with real sensor data. To address this, an effective approach to validating simulated failure scenarios is to locate occurrences of these scenarios within real-world datasets and verify whether the failure persists on the datasets. To this end, we introduce a formal definition of how labeled time series sensor data can match an abstract scenario, represented as a scenario program using the Scenic probabilistic programming language. We present a querying algorithm that, given a scenario program and a labeled dataset, identifies the subset of data that matches the specified scenario. Our experiment shows that our algorithm is more accurate and orders of magnitude faster in querying scenarios than the state-of-the-art commercial vision large language models, and can scale with the duration of queried time series data.

Method: Systematic review categorizing MCER methods into four categories: context-free modeling, sequential context modeling, speaker-differentiated modeling, and speaker-relationship modeling.

Result: Provides classification framework for MCER methods and discusses available datasets, feature extraction techniques, applications in affective computing and human-computer interaction.

Conclusion: The survey helps researchers understand current MCER research status and develop more efficient models by providing systematic categorization and identifying future development directions.

Abstract: Multi-modal conversation emotion recognition (MCER) aims to recognize and track the speaker’s emotional state using text, speech, and visual information in the conversation scene. Analyzing and studying MCER issues is significant to affective computing, intelligent recommendations, and human-computer interaction fields. Unlike the traditional single-utterance multi-modal emotion recognition or single-modal conversation emotion recognition, MCER is a more challenging problem that needs to deal with more complex emotional interaction relationships. The critical issue is learning consistency and complementary semantics for multi-modal feature fusion based on emotional interaction relationships. To solve this problem, people have conducted extensive research on MCER based on deep learning technology, but there is still a lack of systematic review of the modeling methods. Therefore, a timely and comprehensive overview of MCER’s recent advances in deep learning is of great significance to academia and industry. In this survey, we provide a comprehensive overview of MCER modeling methods and roughly divide MCER methods into four categories, i.e., context-free modeling, sequential context modeling, speaker-differentiated modeling, and speaker-relationship modeling. In addition, we further discuss MCER’s publicly available popular datasets, multi-modal feature extraction methods, application areas, existing challenges, and future development directions. We hope that our review can help MCER researchers understand the current research status in emotion recognition, provide some inspiration, and develop more efficient models.

Method: Formulated discussion-graph semantics for first-order logic with equality, generalized Dung’s notion of extensions to handle equivalent graph nodes, and established connections between these two contributions.

Result: Showed that generalized extensions are first-order characterizable within the proposed semantics, with propositional characterizability of all Dung’s extensions and acceptability semantics as immediate consequences.

Conclusion: The work provides a more general formal framework for reasoning about discussion and argumentation, extending classical argumentation theory to first-order logic while maintaining compatibility with existing propositional approaches.

Abstract: We make three contributions. First, we formulate a discussion-graph semantics for first-order logic with equality, enabling reasoning about discussion and argumentation in AI more generally than before. This addresses the current lack of a formal reasoning framework capable of handling diverse discussion and argumentation models. Second, we generalise Dung’s notion of extensions to cases where two or more graph nodes in an argumentation framework are equivalent. Third, we connect these two contributions by showing that the generalised extensions are first-order characterisable within the proposed discussion-graph semantics. Propositional characterisability of all Dung’s extensions is an immediate consequence. We furthermore show that the set of all generalised extensions (acceptability semantics), too, are first-order characterisable. Propositional characterisability of all Dung’s acceptability semantics is an immediate consequence.

Method: Developed a dataset of abstract argumentation frameworks with explanations, fine-tuned LLMs on this data, and compared with chain-of-thought baselines.

Result: Models trained with explanations achieved superior generalization accuracy compared to question-answer only training, and provided transparent self-explanations.

Conclusion: Integrating formal argumentation with process explanations significantly improves LLMs’ conflict resolution capabilities and addresses neural network transparency issues.

Abstract: In recent years, large language models (LLMs) have made significant advancements in developing human-like and engaging dialogue systems. However, in tasks such as consensus-building and persuasion, LLMs often struggle to resolve conflicts arising from incomplete or inconsistent information, revealing their limitations in real-world applications. Given these limitations, abstract argumentation, a specialized logical framework designed to resolve conflicts and inconsistencies, becomes particularly relevant. In this paper, we aim to enhance the conflict-solving capabilities of LLMs by leveraging formal abstract argumentation, integrating language model learning with symbolic computation. To achieve this, we develop and curate a dataset comprising diverse abstract argumentation frameworks, accompanied by detailed explanations of the argument acceptability computation process. Subsequently, we fine-tune LLMs on this dataset, focusing on abstract conflict resolution tasks. As a comparative baseline, LLMs are also evaluated using a chain-of-thought approach, however, they fail to solve the conflict-based arguments effectively. Our experiments demonstrate that process explanations play a crucial role in learning. Models trained with explanations exhibit superior generalization accuracy compared to those trained solely on question-answer pairs. Furthermore, leveraging LLMs’ self-explanation capabilities, our approach provides detailed illustrations that mitigate the lack of transparency typically associated with neural networks.

Method: Combine differentiable dynamics with state predictors in end-to-end computation graphs to create three novel task setups: learning relative odometry, optimal planners, and optimal inverse states.

Result: Successfully developed Analytic World Models (AWMs) that enable efficient end-to-end learning and demonstrate broad applicability in autonomous driving scenarios.

Conclusion: Differentiable simulation enables efficient learning of world models that augment agent decision-making beyond reactive control, providing predictive, prescriptive, and counterfactual capabilities.

Abstract: Differentiable simulators represent an environment’s dynamics as a differentiable function. Within robotics and autonomous driving, this property is used in Analytic Policy Gradients (APG), which relies on backpropagating through the dynamics to train accurate policies for diverse tasks. Here we show that differentiable simulation also has an important role in world modeling, where it can impart predictive, prescriptive, and counterfactual capabilities to an agent. Specifically, we design three novel task setups in which the differentiable dynamics are combined within an end-to-end computation graph not with a policy, but a state predictor. This allows us to learn relative odometry, optimal planners, and optimal inverse states. We collectively call these predictors Analytic World Models (AWMs) and demonstrate how differentiable simulation enables their efficient, end-to-end learning. In autonomous driving scenarios, they have broad applicability and can augment an agent’s decision-making beyond reactive control.

Method: Propose Mix Distillation - a strategy that balances reasoning complexity by combining long and short chain-of-thought examples or reasoning from both larger and smaller models.

Result: Mix Distillation significantly improves small model reasoning performance compared to training on either long or short reasoning data alone.

Conclusion: Direct strong model distillation has limitations, and adapting reasoning complexity is crucial for effective reasoning capability transfer to small models.

Abstract: Large language models (LLMs) excel in complex reasoning tasks, and distilling their reasoning capabilities into smaller models has shown promise. However, we uncover an interesting phenomenon, which we term the Small Model Learnability Gap: small models ($\leq$3B parameters) do not consistently benefit from long chain-of-thought (CoT) reasoning or distillation from larger models. Instead, they perform better when fine-tuned on shorter, simpler reasoning chains that better align with their intrinsic learning capacity. To address this, we propose Mix Distillation, a simple yet effective strategy that balances reasoning complexity by combining long and short CoT examples or reasoning from both larger and smaller models. Our experiments demonstrate that Mix Distillation significantly improves small model reasoning performance compared to training on either data alone. These findings highlight the limitations of direct strong model distillation and underscore the importance of adapting reasoning complexity for effective reasoning capability transfer.

Method: PITA learns a small preference-based guidance policy to modify token probabilities during inference without LLM fine-tuning, using stochastic search and iterative refinement to identify underlying preference distributions.

Result: Evaluation across mathematical reasoning and sentiment classification tasks shows PITA effectively aligns LLM outputs with user preferences without reward models.

Conclusion: PITA provides an efficient and effective alternative to reward model-dependent alignment methods, enabling direct preference integration during inference with reduced computational costs.

Abstract: Inference-time alignment enables large language models (LLMs) to generate outputs aligned with end-user preferences without further training. Recent post-training methods achieve this by using small guidance models to modify token generation during inference. These methods typically optimize a reward function KL-regularized by the original LLM taken as the reference policy. A critical limitation, however, is their dependence on a pre-trained reward model, which requires fitting to human preference feedback–a potentially unstable process. In contrast, we introduce PITA, a novel framework that integrates preference feedback directly into the LLM’s token generation, eliminating the need for a reward model. PITA learns a small preference-based guidance policy to modify token probabilities at inference time without LLM fine-tuning, reducing computational cost and bypassing the pre-trained reward model dependency. The problem is framed as identifying an underlying preference distribution, solved through stochastic search and iterative refinement of the preference-based guidance model. We evaluate PITA across diverse tasks, including mathematical reasoning and sentiment classification, demonstrating its effectiveness in aligning LLM outputs with user preferences.

Method: Developed a KG-based tool retrieval framework using ensembles of 1-hop ego tool graphs to model direct and indirect connections between tools, enabling comprehensive contextual tool selection for multi-step tasks.

Result: Achieved 91.85% tool coverage on micro-average CompleteRecall metric, outperforming the strongest non-KG baseline (89.26%) and demonstrating better performance particularly for sequential tool composition queries.

Conclusion: Structural information modeled in knowledge graphs provides complementary signals to pure similarity matching, especially beneficial for queries requiring sequential tool composition, validating the effectiveness of the proposed approach.

Abstract: Effective tool pre-selection via retrieval is essential for AI agents to select from a vast array of tools when identifying and planning actions in the context of complex user queries. Despite its central role in planning, this aspect remains underexplored in the literature. Traditional approaches rely primarily on similarities between user queries and tool descriptions, which significantly limits retrieval accuracy, specifically when handling multi-step user requests. To address these limitations, we propose a Knowledge Graph (KG)-based tool retrieval framework that captures the semantic relationships between tools and their functional dependencies. Our retrieval algorithm leverages ensembles of 1-hop ego tool graphs to model direct and indirect connections between tools, enabling more comprehensive and contextual tool selection for multi-step tasks. We evaluate our approach on a synthetically generated internal dataset across six defined user classes, extending previous work on coherent dialogue synthesis and tool retrieval benchmarks. Results demonstrate that our tool graph-based method achieves 91.85% tool coverage on the micro-average CompleteRecall metric, compared to 89.26% for re-ranked semantic-lexical hybrid retrieval, the strongest non-KG baseline in our experiments. These findings support our hypothesis that the structural information modeled in the graph provides complementary signals to pure similarity matching, particularly for queries requiring sequential tool composition.

Method: Evaluated multiple LLM-based agent architectures (single-agent to modular designs) across realistic penetration testing scenarios, with targeted augmentations for five core capabilities: Global Context Memory, Inter-Agent Messaging, Context-Conditioned Invocation, Adaptive Planning, and Real-Time Monitoring.

Result: Targeted augmentations substantially improve modular agent performance, especially in complex, multi-step, and real-time penetration testing tasks, though some architectures natively exhibit subsets of these properties.

Conclusion: Systematic augmentation of core functional capabilities is crucial for improving LLM-based penetration testing agents, particularly for handling complex multi-step attack scenarios and real-time responsiveness requirements.

Abstract: Large language models (LLMs) are increasingly used to automate or augment penetration testing, but their effectiveness and reliability across attack phases remain unclear. We present a comprehensive evaluation of multiple LLM-based agents, from single-agent to modular designs, across realistic penetration testing scenarios, measuring empirical performance and recurring failure patterns. We also isolate the impact of five core functional capabilities via targeted augmentations: Global Context Memory (GCM), Inter-Agent Messaging (IAM), Context-Conditioned Invocation (CCI), Adaptive Planning (AP), and Real-Time Monitoring (RTM). These interventions support, respectively: (i) context coherence and retention, (ii) inter-component coordination and state management, (iii) tool use accuracy and selective execution, (iv) multi-step strategic planning, error detection, and recovery, and (v) real-time dynamic responsiveness. Our results show that while some architectures natively exhibit subsets of these properties, targeted augmentations substantially improve modular agent performance, especially in complex, multi-step, and real-time penetration testing tasks.

Method: The benchmark includes problems of varying difficulty levels, with a structured dataset and evaluation methodology for testing various language models.

Result: Initial testing shows a correlation between model performance and the number of tokens used, and reveals performance differences between mathematics and physics tasks.

Conclusion: DOoM provides a valuable benchmark for evaluating Russian language models on STEM problems, with results indicating token count correlation and subject-specific performance variations.

Abstract: This paper introduces DOoM, a new open-source benchmark designed to assess the capabilities of language models in solving mathematics and physics problems in Russian. The benchmark includes problems of varying difficulty, ranging from school-level tasks to university Olympiad and entrance exam questions. In this paper we discuss the motivation behind its creation, describe dataset’s structure and evaluation methodology, and present initial results from testing various models. Analysis of the results shows a correlation between model performance and the number of tokens used, and highlights differences in performance between mathematics and physics tasks.

Method: Created PsychCounsel-Bench, a benchmark with 2,252 single-choice questions based on U.S. national counselor examinations, requiring ~70% accuracy to pass.

Result: Advanced models (GPT-4o, Llama3.3-70B, Gemma3-27B) exceeded the passing threshold, while smaller models (Qwen2.5-7B, Mistral-7B) performed poorly.

Conclusion: Only frontier LLMs currently meet counseling exam standards, showing promise but also challenges for psychology-oriented LLM development.

Abstract: Large Language Models (LLMs) have demonstrated remarkable success across a wide range of industries, primarily due to their impressive generative abilities. Yet, their potential in applications requiring cognitive abilities, such as psychological counseling, remains largely untapped. This paper investigates the key question: \textit{Can LLMs be effectively applied to psychological counseling?} To determine whether an LLM can effectively take on the role of a psychological counselor, the first step is to assess whether it meets the qualifications required for such a role, namely the ability to pass the U.S. National Counselor Certification Exam (NCE). This is because, just as a human counselor must pass a certification exam to practice, an LLM must demonstrate sufficient psychological knowledge to meet the standards required for such a role. To address this, we introduce PsychCounsel-Bench, a benchmark grounded in U.S.national counselor examinations, a licensure test for professional counselors that requires about 70% accuracy to pass. PsychCounsel-Bench comprises approximately 2,252 carefully curated single-choice questions, crafted to require deep understanding and broad enough to cover various sub-disciplines of psychology. This benchmark provides a comprehensive assessment of an LLM’s ability to function as a counselor. Our evaluation shows that advanced models such as GPT-4o, Llama3.3-70B, and Gemma3-27B achieve well above the passing threshold, while smaller open-source models (e.g., Qwen2.5-7B, Mistral-7B) remain far below it. These results suggest that only frontier LLMs are currently capable of meeting counseling exam standards, highlighting both the promise and the challenges of developing psychology-oriented LLMs. We release the proposed dataset for public use: https://github.com/cloversjtu/PsychCounsel-Bench

Method: Created LiveResearchBench with 100 expert-curated tasks requiring real-time web search across daily life, enterprise, and academia domains, and developed DeepEval evaluation suite covering content and report quality metrics.

Result: Comprehensive evaluation of 17 frontier deep research systems revealed current capabilities, recurring failure modes, and identified key system components needed for advancement.

Conclusion: The benchmark and evaluation framework provide a rigorous basis for systematically advancing deep research systems, highlighting both current strengths and areas needing improvement in reliable, insightful information synthesis.

Abstract: Deep research – producing comprehensive, citation-grounded reports by searching and synthesizing information from hundreds of live web sources – marks an important frontier for agentic systems. To rigorously evaluate this ability, four principles are essential: tasks should be (1) user-centric, reflecting realistic information needs, (2) dynamic, requiring up-to-date information beyond parametric knowledge, (3) unambiguous, ensuring consistent interpretation across users, and (4) multi-faceted and search-intensive, requiring search over numerous web sources and in-depth analysis. Existing benchmarks fall short of these principles, often focusing on narrow domains or posing ambiguous questions that hinder fair comparison. Guided by these principles, we introduce LiveResearchBench, a benchmark of 100 expert-curated tasks spanning daily life, enterprise, and academia, each requiring extensive, dynamic, real-time web search and synthesis. Built with over 1,500 hours of human labor, LiveResearchBench provides a rigorous basis for systematic evaluation. To evaluate citation-grounded long-form reports, we introduce DeepEval, a comprehensive suite covering both content- and report-level quality, including coverage, presentation, citation accuracy and association, consistency and depth of analysis. DeepEval integrates four complementary evaluation protocols, each designed to ensure stable assessment and high agreement with human judgments. Using LiveResearchBench and DeepEval, we conduct a comprehensive evaluation of 17 frontier deep research systems, including single-agent web search, single-agent deep research, and multi-agent systems. Our analysis reveals current strengths, recurring failure modes, and key system components needed to advance reliable, insightful deep research.

Method: Developed a multi-agent AI system integrating free-text descriptions with ontology labels using retrieval-augmented generation (RAG) to refine predictions with PubMed literature, reducing hallucinations and enhancing interpretability.

Result: Achieved 77% correct annotations for mouse gene sets among top predictions. Annotated 5,322 brain cell clusters from mouse brain cell atlas, identified region-specific gene co-expression patterns, inferred functional roles of gene ensembles, and identified Basal Ganglia-related cell types with neurologically meaningful descriptions.

Conclusion: BRAINCELL-AID creates a valuable resource for community-driven cell type annotation by providing more accurate and robust gene set annotation that enables novel insights into brain cell function.

Abstract: Single-cell RNA sequencing has transformed our ability to identify diverse cell types and their transcriptomic signatures. However, annotating these signatures-especially those involving poorly characterized genes-remains a major challenge. Traditional methods, such as Gene Set Enrichment Analysis (GSEA), depend on well-curated annotations and often perform poorly in these contexts. Large Language Models (LLMs) offer a promising alternative but struggle to represent complex biological knowledge within structured ontologies. To address this, we present BRAINCELL-AID (BRAINCELL-AID: https://biodataai.uth.edu/BRAINCELL-AID), a novel multi-agent AI system that integrates free-text descriptions with ontology labels to enable more accurate and robust gene set annotation. By incorporating retrieval-augmented generation (RAG), we developed a robust agentic workflow that refines predictions using relevant PubMed literature, reducing hallucinations and enhancing interpretability. Using this workflow, we achieved correct annotations for 77% of mouse gene sets among their top predictions. Applying this approach, we annotated 5,322 brain cell clusters from the comprehensive mouse brain cell atlas generated by the BRAIN Initiative Cell Census Network, enabling novel insights into brain cell function by identifying region-specific gene co-expression patterns and inferring functional roles of gene ensembles. BRAINCELL-AID also identifies Basal Ganglia-related cell types with neurologically meaningful descriptions. Hence, we create a valuable resource to support community-driven cell type annotation.

Method: GHOST uses hierarchical search with novel abstract-path-unfolding algorithm to compute admissible lower bounds, guiding best-first search at both tour level and GCS path level while avoiding unnecessary convex optimization calls.

Result: GHOST is orders-of-magnitude faster than mixed-integer convex programming baselines and uniquely handles complex trajectory planning with high-order continuity constraints and incomplete GCS.

Conclusion: GHOST provides an optimal and efficient solution for GCS-TSP with strong theoretical guarantees and practical performance, enabling complex trajectory planning previously intractable.

Abstract: We study GCS-TSP, a new variant of the Traveling Salesman Problem (TSP) defined over a Graph of Convex Sets (GCS) – a powerful representation for trajectory planning that decomposes the configuration space into convex regions connected by a sparse graph. In this setting, edge costs are not fixed but depend on the specific trajectory selected through each convex region, making classical TSP methods inapplicable. We introduce GHOST, a hierarchical framework that optimally solves the GCS-TSP by combining combinatorial tour search with convex trajectory optimization. GHOST systematically explores tours on a complete graph induced by the GCS, using a novel abstract-path-unfolding algorithm to compute admissible lower bounds that guide best-first search at both the high level (over tours) and the low level (over feasible GCS paths realizing the tour). These bounds provide strong pruning power, enabling efficient search while avoiding unnecessary convex optimization calls. We prove that GHOST guarantees optimality and present a bounded-suboptimal variant for time-critical scenarios. Experiments show that GHOST is orders-of-magnitude faster than unified mixed-integer convex programming baselines for simple cases and uniquely handles complex trajectory planning problems involving high-order continuity constraints and an incomplete GCS.

Method: The authors formalize a five-phase AI lifecycle mapped to LCA stages, specify governance via PDCA cycles, systematize hardware strategies across edge-cloud continuum, and define a calibrated measurement framework combining estimator models with direct metering.

Result: The framework makes energy, carbon, water, and embodied impacts first-class considerations and enables reproducible, provider-agnostic comparisons of AI environmental impacts.

Conclusion: The article provides actionable, evidence-based guidance combining definition, lifecycle processes, hardware strategies, and calibrated measurement for implementing Green AI across research, practice, and policy.

Abstract: Across the Artificial Intelligence (AI) lifecycle - from hardware to development, deployment, and reuse - burdens span energy, carbon, water, and embodied impacts. Cloud provider tools improve transparency but remain heterogeneous and often omit water and value chain effects, limiting comparability and reproducibility. Addressing these multi dimensional burdens requires a lifecycle approach linking phase explicit mapping with system levers (hardware, placement, energy mix, cooling, scheduling) and calibrated measurement across facility, system, device, and workload levels. This article (i) establishes a unified, operational definition of Green AI distinct from Sustainable AI; (ii) formalizes a five phase lifecycle mapped to Life Cycle Assessment (LCA) stages, making energy, carbon, water, and embodied impacts first class; (iii) specifies governance via Plan Do Check Act (PDCA) cycles with decision gateways; (iv) systematizes hardware and system level strategies across the edge cloud continuum to reduce embodied burdens; and (v) defines a calibrated measurement framework combining estimator models with direct metering to enable reproducible, provider agnostic comparisons. Combining definition, lifecycle processes, hardware strategies, and calibrated measurement, this article offers actionable, evidence based guidance for researchers, practitioners, and policymakers.

Method: Sorts vocabulary by model probability, partitions by residue rank mod k to distribute adjacent tokens across classes, applies fixed bias to selected class. Uses entropy-adaptive gate for zero-bit setting and payload digit selection for multi-bit regime.

Result: WaterMod achieves strong watermark detection performance while maintaining generation quality across natural language generation, mathematical reasoning, and code synthesis tasks in both zero-bit and multi-bit settings.

Conclusion: The modular arithmetic approach supports both binary attribution and rich payloads, providing a robust watermarking solution that preserves fluency while enabling fine-grained provenance tracing.

Abstract: Large language models now draft news, legal analyses, and software code with human-level fluency. At the same time, regulations such as the EU AI Act mandate that each synthetic passage carry an imperceptible, machine-verifiable mark for provenance. Conventional logit-based watermarks satisfy this requirement by selecting a pseudorandom green vocabulary at every decoding step and boosting its logits, yet the random split can exclude the highest-probability token and thus erode fluency. WaterMod mitigates this limitation through a probability-aware modular rule. The vocabulary is first sorted in descending model probability; the resulting ranks are then partitioned by the residue rank mod k, which distributes adjacent-and therefore semantically similar-tokens across different classes. A fixed bias of small magnitude is applied to one selected class. In the zero-bit setting (k=2), an entropy-adaptive gate selects either the even or the odd parity as the green list. Because the top two ranks fall into different parities, this choice embeds a detectable signal while guaranteeing that at least one high-probability token remains available for sampling. In the multi-bit regime (k>2), the current payload digit d selects the color class whose ranks satisfy rank mod k = d. Biasing the logits of that class embeds exactly one base-k digit per decoding step, thereby enabling fine-grained provenance tracing. The same modular arithmetic therefore supports both binary attribution and rich payloads. Experimental results demonstrate that WaterMod consistently attains strong watermark detection performance while maintaining generation quality in both zero-bit and multi-bit settings. This robustness holds across a range of tasks, including natural language generation, mathematical reasoning, and code synthesis. Our code and data are available at https://github.com/Shinwoo-Park/WaterMod.

Method: Proposed Path-Neighbor enhanced GNN (PN-GNN) that aggregates node-neighbor embeddings on reasoning paths. Analyzed existing GNN methods’ limitations and theoretically investigated PN-GNN’s logical expressive power.

Result: PN-GNN has strictly stronger expressive power than C-GNN, with (k+1)-hop logical expressiveness strictly superior to k-hop. Validated on 6 synthetic and 2 real-world datasets with competitive KG reasoning performance.

Conclusion: PN-GNN effectively enhances logical rule expressive power without compromising generalization, as confirmed by both theoretical analysis and extensive experiments in KG reasoning tasks.

Abstract: Graph neural networks (GNNs) can effectively model structural information of graphs, making them widely used in knowledge graph (KG) reasoning. However, existing studies on the expressive power of GNNs mainly focuses on simple single-relation graphs, and there is still insufficient discussion on the power of GNN to express logical rules in KGs. How to enhance the logical expressive power of GNNs is still a key issue. Motivated by this, we propose Path-Neighbor enhanced GNN (PN-GNN), a method to enhance the logical expressive power of GNN by aggregating node-neighbor embeddings on the reasoning path. First, we analyze the logical expressive power of existing GNN-based methods and point out the shortcomings of the expressive power of these methods. Then, we theoretically investigate the logical expressive power of PN-GNN, showing that it not only has strictly stronger expressive power than C-GNN but also that its $(k+1)$-hop logical expressiveness is strictly superior to that of $k$-hop. Finally, we evaluate the logical expressive power of PN-GNN on six synthetic datasets and two real-world datasets. Both theoretical analysis and extensive experiments confirm that PN-GNN enhances the expressive power of logical rules without compromising generalization, as evidenced by its competitive performance in KG reasoning tasks.

Method: Proposes information capacity based on the correlation between compression and intelligence, measuring text compression performance relative to computational complexity. Evaluated 49 models on 5 heterogeneous datasets.

Result: Models of varying sizes within a series exhibit consistent information capacity. The metric enables fair efficiency comparisons across model series and accurate performance prediction within a model series. It effectively incorporates tokenizer efficiency, which is often neglected.

Conclusion: Information capacity provides a unified framework for evaluating LLM efficiency that accounts for tokenizer efficiency, pretraining data, and architectural differences like mixture-of-experts, offering consistent and reliable efficiency assessment.

Abstract: Recent years have witnessed the rapid advancements of large language models (LLMs) and their expanding applications, leading to soaring demands for computational resources. The widespread adoption of test-time scaling further aggravates the tension between model capability and resource consumption, highlighting the importance of inference efficiency. However, a unified metric that accurately reflects an LLM’s efficiency across different model sizes and architectures remains absent. Motivated by the correlation between compression and intelligence, we introduce information capacity, a measure of model efficiency based on text compression performance relative to computational complexity. Larger models can predict the next token more accurately, achieving greater compression gains but at higher computational costs. Empirical evaluations on mainstream open-source models show that models of varying sizes within a series exhibit consistent information capacity. This metric enables a fair efficiency comparison across model series and accurate performance prediction within a model series. A distinctive feature of information capacity is that it incorporates tokenizer efficiency, which affects both input and output token counts but is often neglected in LLM evaluations. We assess the information capacity of 49 models on 5 heterogeneous datasets and observe consistent results on the influences of tokenizer efficiency, pretraining data, and the mixture-of-experts architecture.

Method: Uses pretrained multilingual acoustic and linguistic transformer models with dynamic fusion architecture inspired by Mixture of Experts, including robust preprocessing with automatic transcription, LLM-based anomaly detection, and SHAP-based explainability.

Result: Achieved AUC=0.88 and F1=0.72 for classifying cognitively healthy, MCI, and AD individuals, and AUC=0.90 and F1=0.62 for MCI detection, with minimal bias except for adults over 80.

Conclusion: SpeechCARE shows promise for real-world deployment in care settings and EHR integration, with future work focusing on underrepresented populations in New York City.

Abstract: Alzheimer’s disease and related dementias (ADRD) affect one in five adults over 60, yet more than half of individuals with cognitive decline remain undiagnosed. Speech-based assessments show promise for early detection, as phonetic motor planning deficits alter acoustic features (e.g., pitch, tone), while memory and language impairments lead to syntactic and semantic errors. However, conventional speech-processing pipelines with hand-crafted features or general-purpose audio classifiers often exhibit limited performance and generalizability. To address these limitations, we introduce SpeechCARE, a multimodal speech processing pipeline that leverages pretrained, multilingual acoustic and linguistic transformer models to capture subtle speech-related cues associated with cognitive impairment. Inspired by the Mixture of Experts (MoE) paradigm, SpeechCARE employs a dynamic fusion architecture that weights transformer-based acoustic, linguistic, and demographic inputs, allowing integration of additional modalities (e.g., social factors, imaging) and enhancing robustness across diverse tasks. Its robust preprocessing includes automatic transcription, large language model (LLM)-based anomaly detection, and task identification. A SHAP-based explainability module and LLM reasoning highlight each modality’s contribution to decision-making. SpeechCARE achieved AUC = 0.88 and F1 = 0.72 for classifying cognitively healthy, MCI, and AD individuals, with AUC = 0.90 and F1 = 0.62 for MCI detection. Bias analysis showed minimal disparities, except for adults over 80. Mitigation techniques included oversampling and weighted loss. Future work includes deployment in real-world care settings (e.g., VNS Health, Columbia ADRC) and EHR-integrated explainability for underrepresented populations in New York City.

Method: Filtered 4.8M synthetic examples to 12K clean instances, then trained 3B-parameter Vision-Language Model using supervised fine-tuning, chain-of-thought-augmented fine-tuning, and reinforcement learning via Group Relative Policy Optimization.

Result: Models trained with filtered data and lightweight strategies match or surpass larger baselines on ScreenSpot, Multimodal-Mind2Web, and AndroidControl benchmarks.

Conclusion: Principled data curation and robust adaptation can rival large-scale training, enabling compact yet capable multimodal reasoning agents.

Abstract: Visual grounding is the task of localising image regions from natural language queries and is critical for reasoning capable Graphical User Interface agents. Many existing methods rely on massive, noisy synthetic datasets.This work introduces an efficient training pipeline that combines model-based data filtering with parameter-efficient fine-tuning. From 4.8M synthetic examples, 12K clean and diverse instances are curated by first identifying challenging cases, removing misaligned and then selecting a diverse set of multimodal instances. On this data, a 3B-parameter Vision-Language Model is trained under three regimes: supervised fine-tuning, chain-of-thought-augmented fine-tuning, and reinforcement learning via Group Relative Policy Optimization. Models trained with the filtered data and lightweight training strategies match or surpass larger baselines on benchmarks such as ScreenSpot, Multimodal-Mind2Web, and AndroidControl. These results demonstrate that principled data curation and robust adaptation can rival large-scale training, enabling compact yet capable multimodal reasoning agents.

Method: Train Self-Organizing Maps on harmful prompt representations to identify multiple neurons, then subtract the centroid of harmless representations from each neuron to derive multiple refusal directions.

Result: Ablating multiple directions outperforms single-direction baseline and specialized jailbreak algorithms, effectively suppressing refusal behavior in models.

Conclusion: The approach demonstrates the mechanistic implications of using multiple directions for refusal behavior in language models.

Abstract: Refusal refers to the functional behavior enabling safety-aligned language models to reject harmful or unethical prompts. Following the growing scientific interest in mechanistic interpretability, recent work encoded refusal behavior as a single direction in the model’s latent space; e.g., computed as the difference between the centroids of harmful and harmless prompt representations. However, emerging evidence suggests that concepts in LLMs often appear to be encoded as a low-dimensional manifold embedded in the high-dimensional latent space. Motivated by these findings, we propose a novel method leveraging Self-Organizing Maps (SOMs) to extract multiple refusal directions. To this end, we first prove that SOMs generalize the prior work’s difference-in-means technique. We then train SOMs on harmful prompt representations to identify multiple neurons. By subtracting the centroid of harmless representations from each neuron, we derive a set of multiple directions expressing the refusal concept. We validate our method on an extensive experimental setup, demonstrating that ablating multiple directions from models’ internals outperforms not only the single-direction baseline but also specialized jailbreak algorithms, leading to an effective suppression of refusal. Finally, we conclude by analyzing the mechanistic implications of our approach.

Method: Uses hierarchical video parsing as a music conductor, modality-specific encoders with storyboard-guided cross-attention, position/duration encoding for temporal coherence, and frame-level transition-beat aligner for rhythmic synchronization.

Result: Experimental results demonstrate superiority in semantic relevance and rhythmic precision, particularly on the novel video-music dataset with strict transition-beat synchronization requirements.

Conclusion: VeM effectively addresses the limitations of current approaches by generating high-quality soundtracks with comprehensive semantic, temporal, and rhythmic alignment for input videos.

Abstract: Video-to-Music generation seeks to generate musically appropriate background music that enhances audiovisual immersion for videos. However, current approaches suffer from two critical limitations: 1) incomplete representation of video details, leading to weak alignment, and 2) inadequate temporal and rhythmic correspondence, particularly in achieving precise beat synchronization. To address the challenges, we propose Video Echoed in Music (VeM), a latent music diffusion that generates high-quality soundtracks with semantic, temporal, and rhythmic alignment for input videos. To capture video details comprehensively, VeM employs a hierarchical video parsing that acts as a music conductor, orchestrating multi-level information across modalities. Modality-specific encoders, coupled with a storyboard-guided cross-attention mechanism (SG-CAtt), integrate semantic cues while maintaining temporal coherence through position and duration encoding. For rhythmic precision, the frame-level transition-beat aligner and adapter (TB-As) dynamically synchronize visual scene transitions with music beats. We further contribute a novel video-music paired dataset sourced from e-commerce advertisements and video-sharing platforms, which imposes stricter transition-beat synchronization requirements. Meanwhile, we introduce novel metrics tailored to the task. Experimental results demonstrate superiority, particularly in semantic relevance and rhythmic precision.

Method: Displays multiple MIDI tracks on a single time-aligned grid with synchronized audio playback, enabling comparison of pitch, timing, note detection errors, and section-level patterns.

Result: Enables users to identify missed/extra notes, observe onset/offset differences, and compare transcription outputs visually and audibly in real-time.

Conclusion: The open-source library facilitates better AMT model evaluation through interactive comparative visualization and is available at https://github.com/crescent-stdio/wave-roll.

Abstract: WaveRoll is an interactive JavaScript library that enables comparative visualization and synchronized playback of multiple MIDI piano rolls on a browser. It addresses a specific evaluation need in Automatic Music Transcription (AMT), contrasting multiple MIDI outputs produced from the same input. The library displays multiple MIDI tracks on a single, time-aligned grid with synchronized audio, allowing users to compare pitch and timing, identify missed or extra notes, and observe onset and offset differences, as well as section-level patterns. We expect that such comparisons would assist in model evaluation and error analysis, and help readers to understand the model behavior better. The open-source library is available at https://github.com/crescent-stdio/wave-roll

Method: Uses text feature encoder for text/phoneme/tone/BERT features, processes frame-level SSL features into phoneme-level via average pooling and attention mechanisms, and incorporates duration predictor for speech rate alignment.

Result: Outperforms competing systems in naturalness, similarity, and content integrity metrics.

Conclusion: The proposed FabasedVC effectively integrates multiple modalities and duration modeling to achieve superior voice conversion performance across key evaluation dimensions.

Abstract: In voice conversion (VC), it is crucial to preserve complete semantic information while accurately modeling the target speaker’s timbre and prosody. This paper proposes FabasedVC to achieve VC with enhanced similarity in timbre, prosody, and duration to the target speaker, as well as improved content integrity. It is an end-to-end VITS-based VC system that integrates relevant textual modality information, phoneme-level self-supervised learning (SSL) features, and a duration predictor. Specifically, we employ a text feature encoder to encode attributes such as text, phonemes, tones and BERT features. We then process the frame-level SSL features into phoneme-level features using two methods: average pooling and attention mechanism based on each phoneme’s duration. Moreover, a duration predictor is incorporated to better align the speech rate and prosody of the target speaker. Experimental results demonstrate that our method outperforms competing systems in terms of naturalness, similarity, and content integrity.

Method: SACRED-Bench uses three speech-audio composition mechanisms: speech overlap/multi-speaker dialogue, speech-audio mixture, and diverse spoken instruction formats to evade text-only filters.

Result: Experiments show Gemini 2.5 Pro has 66% attack success rate. The proposed SALMONN-Guard defense reduces this to 20% by jointly inspecting speech, audio, and text.

Conclusion: Audio-aware defenses are crucial for multimodal LLM safety, as current safeguards fail against cross-modal speech-audio composition attacks.

Abstract: Recent progress in large language models (LLMs) has enabled understanding of both speech and non-speech audio, but exposing new safety risks emerging from complex audio inputs that are inadequately handled by current safeguards. We introduce SACRED-Bench (Speech-Audio Composition for RED-teaming) to evaluate the robustness of LLMs under complex audio-based attacks. Unlike existing perturbation-based methods that rely on noise optimization or white-box access, SACRED-Bench exploits speech-audio composition mechanisms. SACRED-Bench adopts three mechanisms: (a) speech overlap and multi-speaker dialogue, which embeds harmful prompts beneath or alongside benign speech; (b) speech-audio mixture, which imply unsafe intent via non-speech audio alongside benign speech or audio; and (c) diverse spoken instruction formats (open-ended QA, yes/no) that evade text-only filters. Experiments show that, even Gemini 2.5 Pro, the state-of-the-art proprietary LLM, still exhibits 66% attack success rate in SACRED-Bench test set, exposing vulnerabilities under cross-modal, speech-audio composition attacks. To bridge this gap, we propose SALMONN-Guard, a safeguard LLM that jointly inspects speech, audio, and text for safety judgments, reducing attack success down to 20%. Our results highlight the need for audio-aware defenses for the safety of multimodal LLMs. The benchmark and SALMONN-Guard checkpoints can be found at https://huggingface.co/datasets/tsinghua-ee/SACRED-Bench. Warning: this paper includes examples that may be offensive or harmful.

Method: Hybrid training framework with dual-stage masking (MaskedSpec and MaskedFeature) for complementary regularization, compression-aware strategy for low-resource scenarios, and unified pipeline combining self-supervised feature extractor with ResNet classification head.

Result: Achieved state-of-the-art results: 4.08% EER on ASVspoof5 Challenge (reduced to 2.71% with fusion), 0.18% EER on ASVspoof2019 evaluation set, and 2.92% EER on ASVspoof2021 DF task.

Conclusion: The proposed hybrid framework with novel augmentation strategies effectively advances deepfake speech detection performance, demonstrating strong generalization across different datasets and conditions.

Abstract: Deepfake speech detection presents a growing challenge as generative audio technologies continue to advance. We propose a hybrid training framework that advances detection performance through novel augmentation strategies. First, we introduce a dual-stage masking approach that operates both at the spectrogram level (MaskedSpec) and within the latent feature space (MaskedFeature), providing complementary regularization that improves tolerance to localized distortions and enhances generalization learning. Second, we introduce compression-aware strategy during self-supervised to increase variability in low-resource scenarios while preserving the integrity of learned representations, thereby improving the suitability of pretrained features for deepfake detection. The framework integrates a learnable self-supervised feature extractor with a ResNet classification head in a unified training pipeline, enabling joint adaptation of acoustic representations and discriminative patterns. On the ASVspoof5 Challenge (Track~1), the system achieves state-of-the-art results with an Equal Error Rate (EER) of 4.08% under closed conditions, further reduced to 2.71% through fusion of models with diverse pretrained feature extractors. when trained on ASVspoof2019, our system obtaining leading performance on the ASVspoof2019 evaluation set (0.18% EER) and the ASVspoof2021 DF task (2.92% EER).

Method: Created a daily-oriented text audio dataset with 9,300 recordings from 34 participants, enhanced using AI techniques including timbre synthesis, speed variation, and noise addition to increase complexity and scalability.

Result: Developed DOTA-ME-CS dataset with 18.54 hours of carefully curated audio data that ensures both diversity and quality for bilingual speech recognition research.

Conclusion: The dataset provides a robust resource for addressing code-switching ASR challenges and demonstrates potential for future research, with plans for public release of both dataset and code.

Abstract: Code-switching, the alternation between two or more languages within communication, poses great challenges for Automatic Speech Recognition (ASR) systems. Existing models and datasets are limited in their ability to effectively handle these challenges. To address this gap and foster progress in code-switching ASR research, we introduce the DOTA-ME-CS: Daily oriented text audio Mandarin-English code-switching dataset, which consists of 18.54 hours of audio data, including 9,300 recordings from 34 participants. To enhance the dataset’s diversity, we apply artificial intelligence (AI) techniques such as AI timbre synthesis, speed variation, and noise addition, thereby increasing the complexity and scalability of the task. The dataset is carefully curated to ensure both diversity and quality, providing a robust resource for researchers addressing the intricacies of bilingual speech recognition with detailed data analysis. We further demonstrate the dataset’s potential in future research. The DOTA-ME-CS dataset, along with accompanying code, will be made publicly available.

Method: Integrates Dasheng audio encoder with publicly available datasets, focusing on general audio captions that fuse speech, sound and music information into unified textual representations.

Result: Achieves up to 4x speedup in time-to-first-token and up to 20x higher throughput than comparable models, with fully transparent and reproducible results.

Conclusion: MiDashengLM demonstrates that open-source models using general audio captions can achieve efficient and comprehensive audio understanding while maintaining full transparency.

Abstract: Current approaches for large audio language models (LALMs) often rely on closed data sources or proprietary models, limiting their generalization and accessibility. This paper introduces MiDashengLM, a novel open audio-language model designed for efficient and comprehensive audio understanding through the use of general audio captions using our novel ACAVCaps training dataset. MiDashengLM exclusively relies on publicly available pretraining and supervised fine-tuning (SFT) datasets, ensuring full transparency and reproducibility. At its core, MiDashengLM integrates Dasheng, an open-source audio encoder, specifically engineered to process diverse auditory information effectively. Unlike previous works primarily focused on Automatic Speech Recognition (ASR) based audio-text alignment, our strategy centers on general audio captions, fusing speech, sound and music information into one textual representation, enabling a holistic textual representation of complex audio scenes. Lastly, MiDashengLM provides an up to 4x speedup in terms of time-to-first-token (TTFT) and up to 20x higher throughput than comparable models. Checkpoints are available online at https://huggingface.co/mispeech/midashenglm-7b and https://github.com/xiaomi-research/dasheng-lm.

Method: Proposed multi-view spiking temporal-aware self-attention (MSTASA) module combined with spiking contextual refinement channel MLP (SCR-MLP) in a fully spike-driven transformer architecture called SpikCommander.

Result: Outperforms state-of-the-art SNN approaches on Spiking Heidelberg Dataset, Spiking Speech Commands, and Google Speech Commands V2 datasets with fewer parameters under comparable time steps.

Conclusion: SpikCommander demonstrates effectiveness and efficiency for robust speech command recognition through enhanced temporal context modeling and channel-wise feature integration.

Abstract: Spiking neural networks (SNNs) offer a promising path toward energy-efficient speech command recognition (SCR) by leveraging their event-driven processing paradigm. However, existing SNN-based SCR methods often struggle to capture rich temporal dependencies and contextual information from speech due to limited temporal modeling and binary spike-based representations. To address these challenges, we first introduce the multi-view spiking temporal-aware self-attention (MSTASA) module, which combines effective spiking temporal-aware attention with a multi-view learning framework to model complementary temporal dependencies in speech commands. Building on MSTASA, we further propose SpikCommander, a fully spike-driven transformer architecture that integrates MSTASA with a spiking contextual refinement channel MLP (SCR-MLP) to jointly enhance temporal context modeling and channel-wise feature integration. We evaluate our method on three benchmark datasets: the Spiking Heidelberg Dataset (SHD), the Spiking Speech Commands (SSC), and the Google Speech Commands V2 (GSC). Extensive experiments demonstrate that SpikCommander consistently outperforms state-of-the-art (SOTA) SNN approaches with fewer parameters under comparable time steps, highlighting its effectiveness and efficiency for robust speech command recognition.

Method: Constructs directed bipartite graph with common and rare code nodes, using probability-based bias from co-occurrence patterns in attention module. Uses LLM for code descriptions to enrich embeddings.

Result: Achieves state-of-the-art performance on three benchmark datasets with notable improvements in Macro-F1 for long-tail classification.

Conclusion: The method effectively addresses long-tail ICD coding by leveraging co-occurrence relationships and LLM-generated descriptions to enhance rare code representations.

Abstract: Automated International Classification of Diseases (ICD) coding aims to assign multiple disease codes to clinical documents, constituting a crucial multi-label text classification task in healthcare informatics. However, the task is challenging due to its large label space (10,000 to 20,000 codes) and long-tail distribution, where a few codes dominate while many rare codes lack sufficient training data. To address this, we propose a learning method that models fine-grained co-occurrence relationships among codes. Specifically, we construct a Directed Bipartite Graph Encoder with disjoint sets of common and rare code nodes. To facilitate a one-way information flow, edges are directed exclusively from common to rare codes. The nature of these connections is defined by a probability-based bias, which is derived from the conditional probability of a common code co-occurring given the presence of a rare code. This bias is then injected into the encoder’s attention module, a process we term Co-occurrence Encoding. This structure empowers the graph encoder to enrich rare code representations by aggregating latent comorbidity information reflected in the statistical co-occurrence of their common counterparts. To ensure high-quality input to the graph, we utilize a large language model (LLM) to generate comprehensive descriptions for codes, enriching initial embeddings with clinical context and comorbidity information, serving as external knowledge for the statistical co-occurrence relationships in the code system. Experiments on three automated ICD coding benchmark datasets demonstrate that our method achieves state-of-the-art performance with particularly notable improvements in Macro-F1, which is the key metric for long-tail classification.

Method: Hybrid GNN and Temporal Transformer framework integrating multi-source data (social media, marketing expenditure, engagement logs, sentiment) with dual-channel architecture for diffusion structure and temporal influence learning, plus causal inference modules.

Result: Outperforms existing baselines in all six metrics on large-scale real-world datasets from Twitter, TikTok, and YouTube advertising platforms.

Conclusion: The DSS provides interpretable real-time insights into AIGC-driven content dissemination and market growth patterns, enhancing marketing decision-making.

Abstract: The rapid proliferation of AI-generated content (AIGC) has reshaped the dynamics of digital marketing and online consumer behavior. However, predicting the diffusion trajectory and market impact of such content remains challenging due to data heterogeneity, non linear propagation mechanisms, and evolving consumer interactions. This study proposes an AI driven Decision Support System (DSS) that integrates multi source data including social media streams, marketing expenditure records, consumer engagement logs, and sentiment dynamics using a hybrid Graph Neural Network (GNN) and Temporal Transformer framework. The model jointly learns the content diffusion structure and temporal influence evolution through a dual channel architecture, while causal inference modules disentangle the effects of marketing stimuli on return on investment (ROI) and market visibility. Experiments on large scale real-world datasets collected from multiple online platforms such as Twitter, TikTok, and YouTube advertising show that our system outperforms existing baselines in all six metrics. The proposed DSS enhances marketing decisions by providing interpretable real-time insights into AIGC driven content dissemination and market growth patterns.

Method: Developed several discrete-time deep mixture-of-experts architectures for survival analysis, with varying levels of expert expressiveness, including models with more expressive experts that tailor predictions per patient rather than relying on fixed group prototypes.

Result: Found that more expressive experts that customize predictions for individual patients outperform experts using fixed group prototypes, achieving all three desiderata: clustering, calibration, and predictive accuracy.

Conclusion: The key differentiator between mixture-of-experts architectures is expert expressiveness, with patient-tailored predictions enabling simultaneous achievement of clustering, calibration, and accuracy goals in survival analysis.

Abstract: Deep mixture-of-experts models have attracted a lot of attention for survival analysis problems, particularly for their ability to cluster similar patients together. In practice, grouping often comes at the expense of key metrics such calibration error and predictive accuracy. This is due to the restrictive inductive bias that mixture-of-experts imposes, that predictions for individual patients must look like predictions for the group they’re assigned to. Might we be able to discover patient group structure, where it exists, while improving calibration and predictive accuracy? In this work, we introduce several discrete-time deep mixture-of-experts (MoE) based architectures for survival analysis problems, one of which achieves all desiderata: clustering, calibration, and predictive accuracy. We show that a key differentiator between this array of MoEs is how expressive their experts are. We find that more expressive experts that tailor predictions per patient outperform experts that rely on fixed group prototypes.

Method: Hybrid architecture with two RNNs: one for mode prediction and another for filtering based on mode-augmented KalmanNet, trained jointly using alternating least squares strategy without mode supervision.

Result: Outperforms classical model-based filters (interacting multiple models, particle filters) and model-free deep learning baselines, especially in non-stationary and high-noise regimes. Shows improvement over KalmanNet, more pronounced in complex systems or long trajectories.

Conclusion: JMFNet provides consistent and reliable performance with low sensitivity to initial conditions, hyperparameters, and incorrect model knowledge, making it suitable for practical applications in jump Markov systems.

Abstract: This paper presents the Jump Markov Filtering Network (JMFNet), a novel model-based deep learning framework for real-time state-state estimation in jump Markov systems with unknown noise statistics and mode transition dynamics. A hybrid architecture comprising two Recurrent Neural Networks (RNNs) is proposed: one for mode prediction and another for filtering that is based on a mode-augmented version of the recently presented KalmanNet architecture. The proposed RNNs are trained jointly using an alternating least squares strategy that enables mutual adaptation without supervision of the latent modes. Extensive numerical experiments on linear and nonlinear systems, including target tracking, pendulum angle tracking, Lorenz attractor dynamics, and a real-life dataset demonstrate that the proposed JMFNet framework outperforms classical model-based filters (e.g., interacting multiple models and particle filters) as well as model-free deep learning baselines, particularly in non-stationary and high-noise regimes. It is also showcased that JMFNet achieves a small yet meaningful improvement over the KalmanNet framework, which becomes much more pronounced in complicated systems or long trajectories. Finally, the method’s performance is empirically validated to be consistent and reliable, exhibiting low sensitivity to initial conditions, hyperparameter selection, as well as to incorrect model knowledge

Method: Dynamical systems analysis modeling learning as continuous-time gradient flow, proving unstable equilibria for IGM-violating solutions and stable attractors for IGM-consistent ones.

Result: Unconstrained non-monotonic factorization reliably recovers IGM-optimal solutions and consistently outperforms monotonic baselines in both synthetic games and MARL benchmarks.

Conclusion: Non-monotonic value decomposition naturally achieves IGM consistency through learning dynamics, providing insights for future value-based MARL algorithm design.

Abstract: Value decomposition is a central approach in multi-agent reinforcement learning (MARL), enabling centralized training with decentralized execution by factorizing the global value function into local values. To ensure individual-global-max (IGM) consistency, existing methods either enforce monotonicity constraints, which limit expressive power, or adopt softer surrogates at the cost of algorithmic complexity. In this work, we present a dynamical systems analysis of non-monotonic value decomposition, modeling learning dynamics as continuous-time gradient flow. We prove that, under approximately greedy exploration, all zero-loss equilibria violating IGM consistency are unstable saddle points, while only IGM-consistent solutions are stable attractors of the learning dynamics. Extensive experiments on both synthetic matrix games and challenging MARL benchmarks demonstrate that unconstrained, non-monotonic factorization reliably recovers IGM-optimal solutions and consistently outperforms monotonic baselines. Additionally, we investigate the influence of temporal-difference targets and exploration strategies, providing actionable insights for the design of future value-based MARL algorithms.

Method: Apply group averaging at evaluation time over small symmetry groups to well-established benchmark ML models of differential equations, without modifying model structure or training.

Result: Group averaging always decreased average evaluation loss, with improvements up to 37% in VRMSE. Produced visually better predictions for continuous dynamics.

Conclusion: Group averaging is a cheap, simple way to improve model accuracy with no disadvantages under common circumstances; ML4PS community should adopt this approach.

Abstract: Many machine learning tasks in the natural sciences are precisely equivariant to particular symmetries. Nonetheless, equivariant methods are often not employed, perhaps because training is perceived to be challenging, or the symmetry is expected to be learned, or equivariant implementations are seen as hard to build. Group averaging is an available technique for these situations. It happens at test time; it can make any trained model precisely equivariant at a (often small) cost proportional to the size of the group; it places no requirements on model structure or training. It is known that, under mild conditions, the group-averaged model will have a provably better prediction accuracy than the original model. Here we show that an inexpensive group averaging can improve accuracy in practice. We take well-established benchmark machine learning models of differential equations in which certain symmetries ought to be obeyed. At evaluation time, we average the models over a small group of symmetries. Our experiments show that this procedure always decreases the average evaluation loss, with improvements of up to 37% in terms of the VRMSE. The averaging produces visually better predictions for continuous dynamics. This short paper shows that, under certain common circumstances, there are no disadvantages to imposing exact symmetries; the ML4PS community should consider group averaging as a cheap and simple way to improve model accuracy.

Method: Uses a guided denoising diffusion probabilistic model (DDPM) with surrogate gradients from trained neural networks predicting pressure drop and temperature to guide geometry generation.

Result: Generated heat sinks achieve up to 10% lower pressure drops compared to traditional CMA-ES optimization while maintaining temperature constraints.

Conclusion: The method provides a computationally efficient foundation for generative electronics cooling design that doesn’t require retraining for new constraints.

Abstract: This study presents a generative optimization framework based on a guided denoising diffusion probabilistic model (DDPM) that leverages surrogate gradients to generate heat sink designs minimizing pressure drop while maintaining surface temperatures below a specified threshold. Geometries are represented using boundary representations of multiple fins, and a multi-fidelity approach is employed to generate training data. Using this dataset, along with vectors representing the boundary representation geometries, we train a denoising diffusion probabilistic model to generate heat sinks with characteristics consistent with those observed in the data. We train two different residual neural networks to predict the pressure drop and surface temperature for each geometry. We use the gradients of these surrogate models with respect to the design variables to guide the geometry generation process toward satisfying the low-pressure and surface temperature constraints. This inference-time guidance directs the generative process toward heat sink designs that not only prevent overheating but also achieve lower pressure drops compared to traditional optimization methods such as CMA-ES. In contrast to traditional black-box optimization approaches, our method is scalable, provided sufficient training data is available. Unlike traditional topology optimization methods, once the model is trained and the heat sink world model is saved, inference under new constraints (e.g., temperature) is computationally inexpensive and does not require retraining. Samples generated using the guided diffusion model achieve pressure drops up to 10 percent lower than the limits obtained by traditional black-box optimization methods. This work represents a step toward building a foundational generative model for electronics cooling.

Method: Proposes the Generation-Execution-Feedback (GEF) loop framework and systematically reviews environment scaling methods across task generation, task execution, and feedback stages.

Result: Organizes fragmented advances in environment scaling, analyzes benchmarks and implementation strategies, and provides a comprehensive survey from an environment-centric perspective.

Conclusion: Environment scaling toward greater complexity, realism, and interactivity is crucial for advancing agent intelligence through experiential learning in the GEF loop paradigm.

Abstract: LLM-based agents can autonomously accomplish complex tasks across various domains. However, to further cultivate capabilities such as adaptive behavior and long-term decision-making, training on static datasets built from human-level knowledge is insufficient. These datasets are costly to construct and lack both dynamism and realism. A growing consensus is that agents should instead interact directly with environments and learn from experience through reinforcement learning. We formalize this iterative process as the Generation-Execution-Feedback (GEF) loop, where environments generate tasks to challenge agents, return observations in response to agents’ actions during task execution, and provide evaluative feedback on rollouts for subsequent learning. Under this paradigm, environments function as indispensable producers of experiential data, highlighting the need to scale them toward greater complexity, realism, and interactivity. In this survey, we systematically review representative methods for environment scaling from a pioneering environment-centric perspective and organize them along the stages of the GEF loop, namely task generation, task execution, and feedback. We further analyze benchmarks, implementation strategies, and applications, consolidating fragmented advances and outlining future research directions for agent intelligence.

Method: DR-GNN uses an operator-level Control Data Flow Graph (CDFG) to represent RTL circuits, enabling capture of dynamic dependencies and runtime execution. It was trained on a comprehensive dataset of 6,300 Verilog designs and 63,000 simulation traces.

Result: DR-GNN outperforms existing models in branch hit prediction and toggle rate prediction. Its learned representations also transfer effectively to power estimation and assertion prediction tasks.

Conclusion: DR-GNN successfully addresses the limitation of static-only circuit representations by incorporating dynamic execution behavior, demonstrating superior performance across multiple circuit analysis tasks.

Abstract: There is a growing body of work on using Graph Neural Networks (GNNs) to learn representations of circuits, focusing primarily on their static characteristics. However, these models fail to capture circuit runtime behavior, which is crucial for tasks like circuit verification and optimization. To address this limitation, we introduce DR-GNN (DynamicRTL-GNN), a novel approach that learns RTL circuit representations by incorporating both static structures and multi-cycle execution behaviors. DR-GNN leverages an operator-level Control Data Flow Graph (CDFG) to represent Register Transfer Level (RTL) circuits, enabling the model to capture dynamic dependencies and runtime execution. To train and evaluate DR-GNN, we build the first comprehensive dynamic circuit dataset, comprising over 6,300 Verilog designs and 63,000 simulation traces. Our results demonstrate that DR-GNN outperforms existing models in branch hit prediction and toggle rate prediction. Furthermore, its learned representations transfer effectively to related dynamic circuit tasks, achieving strong performance in power estimation and assertion prediction.

Method: Formulates personalization as a Constrained Markov Decision Process (CMDP) with multi-objective reward function balancing engagement and well-being, using emotion-informed state representation and safety-constrained RL algorithms.

Result: Conceptual framework that operationalizes empathy and responsibility in machine learning policy optimization, bridging safe RL, affective computing and responsible AI.

Conclusion: Initiates methodological conversation about ethically aligned reinforcement learning for emotionally aware and trustworthy personalization systems in human-centric domains.

Abstract: Personalized decision systems in healthcare and behavioral support often rely on static rule-based or engagement-maximizing heuristics that overlook users’ emotional context and ethical constraints. Such approaches risk recommending insensitive or unsafe interventions, especially in domains involving serious mental illness, substance use disorders, or depression. To address this limitation, we propose a Responsible Reinforcement Learning (RRL) framework that integrates emotional and contextual understanding with ethical considerations into the sequential decision-making process. RRL formulates personalization as a Constrained Markov Decision Process (CMDP), where the agent optimizes engagement and adherence while ensuring emotional alignment and ethical safety. We introduce a multi-objective reward function that explicitly balances short-term behavioral engagement with long-term user well-being, and define an emotion-informed state representation that captures fluctuations in emotional readiness, affect, and risk. The proposed architecture can be instantiated with any RL algorithm (e.g., DQN, PPO) augmented with safety constraints or Lagrangian regularization. Conceptually, this framework operationalizes empathy and responsibility within machine learning policy optimization, bridging safe RL, affective computing and responsible AI. We discuss the implications of this approach for human-centric domains such as behavioral health, education, and digital therapeutics, and outline simulation-based validation paths for future empirical work. This paper aims to initiate a methodological conversation about ethically aligned reinforcement learning for emotionally aware and trustworthy personalization systems.

Method: Partitions total attention workload into balanced, non-overlapping distance bands and assigns each head a unique segment, transforming H independent O(N²) computations into a single collaborative O(N²) computation with structured inductive bias for functional specialization.

Result: Achieves ~2x training throughput increase while performing on par with standard dense attention (even surpassing on select metrics) and consistently outperforming Longformer, Reformer, and BigBird across all evaluation metrics on OLMoE-1B-7B and 0.25B-1.75B models.

Conclusion: SPAttention resolves the efficiency-performance trade-off through principled structural sparsity, fundamentally reducing attention complexity while maintaining or improving performance compared to both dense and existing sparse attention methods.

Abstract: The design of Large Language Models (LLMs) has long been hampered by a fundamental conflict within their core attention mechanism: its remarkable expressivity is built upon a computational complexity of $O(H \cdot N^2)$ that grows quadratically with the context size ($N$) and linearly with the number of heads ($H$). This standard implementation harbors significant computational redundancy, as all heads independently compute attention over the same sequence space. Existing sparse methods, meanwhile, often trade information integrity for computational efficiency. To resolve this efficiency-performance trade-off, we propose SPAttention, whose core contribution is the introduction of a new paradigm we term Principled Structural Sparsity. SPAttention does not merely drop connections but instead reorganizes the computational task by partitioning the total attention workload into balanced, non-overlapping distance bands, assigning each head a unique segment. This approach transforms the multi-head attention mechanism from $H$ independent $O(N^2)$ computations into a single, collaborative $O(N^2)$ computation, fundamentally reducing complexity by a factor of $H$. The structured inductive bias compels functional specialization among heads, enabling a more efficient allocation of computational resources from redundant modeling to distinct dependencies across the entire sequence span. Extensive empirical validation on the OLMoE-1B-7B and 0.25B-1.75B model series demonstrates that while delivering an approximately two-fold increase in training throughput, its performance is on par with standard dense attention, even surpassing it on select key metrics, while consistently outperforming representative sparse attention methods including Longformer, Reformer, and BigBird across all evaluation metrics.

Method: Alternating framework with (1) acquisition-driven search leveraging inter-task synergies using task-aware Gaussian processes, and (2) generative solution sampling via conditional generative models.

Result: Achieves direct solution prediction for unseen parameterized EMOPs without additional expensive evaluations, with theoretical justification for faster convergence and empirical verification on synthetic and real-world benchmarks.

Conclusion: The proposed alternating framework effectively learns inverse models for parametric expensive multi-objective optimization problems, enabling efficient optimization across related tasks and direct solution prediction.

Abstract: Many real-world applications require solving families of expensive multi-objective optimization problems~(EMOPs) under varying operational conditions. This gives rise to parametric expensive multi-objective optimization problems (P-EMOPs) where each task parameter defines a distinct optimization instance. Current multi-objective Bayesian optimization methods have been widely used for finding finite sets of Pareto optimal solutions for individual tasks. However, P-EMOPs present a fundamental challenge: the continuous task parameter space can contain infinite distinct problems, each requiring separate expensive evaluations. This demands learning an inverse model that can directly predict optimized solutions for any task-preference query without expensive re-evaluation. This paper introduces the first parametric multi-objective Bayesian optimizer that learns this inverse model by alternating between (1) acquisition-driven search leveraging inter-task synergies and (2) generative solution sampling via conditional generative models. This approach enables efficient optimization across related tasks and finally achieves direct solution prediction for unseen parameterized EMOPs without additional expensive evaluations. We theoretically justify the faster convergence by leveraging inter-task synergies through task-aware Gaussian processes. Meanwhile, empirical studies in synthetic and real-world benchmarks further verify the effectiveness of our alternating framework.

Method: Analyzed regret bounds for multiple agents in the same Markov decision process with potentially different reward functions, comparing scenarios where agents share observations versus learning individually.

Result: When agents share observations, the total regret of all agents is reduced by a factor of √N compared to when each agent relies only on their own collected information.

Conclusion: Regret analysis in multi-agent settings provides theoretical bounds demonstrating significant benefits of observation sharing in transfer learning, with √N improvement in total regret.

Abstract: We present an approach for the quantification of the usefulness of transfer in reinforcement learning via regret bounds for a multi-agent setting. Considering a number of $\aleph$ agents operating in the same Markov decision process, however possibly with different reward functions, we consider the regret each agent suffers with respect to an optimal policy maximizing her average reward. We show that when the agents share their observations the total regret of all agents is smaller by a factor of $\sqrt{\aleph}$ compared to the case when each agent has to rely on the information collected by herself. This result demonstrates how considering the regret in multi-agent settings can provide theoretical bounds on the benefit of sharing observations in transfer learning.

Method: UCB-QRL is an iterative algorithm that estimates transition probabilities and optimizes quantile value functions over confidence balls around these estimates.

Result: The algorithm achieves a high-probability regret bound of O((2/κ)^{H+1}H√(SATH log(2SATH/δ))) for episodic MDPs with S states, A actions, T episodes, and H horizons.

Conclusion: UCB-QRL provides a principled approach for risk-sensitive RL with quantile objectives and establishes theoretical guarantees for its performance.

Abstract: Reinforcement Learning (RL) has achieved tremendous success in recent years. However, the classical foundations of RL do not account for the risk sensitivity of the objective function, which is critical in various fields, including healthcare and finance. A popular approach to incorporate risk sensitivity is to optimize a specific quantile of the cumulative reward distribution. In this paper, we develop UCB-QRL, an optimistic learning algorithm for the $τ$-quantile objective in finite-horizon Markov decision processes (MDPs). UCB-QRL is an iterative algorithm in which, at each iteration, we first estimate the underlying transition probability and then optimize the quantile value function over a confidence ball around this estimate. We show that UCB-QRL yields a high-probability regret bound $\mathcal O\left((2/κ)^{H+1}H\sqrt{SATH\log(2SATH/δ)}\right)$ in the episodic setting with $S$ states, $A$ actions, $T$ episodes, and $H$ horizons. Here, $κ>0$ is a problem-dependent constant that captures the sensitivity of the underlying MDP’s quantile value.

Method: Systematic self-supervised pre-training on single real tables and evaluation across diverse benchmarks, analyzing what data aspects are crucial for building generalizable tabular foundation models.

Result: Simple self-supervised pre-training on just one real table produces surprisingly strong transfer across heterogeneous benchmarks, with the number and quality of tasks constructed from a dataset being key to downstream performance.

Conclusion: The number and quality of tasks that can be constructed from a dataset, rather than dataset size alone, is the key factor for building effective tabular foundation models that generalize across domains.

Abstract: Deep tabular modelling increasingly relies on in-context learning where, during inference, a model receives a set of $(x,y)$ pairs as context and predicts labels for new inputs without weight updates. We challenge the prevailing view that broad generalization here requires pre-training on large synthetic corpora (e.g., TabPFN priors) or a large collection of real data (e.g., TabDPT training datasets), discovering that a relatively small amount of data suffices for generalization. We find that simple self-supervised pre-training on just a \emph{single} real table can produce surprisingly strong transfer across heterogeneous benchmarks. By systematically pre-training and evaluating on many diverse datasets, we analyze what aspects of the data are most important for building a Tabular Foundation Model (TFM) generalizing across domains. We then connect this to the pre-training procedure shared by most TFMs and show that the number and quality of \emph{tasks} one can construct from a dataset is key to downstream performance.

Method: GEM+ integrates AIM’s adaptive ‘select-measure-generate’ framework with GEM’s scalable generator neural network, combining systematic optimization under privacy constraints with improved computational efficiency.

Result: GEM+ outperforms AIM in both utility and scalability, achieving state-of-the-art results while efficiently handling datasets with over a hundred columns where AIM fails due to memory and computational limitations.

Conclusion: GEM+ provides a superior solution for differentially private synthetic tabular data generation that combines the strengths of both adaptive measurement frameworks and scalable neural networks, enabling practical application to real-world high-dimensional datasets.

Abstract: State-of-the-art differentially private synthetic tabular data has been defined by adaptive ‘select-measure-generate’ frameworks, exemplified by methods like AIM. These approaches iteratively measure low-order noisy marginals and fit graphical models to produce synthetic data, enabling systematic optimisation of data quality under privacy constraints. Graphical models, however, are inefficient for high-dimensional data because they require substantial memory and must be retrained from scratch whenever the graph structure changes, leading to significant computational overhead. Recent methods, like GEM, overcome these limitations by using generator neural networks for improved scalability. However, empirical comparisons have mostly focused on small datasets, limiting real-world applicability. In this work, we introduce GEM+, which integrates AIM’s adaptive measurement framework with GEM’s scalable generator network. Our experiments show that GEM+ outperforms AIM in both utility and scalability, delivering state-of-the-art results while efficiently handling datasets with over a hundred columns, where AIM fails due to memory and computational overheads.

Method: Sequentially trains an ensemble of GFlowNets, each optimizing a residual reward that compensates for mass captured by previous models, reactivating learning signals in underexplored regions.

Result: Achieves substantially better exploration and sample diversity on multimodal synthetic benchmarks and peptide design tasks while maintaining training stability and simplicity.

Conclusion: Boosted GFlowNets provide monotone non-degradation guarantees and significantly improve coverage of high-reward regions without compromising the stability of standard GFlowNet training.

Abstract: Generative Flow Networks (GFlowNets) are powerful samplers for compositional objects that, by design, sample proportionally to a given non-negative reward. Nonetheless, in practice, they often struggle to explore the reward landscape evenly: trajectories toward easy-to-reach regions dominate training, while hard-to-reach modes receive vanishing or uninformative gradients, leading to poor coverage of high-reward areas. We address this imbalance with Boosted GFlowNets, a method that sequentially trains an ensemble of GFlowNets, each optimizing a residual reward that compensates for the mass already captured by previous models. This residual principle reactivates learning signals in underexplored regions and, under mild assumptions, ensures a monotone non-degradation property: adding boosters cannot worsen the learned distribution and typically improves it. Empirically, Boosted GFlowNets achieve substantially better exploration and sample diversity on multimodal synthetic benchmarks and peptide design tasks, while preserving the stability and simplicity of standard trajectory-balance training.

Method: Integrates shadow Q model for reward estimation, GAN for imperceptible perturbations, and world model for environment simulation. Uses two-stage iterative training alternating between shadow model learning and generator refinement.

Result: Significantly reduces cumulative rewards on MuJoCo and Atari benchmarks while preserving visual fidelity and greatly decreasing environment interactions compared to prior methods.

Conclusion: SEBA provides an effective and efficient black-box adversarial attack framework for visual RL that addresses the limitations of existing approaches in continuous control settings.

Abstract: Visual reinforcement learning has achieved remarkable progress in visual control and robotics, but its vulnerability to adversarial perturbations remains underexplored. Most existing black-box attacks focus on vector-based or discrete-action RL, and their effectiveness on image-based continuous control is limited by the large action space and excessive environment queries. We propose SEBA, a sample-efficient framework for black-box adversarial attacks on visual RL agents. SEBA integrates a shadow Q model that estimates cumulative rewards under adversarial conditions, a generative adversarial network that produces visually imperceptible perturbations, and a world model that simulates environment dynamics to reduce real-world queries. Through a two-stage iterative training procedure that alternates between learning the shadow model and refining the generator, SEBA achieves strong attack performance while maintaining efficiency. Experiments on MuJoCo and Atari benchmarks show that SEBA significantly reduces cumulative rewards, preserves visual fidelity, and greatly decreases environment interactions compared to prior black-box and white-box methods.

Method: Constrained RL framework with natural and interpretable reward/constraint signals, dynamically balancing trade-offs among them during training.

Result: Theoretical guarantees established and numerical experiments on Text2SQL datasets show improvement over state-of-the-art RL-trained LLMs.

Conclusion: The constrained RL framework effectively addresses reward sensitivity and hacking issues in Text2SQL, providing better performance than existing RL approaches.

Abstract: Reinforcement learning (RL) has demonstrated significant promise in enhancing the reasoning capabilities of Text2SQL LLMs, especially with advanced algorithms such as GRPO and DAPO. However, the performance of these methods is highly sensitive to the design of reward functions. Inappropriate rewards can lead to reward hacking, where models exploit loopholes in the reward structure to achieve high scores without genuinely solving the task. This work considers a constrained RL framework for Text2SQL that incorporates natural and interpretable reward and constraint signals, while dynamically balancing trade-offs among them during the training. We establish the theoretical guarantees of our constrained RL framework and our numerical experiments on the well-known Text2SQL datasets substantiate the improvement of our approach over the state-of-the-art RL-trained LLMs.

Method: Conducted extensive evaluation including capacity measurements on 123 datasets, designed the first dedicated hardware accelerator for MCR, and compared performance against binary spatter codes and other representations.

Result: MCR outperforms binary/integer vectors in capacity, matches binary spatter code performance with 4x less memory, achieves 3.08x faster execution and 2.68x lower energy consumption when accuracy-matched against binary spatter codes.

Conclusion: MCR provides a unique balance of capacity, accuracy, and hardware efficiency. Its modular arithmetic and higher precision enable lower dimensionality, making it a faster, more energy-efficient alternative to existing models when implemented with dedicated hardware.

Abstract: The modular composite representation (MCR) is a computing model that represents information with high-dimensional integer vectors using modular arithmetic. Originally proposed as a generalization of the binary spatter code model, it aims to provide higher representational power while remaining a lighter alternative to models requiring high-precision components. Despite this potential, MCR has received limited attention. Systematic analyses of its trade-offs and comparisons with other models are lacking, sustaining the perception that its added complexity outweighs the improved expressivity. In this work, we revisit MCR by presenting its first extensive evaluation, demonstrating that it achieves a unique balance of capacity, accuracy, and hardware efficiency. Experiments measuring capacity demonstrate that MCR outperforms binary and integer vectors while approaching complex-valued representations at a fraction of their memory footprint. Evaluation on 123 datasets confirms consistent accuracy gains and shows that MCR can match the performance of binary spatter codes using up to 4x less memory. We investigate the hardware realization of MCR by showing that it maps naturally to digital logic and by designing the first dedicated accelerator. Evaluations on basic operations and 7 selected datasets demonstrate a speedup of up to 3 orders of magnitude and significant energy reductions compared to software implementation. When matched for accuracy against binary spatter codes, MCR achieves on average 3.08x faster execution and 2.68x lower energy consumption. These findings demonstrate that, although MCR requires more sophisticated operations than binary spatter codes, its modular arithmetic and higher per-component precision enable lower dimensionality. When realized with dedicated hardware, this results in a faster, more energy-efficient, and high-precision alternative to existing models.

Method: Condition neural models on vector encodings representing PDE terms and coefficients, trained on a family of 1D PDEs from APEBench suite using four distinct modeling techniques.

Result: Achieves strong performance on parameter sets held out from training distribution, maintains stability for rollout beyond training window, and generalizes to entirely unseen PDEs.

Conclusion: The framework enables generalization across PDE families and is part of broader effort to automate creation of expert-level empirical software for scientific tasks.

Abstract: Solving partial differential equations (PDEs) can be prohibitively expensive using traditional numerical methods. Deep learning-based surrogate models typically specialize in a single PDE with fixed parameters. We present a framework for equation-aware emulation that generalizes to unseen PDEs, conditioning a neural model on a vector encoding representing the terms in a PDE and their coefficients. We present a baseline of four distinct modeling technqiues, trained on a family of 1D PDEs from the APEBench suite. Our approach achieves strong performance on parameter sets held out from the training distribution, with strong stability for rollout beyond the training window, and generalization to an entirely unseen PDE. This work was developed as part of a broader effort exploring AI systems that automate the creation of expert-level empirical software for scorable scientific tasks. The data and codebase are available at https://github.com/google-research/generalized-pde-emulator.

Method: Uses Conditional Flow Matching (CFM) to learn a direct noise-to-data mapping instead of iterative sampling, enabling rapid high-fidelity sample generation with fewer function evaluations.

Result: FlowCast establishes new state-of-the-art in predictive accuracy and is significantly more efficient than diffusion models, maintaining high performance with fewer sampling steps.

Conclusion: CFM is positioned as a powerful and practical alternative to diffusion models for high-dimensional spatiotemporal forecasting tasks.

Abstract: Radar-based precipitation nowcasting, the task of forecasting short-term precipitation fields from previous radar images, is a critical problem for flood risk management and decision-making. While deep learning has substantially advanced this field, two challenges remain fundamental: the uncertainty of atmospheric dynamics and the efficient modeling of high-dimensional data. Diffusion models have shown strong promise by producing sharp, reliable forecasts, but their iterative sampling process is computationally prohibitive for time-critical applications. We introduce FlowCast, the first model to apply Conditional Flow Matching (CFM) to precipitation nowcasting. Unlike diffusion, CFM learns a direct noise-to-data mapping, enabling rapid, high-fidelity sample generation with drastically fewer function evaluations. Our experiments demonstrate that FlowCast establishes a new state-of-the-art in predictive accuracy. A direct comparison further reveals the CFM objective is both more accurate and significantly more efficient than a diffusion objective on the same architecture, maintaining high performance with significantly fewer sampling steps. This work positions CFM as a powerful and practical alternative for high-dimensional spatiotemporal forecasting.

Method: Propose Hydra, a novel mitigation method inspired by multi-head neural networks, adapted for SFL setting to address sequence-dependent forgetting.

Result: Extensive evaluations show Hydra outperforms baseline methods and existing approaches from literature in handling catastrophic forgetting in SFL.

Conclusion: Hydra effectively mitigates catastrophic forgetting in SFL, addressing both local update drift and server-side sequence sensitivity through multi-head neural network adaptation.

Abstract: In Split Federated Learning (SFL), the clients collaboratively train a model with the help of a server by splitting the model into two parts. Part-1 is trained locally at each client and aggregated by the aggregator at the end of each round. Part-2 is trained at a server that sequentially processes the intermediate activations received from each client. We study the phenomenon of catastrophic forgetting (CF) in SFL in the presence of data heterogeneity. In detail, due to the nature of SFL, local updates of part-1 may drift away from global optima, while part-2 is sensitive to the processing sequence, similar to forgetting in continual learning (CL). Specifically, we observe that the trained model performs better in classes (labels) seen at the end of the sequence. We investigate this phenomenon with emphasis on key aspects of SFL, such as the processing order at the server and the cut layer. Based on our findings, we propose Hydra, a novel mitigation method inspired by multi-head neural networks and adapted for the SFL’s setting. Extensive numerical evaluations show that Hydra outperforms baselines and methods from the literature.

Method: Simplifies previous two-phase approaches by introducing single-phase adaptation (SPARC) that trains context encoder and history adaptation module together, tested on Gran Turismo 7 racing simulator and wind-perturbed MuJoCo environments.

Result: SPARC achieves reliable and robust out-of-distribution generalization across varying contexts in high-fidelity simulations.

Conclusion: Single-phase adaptation provides an effective and simplified approach for robust control in contextual reinforcement learning, enabling better generalization to unseen environments.

Abstract: Generalization to unseen environments is a significant challenge in the field of robotics and control. In this work, we focus on contextual reinforcement learning, where agents act within environments with varying contexts, such as self-driving cars or quadrupedal robots that need to operate in different terrains or weather conditions than they were trained for. We tackle the critical task of generalizing to out-of-distribution (OOD) settings, without access to explicit context information at test time. Recent work has addressed this problem by training a context encoder and a history adaptation module in separate stages. While promising, this two-phase approach is cumbersome to implement and train. We simplify the methodology and introduce SPARC: single-phase adaptation for robust control. We test SPARC on varying contexts within the high-fidelity racing simulator Gran Turismo 7 and wind-perturbed MuJoCo environments, and find that it achieves reliable and robust OOD generalization.

Method: TawPipe groups devices based on topology to optimize intra-node collective and inter-node P2P communication, assigns each device a fixed shard of model weights and gradients to avoid redundant transfers, and overlaps communication with computation to hide latency.

Result: Experiments on up to 24 GPUs with LLaMA-style models show TawPipe achieves superior throughput and scalability compared to state-of-the-art baselines, with significantly reduced cross-node traffic.

Conclusion: TawPipe effectively addresses communication bottlenecks in distributed LLM training by exploiting hierarchical bandwidth and topology awareness, delivering better performance than existing pipeline parallelism approaches.

Abstract: Training large language models (LLMs) is fundamentally constrained by limited device memory and costly inter-device communication. Although pipeline parallelism alleviates memory pressure by partitioning models across devices, it incurs activation communication overhead that scales linearly with sequence length, limiting efficiency in long-context training. Recent weight-passing approaches (e.g., WeiPipe) mitigate this by transmitting model weights instead of activations, but suffer from redundant peer-to-peer (P2P) transfers and underutilized intra-node bandwidth. We propose TawPipe–topology-aware weight pipeline parallelism, which exploits hierarchical bandwidth in distributed clusters for improved communication efficiency. TawPipe: (i) groups devices based on topology to optimize intra-node collective and inter-node P2P communication; (ii) assigns each device a fixed shard of model weights and gradients, avoiding redundant transfers; and (iii) overlaps communication with computation to hide latency. Unlike global collective operations used in fully sharded data parallelism (FSDP), TawPipe confines most communication within node boundaries, significantly reducing cross-node traffic. Extensive experiments on up to 24 GPUs with LLaMA-style models show that TawPipe achieves superior throughput and scalability compared to state-of-the-art baselines.

Method: Introduces macro-contextual retrieval - a framework that jointly embeds macro indicators (CPI, unemployment, yield spread, GDP growth) and financial news sentiment in a shared similarity space, enabling retrieval of historically analogous macroeconomic regimes for each prediction without retraining.

Result: Achieved positive out-of-sample trading outcomes (AAPL: PF=1.18, Sharpe=0.95; XOM: PF=1.16, Sharpe=0.61) and consistently narrowed the CV to OOD performance gap, while static baselines collapsed under regime shifts. Retrieved neighbors form interpretable evidence chains corresponding to recognizable macro contexts.

Conclusion: Macro-aware retrieval yields robust, explainable forecasts under distributional change by operationalizing the principle that financial history often rhymes, providing both performance gains and transparency through interpretable evidence chains.

Abstract: Financial markets are inherently non-stationary: structural breaks and macroeconomic regime shifts often cause forecasting models to fail when deployed out of distribution (OOD). Conventional multimodal approaches that simply fuse numerical indicators and textual sentiment rarely adapt to such shifts. We introduce macro-contextual retrieval, a retrieval-augmented forecasting framework that grounds each prediction in historically analogous macroeconomic regimes. The method jointly embeds macro indicators (e.g., CPI, unemployment, yield spread, GDP growth) and financial news sentiment in a shared similarity space, enabling causal retrieval of precedent periods during inference without retraining. Trained on seventeen years of S&P 500 data (2007-2023) and evaluated OOD on AAPL (2024) and XOM (2024), the framework consistently narrows the CV to OOD performance gap. Macro-conditioned retrieval achieves the only positive out-of-sample trading outcomes (AAPL: PF=1.18, Sharpe=0.95; XOM: PF=1.16, Sharpe=0.61), while static numeric, text-only, and naive multimodal baselines collapse under regime shifts. Beyond metric gains, retrieved neighbors form interpretable evidence chains that correspond to recognizable macro contexts, such as inflationary or yield-curve inversion phases, supporting causal interpretability and transparency. By operationalizing the principle that “financial history may not repeat, but it often rhymes,” this work demonstrates that macro-aware retrieval yields robust, explainable forecasts under distributional change. All datasets, models, and source code are publicly available.

Method: Theoretical analysis comparing malicious noise (random corruption) and nasty noise (adversarial corruption) models, using cryptographic assumptions to prove separation results and defining ICE (ignore contradictory examples) algorithms to bridge the gap.

Result: Distribution-independent learning shows equivalence between malicious and nasty noise, while fixed-distribution learning shows arbitrarily large separation. ICE algorithms can tolerate nasty noise at half the rate of malicious noise, and this factor is necessary.

Conclusion: The relationship between malicious and nasty noise depends critically on the learning setting, with dramatic differences between distribution-independent and fixed-distribution scenarios, but ICE algorithms provide a natural bridge between the two noise models.

Abstract: We consider the relative abilities and limitations of computationally efficient algorithms for learning in the presence of noise, under two well-studied and challenging adversarial noise models for learning Boolean functions: malicious noise, in which an adversary can arbitrarily corrupt a random subset of examples given to the learner; and nasty noise, in which an adversary can arbitrarily corrupt an adversarially chosen subset of examples given to the learner. We consider both the distribution-independent and fixed-distribution settings. Our main results highlight a dramatic difference between these two settings: For distribution-independent learning, we prove a strong equivalence between the two noise models: If a class ${\cal C}$ of functions is efficiently learnable in the presence of $η$-rate malicious noise, then it is also efficiently learnable in the presence of $η$-rate nasty noise. In sharp contrast, for the fixed-distribution setting we show an arbitrarily large separation: Under a standard cryptographic assumption, for any arbitrarily large value $r$ there exists a concept class for which there is a ratio of $r$ between the rate $η_{malicious}$ of malicious noise that polynomial-time learning algorithms can tolerate, versus the rate $η_{nasty}$ of nasty noise that such learning algorithms can tolerate. To offset the negative result for the fixed-distribution setting, we define a broad and natural class of algorithms, namely those that ignore contradictory examples (ICE). We show that for these algorithms, malicious noise and nasty noise are equivalent up to a factor of two in the noise rate: Any efficient ICE learner that succeeds with $η$-rate malicious noise can be converted to an efficient learner that succeeds with $η/2$-rate nasty noise. We further show that the above factor of two is necessary, again under a standard cryptographic assumption.

Method: Decomposes 30-second epochs into 3-second tokens using CNN tokenizer, uses dual-level Transformers for local and extended context modeling, and fuses EEG/EOG modalities via Graph Convolutional Network.

Result: Achieved 85.3% accuracy on Sleep-EDF and 81.9% on ISRUC datasets, with state-of-the-art performance matching or exceeding published baselines in overall and per-class metrics.

Conclusion: Framing sleep as a compositional language enables unified hierarchical sequence modeling and multimodal fusion, advancing automated sleep staging toward more transparent and clinically meaningful applications.

Abstract: Automated sleep stage classification from polysomnography remains limited by the lack of expressive temporal hierarchies, challenges in multimodal EEG and EOG fusion, and the limited interpretability of deep learning models. We propose NeuroLingua, a language-inspired framework that conceptualizes sleep as a structured physiological language. Each 30-second epoch is decomposed into overlapping 3-second subwindows (“tokens”) using a CNN-based tokenizer, enabling hierarchical temporal modeling through dual-level Transformers: intra-segment encoding of local dependencies and inter-segment integration across seven consecutive epochs (3.5 minutes) for extended context. Modality-specific embeddings from EEG and EOG channels are fused via a Graph Convolutional Network, facilitating robust multimodal integration. NeuroLingua is evaluated on the Sleep-EDF Expanded and ISRUC-Sleep datasets, achieving state-of-the-art results on Sleep-EDF (85.3% accuracy, 0.800 macro F1, and 0.796 Cohen’s kappa) and competitive performance on ISRUC (81.9% accuracy, 0.802 macro F1, and 0.755 kappa), matching or exceeding published baselines in overall and per-class metrics. The architecture’s attention mechanisms enhance the detection of clinically relevant sleep microevents, providing a principled foundation for future interpretability, explainability, and causal inference in sleep research. By framing sleep as a compositional language, NeuroLingua unifies hierarchical sequence modeling and multimodal fusion, advancing automated sleep staging toward more transparent and clinically meaningful applications.

Method: Demonstrates out-of-context and in-context attacks where adversaries inject malicious tokens during decentralized GRPO training. Tests on math and coding tasks.

Result: Achieves up to 100% attack success rate in just 50 iterations, effectively poisoning benign nodes’ local LLM training.

Conclusion: Proposes two defense strategies that can achieve 100% stop rates against these attacks, making the attacks impossible when properly defended.

Abstract: Group Relative Policy Optimization (GRPO) has demonstrated great utilization in post-training of Large Language Models (LLMs). In GRPO, prompts are answered by the model and, through reinforcement learning, preferred completions are learnt. Owing to the small communication volume, GRPO is inherently suitable for decentralised training as the prompts can be concurrently answered by multiple nodes and then exchanged in the forms of strings. In this work, we present the first adversarial attack in decentralised GRPO. We demonstrate that malicious parties can poison such systems by injecting arbitrary malicious tokens in benign models in both out-of-context and in-context attacks. Using empirical examples of math and coding tasks, we show that adversarial attacks can easily poison the benign nodes, polluting their local LLM post-training, achieving attack success rates up to 100% in as few as 50 iterations. We propose two ways to defend against these attacks, depending on whether all users train the same model or different models. We show that these defenses can achieve stop rates of up to 100%, making the attack impossible.

Method: Proposed a theoretical framework showing JEPA’s predictive objective learns Koopman operator invariants, validated on synthetic data with known dynamics, and identified that constraining the linear predictor to be near-identity is the key inductive bias.

Result: Demonstrated that the near-identity constraint forces the encoder to learn regime indicator functions (Koopman eigenfunctions), and this constraint is critical for selecting interpretable solutions from mathematically equivalent but entangled optima.

Conclusion: This work demystifies JEPA behavior, provides principled connection between self-supervised learning and dynamical systems theory, and informs design of more robust and interpretable time-series models.

Abstract: Joint-Embedding Predictive Architectures (JEPAs), a powerful class of self-supervised models, exhibit an unexplained ability to cluster time-series data by their underlying dynamical regimes. We propose a novel theoretical explanation for this phenomenon, hypothesizing that JEPA’s predictive objective implicitly drives it to learn the invariant subspace of the system’s Koopman operator. We prove that an idealized JEPA loss is minimized when the encoder represents the system’s regime indicator functions, which are Koopman eigenfunctions. This theory was validated on synthetic data with known dynamics, demonstrating that constraining the JEPA’s linear predictor to be a near-identity operator is the key inductive bias that forces the encoder to learn these invariants. We further discuss that this constraint is critical for selecting this interpretable solution from a class of mathematically equivalent but entangled optima, revealing the predictor’s role in representation disentanglement. This work demystifies a key behavior of JEPAs, provides a principled connection between modern self-supervised learning and dynamical systems theory, and informs the design of more robust and interpretable time-series models.

Method: Uses a dual-stream architecture with classification branch for stepwise trend learning and regression branch for deviation estimation, plus multi-task loss with uncertainty-aware weighting and four variants incorporating CNNs, LSTMs, and Transformers.

Result: CaReTS outperforms state-of-the-art algorithms in forecasting accuracy and achieves higher trend classification performance on real-world datasets.

Conclusion: The proposed framework successfully provides both accurate predictions and interpretable insights by disentangling trends from deviations through multi-task learning.

Abstract: Recent advances in deep forecasting models have achieved remarkable performance, yet most approaches still struggle to provide both accurate predictions and interpretable insights into temporal dynamics. This paper proposes CaReTS, a novel multi-task learning framework that combines classification and regression tasks for multi-step time series forecasting problems. The framework adopts a dual-stream architecture, where a classification branch learns the stepwise trend into the future, while a regression branch estimates the corresponding deviations from the latest observation of the target variable. The dual-stream design provides more interpretable predictions by disentangling macro-level trends from micro-level deviations in the target variable. To enable effective learning in output prediction, deviation estimation, and trend classification, we design a multi-task loss with uncertainty-aware weighting to adaptively balance the contribution of each task. Furthermore, four variants (CaReTS1–4) are instantiated under this framework to incorporate mainstream temporal modelling encoders, including convolutional neural networks (CNNs), long short-term memory networks (LSTMs), and Transformers. Experiments on real-world datasets demonstrate that CaReTS outperforms state-of-the-art (SOTA) algorithms in forecasting accuracy, while achieving higher trend classification performance.

Method: Proposed an efficient algorithm for feasible BAI that allows choosing between testing an arm’s performance or any of its N feasibility constraints, focusing on fixed confidence setting to identify the feasible arm with highest performance with probability at least 1-δ.

Result: The algorithm naturally adapts to problem difficulty by eliminating arms based on worse performance or infeasibility, with theoretical analysis showing asymptotic optimality and empirical results demonstrating superior performance over state-of-the-art BAI algorithms on synthetic and real-world datasets.

Conclusion: The proposed algorithm effectively handles the practical challenge of separate performance and feasibility testing in BAI problems, providing both theoretical guarantees and empirical advantages over existing methods.

Abstract: Best arm identification (BAI) aims to identify the highest-performance arm among a set of $K$ arms by collecting stochastic samples from each arm. In real-world problems, the best arm needs to satisfy additional feasibility constraints. While there is limited prior work on BAI with feasibility constraints, they typically assume the performance and constraints are observed simultaneously on each pull of an arm. However, this assumption does not reflect most practical use cases, e.g., in drug discovery, we wish to find the most potent drug whose toxicity and solubility are below certain safety thresholds. These safety experiments can be conducted separately from the potency measurement. Thus, this requires designing BAI algorithms that not only decide which arm to pull but also decide whether to test for the arm’s performance or feasibility. In this work, we study feasible BAI which allows a decision-maker to choose a tuple $(i,\ell)$, where $i\in [K]$ denotes an arm and $\ell$ denotes whether she wishes to test for its performance ($\ell=0$) or any of its $N$ feasibility constraints ($\ell\in[N]$). We focus on the fixed confidence setting, which is to identify the \textit{feasible} arm with the \textit{highest performance}, with a probability of at least $1-δ$. We propose an efficient algorithm and upper-bound its sample complexity, showing our algorithm can naturally adapt to the problem’s difficulty and eliminate arms by worse performance or infeasibility, whichever is easier. We complement this upper bound with a lower bound showing that our algorithm is \textit{asymptotically ($δ\rightarrow 0$) optimal}. Finally, we empirically show that our algorithm outperforms other state-of-the-art BAI algorithms in both synthetic and real-world datasets.

Method: Analyzed gradient flow in overparameterized optimization problems interpreted as neural networks with linear activations, using proper real analytic cost functions and specializing to scalar-valued cases.

Result: Showed global convergence for any proper real analytic cost function, demonstrated universal structural features across all admissible costs, and revealed convergence acceleration based on initialization imbalance.

Conclusion: Compositional structure fundamentally shapes optimization landscapes, with universal properties that may extend beyond linear activations to networks with sigmoidal activations.

Abstract: This paper investigates how the compositional structure of neural networks shapes their optimization landscape and training dynamics. We analyze the gradient flow associated with overparameterized optimization problems, which can be interpreted as training a neural network with linear activations. Remarkably, we show that the global convergence properties can be derived for any cost function that is proper and real analytic. We then specialize the analysis to scalar-valued cost functions, where the geometry of the landscape can be fully characterized. In this setting, we demonstrate that key structural features – such as the location and stability of saddle points – are universal across all admissible costs, depending solely on the overparameterized representation rather than on problem-specific details. Moreover, we show that convergence can be arbitrarily accelerated depending on the initialization, as measured by an imbalance metric introduced in this work. Finally, we discuss how these insights may generalize to neural networks with sigmoidal activations, showing through a simple example which geometric and dynamical properties persist beyond the linear case.

Method: Step-wise Momentum Fusion (SMoFi) synchronizes momentum buffers across server-side optimizers and uses staleness-aware alignment to control gradient divergence by constraining gradient updates of server-side submodels.

Result: Extensive validations show SMoFi improves global model accuracy up to 7.1% and convergence speed up to 10.25×, with greater impact when more clients and deeper models are involved.

Conclusion: SMoFi is particularly suitable for model training in resource-constrained contexts due to its effectiveness in handling data heterogeneity and improving performance.

Abstract: Split Federated Learning is a system-efficient federated learning paradigm that leverages the rich computing resources at a central server to train model partitions. Data heterogeneity across silos, however, presents a major challenge undermining the convergence speed and accuracy of the global model. This paper introduces Step-wise Momentum Fusion (SMoFi), an effective and lightweight framework that counteracts gradient divergence arising from data heterogeneity by synchronizing the momentum buffers across server-side optimizers. To control gradient divergence over the training process, we design a staleness-aware alignment mechanism that imposes constraints on gradient updates of the server-side submodel at each optimization step. Extensive validations on multiple real-world datasets show that SMoFi consistently improves global model accuracy (up to 7.1%) and convergence speed (up to 10.25$\times$). Furthermore, SMoFi has a greater impact with more clients involved and deeper learning models, making it particularly suitable for model training in resource-constrained contexts.

Method: The algorithm uses a novel duality framework to analyze moment tensor structure, combining refined Jennrich’s Algorithm with PCA over random projections of the moment tensor and gradient-descent-based non-convex optimization.

Result: The algorithm achieves poly(d,1/γ) time complexity under factorizable distributions, establishing a strong separation between CSQ and SQ for this learning problem.

Conclusion: The work demonstrates that by leveraging more general statistical queries (SQ) rather than just correlational ones, polynomial-time learning is achievable for intersections of halfspaces under natural distributional assumptions.

Abstract: Learning intersections of halfspaces is a central problem in Computational Learning Theory. Even for just two halfspaces, it remains a major open question whether learning is possible in polynomial time with respect to the margin $γ$ of the data points and their dimensionality $d$. The best-known algorithms run in quasi-polynomial time $d^{O(\log(1/γ))}$, and it has been shown that this complexity is unavoidable for any algorithm relying solely on correlational statistical queries (CSQ). In this work, we introduce a novel algorithm that provably circumvents the CSQ hardness barrier. Our approach applies to a broad class of distributions satisfying a natural, previously studied, factorizability assumption. Factorizable distributions lie between distribution-specific and distribution-free settings, and significantly extend previously known tractable cases. Under these distributions, we show that CSQ-based methods still require quasipolynomial time even for weakly learning, whereas our algorithm achieves $poly(d,1/γ)$ time by leveraging more general statistical queries (SQ), establishing a strong separation between CSQ and SQ for this simple realizable PAC learning problem. Our result is grounded in a rigorous analysis utilizing a novel duality framework that characterizes the moment tensor structure induced by the marginal distributions. Building on these structural insights, we propose new, efficient learning algorithms. These algorithms combine a refined variant of Jennrich’s Algorithm with PCA over random projections of the moment tensor, along with a gradient-descent-based non-convex optimization framework.

Method: Proposed ACT pipeline where LLMs serve as annotators and critical judges to identify potential errors. Human reviewers focus only on suspicious cases. Includes theoretical analysis of loss function modification for training models on ACT data.

Result: Performance gap reduced to less than 2% on most benchmark datasets compared to fully human-annotated data, while saving up to 90% of human annotation costs. Applicable across NLP, CV, and multimodal domains using MLLMs.

Conclusion: ACT pipeline successfully balances annotation quality and efficiency by leveraging LLMs for both annotation and error detection, enabling near-human performance with dramatically reduced human effort across multiple domains.

Abstract: Supervised learning relies on high-quality labeled data, but obtaining such data through human annotation is both expensive and time-consuming. Recent work explores using large language models (LLMs) for annotation, but LLM-generated labels still fall short of human-level quality. To address this problem, we propose the Annotation with Critical Thinking (ACT) data pipeline, where LLMs serve not only as annotators but also as judges to critically identify potential errors. Human effort is then directed towards reviewing only the most “suspicious” cases, significantly improving the human annotation efficiency. Our major contributions are as follows: (1) ACT is applicable to a wide range of domains, including natural language processing (NLP), computer vision (CV), and multimodal understanding, by leveraging multimodal-LLMs (MLLMs). (2) Through empirical studies, we derive 7 insights on how to enhance annotation quality while efficiently reducing the human cost, and then translate these findings into user-friendly guidelines. (3) We theoretically analyze how to modify the loss function so that models trained on ACT data achieve similar performance to those trained on fully human-annotated data. Our experiments show that the performance gap can be reduced to less than 2% on most benchmark datasets while saving up to 90% of human costs.

Method: Uses a steering vector computed from verifier’s hidden states and injected into pretrained drafter for dynamic alignment, avoiding the need for offline methods like distillation.

Result: Boosts accepted tokens by up to 35% under standard sampling and 22% under greedy sampling with negligible computational overhead.

Conclusion: The approach effectively improves pretrained drafter alignment, can be retrofitted to existing models, and enables rapid adoption without significant computational cost.

Abstract: Speculative decoding accelerates language model inference by separating generation into fast drafting and parallel verification. Its main limitation is drafter-verifier misalignment, which limits token acceptance and reduces overall effectiveness. While small drafting heads trained from scratch compensate with speed, they struggle when verification dominates latency or when inputs are out of distribution. In contrast, pretrained drafters, though slower, achieve higher acceptance rates thanks to stronger standalone generation capabilities, making them competitive when drafting latency is negligible relative to verification or communication overhead. In this work, we aim to improve the acceptance rates of pretrained drafters by introducing a lightweight dynamic alignment mechanism: a steering vector computed from the verifier’s hidden states and injected into the pretrained drafter. Compared to existing offline alignment methods such as distillation, our approach boosts the number of accepted tokens by up to 35% under standard sampling and 22% under greedy sampling, all while incurring negligible computational overhead. Importantly, our approach can be retrofitted to existing architectures and pretrained models, enabling rapid adoption.

Method: Proposes ConSurv with two key components: Multi-staged Mixture of Experts (MS-MoE) to capture task-shared and task-specific knowledge at different network stages, and Feature Constrained Replay (FCR) to mitigate forgetting by restricting feature deviation of previous data at encoder and fusion levels.

Result: Extensive experiments show ConSurv outperforms competing methods across multiple metrics on the new MSAIL benchmark integrating four datasets for multimodal survival analysis incremental learning.

Conclusion: ConSurv successfully addresses catastrophic forgetting and complex inter-modal interactions in multimodal continual learning for survival analysis, demonstrating superior performance over existing methods.

Abstract: Survival prediction of cancers is crucial for clinical practice, as it informs mortality risks and influences treatment plans. However, a static model trained on a single dataset fails to adapt to the dynamically evolving clinical environment and continuous data streams, limiting its practical utility. While continual learning (CL) offers a solution to learn dynamically from new datasets, existing CL methods primarily focus on unimodal inputs and suffer from severe catastrophic forgetting in survival prediction. In real-world scenarios, multimodal inputs often provide comprehensive and complementary information, such as whole slide images and genomics; and neglecting inter-modal correlations negatively impacts the performance. To address the two challenges of catastrophic forgetting and complex inter-modal interactions between gigapixel whole slide images and genomics, we propose ConSurv, the first multimodal continual learning (MMCL) method for survival analysis. ConSurv incorporates two key components: Multi-staged Mixture of Experts (MS-MoE) and Feature Constrained Replay (FCR). MS-MoE captures both task-shared and task-specific knowledge at different learning stages of the network, including two modality encoders and the modality fusion component, learning inter-modal relationships. FCR further enhances learned knowledge and mitigates forgetting by restricting feature deviation of previous data at different levels, including encoder-level features of two modalities and the fusion-level representations. Additionally, we introduce a new benchmark integrating four datasets, Multimodal Survival Analysis Incremental Learning (MSAIL), for comprehensive evaluation in the CL setting. Extensive experiments demonstrate that ConSurv outperforms competing methods across multiple metrics.

Method: Reviews evaluation methods for forgetting, resilience against adversarial attacks, technical solutions (differential privacy, homomorphic encryption, federated learning, ephemeral memory), and institutional safeguards (auditing, regulatory frameworks).

Result: Steady progress but robust and verifiable unlearning remains unresolved; current approaches have limitations in efficiency, security, and transparency.

Conclusion: Need efficient techniques avoiding costly retraining, stronger defenses against adversarial recovery, and governance structures reinforcing accountability for safe LLM deployment in sensitive applications.

Abstract: The growing use of large language models in sensitive domains has exposed a critical weakness: the inability to ensure that private information can be permanently forgotten. Yet these systems still lack reliable mechanisms to guarantee that sensitive information can be permanently removed once it has been used. Retraining from the beginning is prohibitively costly, and existing unlearning methods remain fragmented, difficult to verify, and often vulnerable to recovery. This paper surveys recent research on machine unlearning for LLMs and considers how far current approaches can address these challenges. We review methods for evaluating whether forgetting has occurred, the resilience of unlearned models against adversarial attacks, and mechanisms that can support user trust when model complexity or proprietary limits restrict transparency. Technical solutions such as differential privacy, homomorphic encryption, federated learning, and ephemeral memory are examined alongside institutional safeguards including auditing practices and regulatory frameworks. The review finds steady progress, but robust and verifiable unlearning is still unresolved. Efficient techniques that avoid costly retraining, stronger defenses against adversarial recovery, and governance structures that reinforce accountability are needed if LLMs are to be deployed safely in sensitive applications. By integrating technical and organizational perspectives, this study outlines a pathway toward AI systems that can be required to forget, while maintaining both privacy and public trust.

Method: UGCS identifies hard question-answer pairs using per-sample uncertainty and ranks checkpoints by their performance on these challenging cases. It averages rewards of top-uncertain samples over a short training window without additional computation.

Result: Experiments across three datasets and three LLMs show UGCS consistently identifies checkpoints with stronger generalization, outperforming traditional strategies based on training or validation performance.

Conclusion: Models that solve their hardest tasks with low uncertainty are the most reliable overall, demonstrating the effectiveness of uncertainty-guided checkpoint selection.

Abstract: Reinforcement learning (RL) finetuning is crucial to aligning large language models (LLMs), but the process is notoriously unstable and exhibits high variance across model checkpoints. In practice, selecting the best checkpoint is challenging: evaluating checkpoints on the validation set during training is computationally expensive and requires a good validation set, while relying on the final checkpoint provides no guarantee of good performance. We introduce an uncertainty-guided approach for checkpoint selection (UGCS) that avoids these pitfalls. Our method identifies hard question-answer pairs using per-sample uncertainty and ranks checkpoints by how well they handle these challenging cases. By averaging the rewards of the top-uncertain samples over a short training window, our method produces a stable and discriminative signal without additional forward passes or significant computation overhead. Experiments across three datasets and three LLMs demonstrate that it consistently identifies checkpoints with stronger generalization, outperforming traditional strategies such as relying on training or validation performance. These results highlight that models solving their hardest tasks with low uncertainty are the most reliable overall.

Method: Proposes a fully differentiable, exemplar-free expandable method with two memories: common features memory for cross-task sharing and discriminative characteristics memory for sample-specific learning. Uses memory adjustment module for adaptive pruning/expansion and orthogonal regularization for geometric separation.

Result: Outperforms 14 state-of-the-art methods on CIFAR-10, CIFAR-100, and Tiny-ImageNet with final accuracies of 55.13%, 37.24%, and 30.11% respectively. Achieves feature extraction closest to the upper bound.

Conclusion: The method effectively integrates and utilizes knowledge across sequential tasks, establishing a new milestone in continual learning through autonomous learning of latent representations and prevention of interference.

Abstract: Continual learning methods used to force neural networks to process sequential tasks in isolation, preventing them from leveraging useful inter-task relationships and causing them to repeatedly relearn similar features or overly differentiate them. To address this problem, we propose a fully differentiable, exemplar-free expandable method composed of two complementary memories: One learns common features that can be used across all tasks, and the other combines the shared features to learn discriminative characteristics unique to each sample. Both memories are differentiable so that the network can autonomously learn latent representations for each sample. For each task, the memory adjustment module adaptively prunes critical slots and minimally expands capacity to accommodate new concepts, and orthogonal regularization enforces geometric separation between preserved and newly learned memory components to prevent interference. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet show that the proposed method outperforms 14 state-of-the-art methods for class-incremental learning, achieving final accuracies of 55.13%, 37.24%, and 30.11%, respectively. Additional analysis confirms that, through effective integration and utilization of knowledge, the proposed method can increase average performance across sequential tasks, and it produces feature extraction results closest to the upper bound, thus establishing a new milestone in continual learning.

Method: Select representative anchors using fair sampling, apply any fair clustering algorithm to anchors, then propagate labels via optimization with group-label joint constraints using ADMM.

Result: AFCF reduces computational time by orders of magnitude while maintaining clustering performance and fairness guarantees on large-scale benchmarks.

Conclusion: AFCF provides an efficient and scalable solution for fair clustering that bridges the gap between fairness requirements and computational feasibility.

Abstract: Fair clustering is crucial for mitigating bias in unsupervised learning, yet existing algorithms often suffer from quadratic or super-quadratic computational complexity, rendering them impractical for large-scale datasets. To bridge this gap, we introduce the Anchor-based Fair Clustering Framework (AFCF), a novel, general, and plug-and-play framework that empowers arbitrary fair clustering algorithms with linear-time scalability. Our approach first selects a small but representative set of anchors using a novel fair sampling strategy. Then, any off-the-shelf fair clustering algorithm can be applied to this small anchor set. The core of our framework lies in a novel anchor graph construction module, where we formulate an optimization problem to propagate labels while preserving fairness. This is achieved through a carefully designed group-label joint constraint, which we prove theoretically ensures that the fairness of the final clustering on the entire dataset matches that of the anchor clustering. We solve this optimization efficiently using an ADMM-based algorithm. Extensive experiments on multiple large-scale benchmarks demonstrate that AFCF drastically accelerates state-of-the-art methods, which reduces computational time by orders of magnitude while maintaining strong clustering performance and fairness guarantees.

Method: SE-Diff integrates an ODE-based ECG simulator via beat decoder and simulator-consistent constraints, and uses LLM-powered experience retrieval to inject clinical knowledge into the diffusion process.

Result: SE-Diff outperforms baselines in signal fidelity and text-ECG semantic alignment, and the integrated knowledge also benefits downstream ECG classification tasks.

Conclusion: SE-Diff effectively combines physiological simulator knowledge and clinical experience for comprehensive ECG generation, demonstrating superiority over existing methods.

Abstract: Cardiovascular disease (CVD) is a leading cause of mortality worldwide. Electrocardiograms (ECGs) are the most widely used non-invasive tool for cardiac assessment, yet large, well-annotated ECG corpora are scarce due to cost, privacy, and workflow constraints. Generating ECGs can be beneficial for the mechanistic understanding of cardiac electrical activity, enable the construction of large, heterogeneous, and unbiased datasets, and facilitate privacy-preserving data sharing. Generating realistic ECG signals from clinical context is important yet underexplored. Recent work has leveraged diffusion models for text-to-ECG generation, but two challenges remain: (i) existing methods often overlook the physiological simulator knowledge of cardiac activity; and (ii) they ignore broader, experience-based clinical knowledge grounded in real-world practice. To address these gaps, we propose SE-Diff, a novel physiological simulator and experience enhanced diffusion model for comprehensive ECG generation. SE-Diff integrates a lightweight ordinary differential equation (ODE)-based ECG simulator into the diffusion process via a beat decoder and simulator-consistent constraints, injecting mechanistic priors that promote physiologically plausible waveforms. In parallel, we design an LLM-powered experience retrieval-augmented strategy to inject clinical knowledge, providing more guidance for ECG generation. Extensive experiments on real-world ECG datasets demonstrate that SE-Diff improves both signal fidelity and text-ECG semantic alignment over baselines, proving its superiority for text-to-ECG generation. We further show that the simulator-based and experience-based knowledge also benefit downstream ECG classification.

Method: Developed Simultaneous Weight and Sample Tailoring (SWaST) that alternately performs weight pruning and coreset selection. Integrated a state preservation mechanism to prevent critical double-loss phenomenon where important weights and supportive samples are mistakenly eliminated simultaneously.

Result: Extensive experiments show strong synergy between pruning and coreset selection, delivering accuracy boosts up to 17.83% alongside 10% to 90% FLOPs reductions across varying prune rates and coreset sizes.

Conclusion: The interplay between weight pruning and coreset selection can be effectively harnessed through SWaST, achieving both computational efficiency and improved model accuracy when properly stabilized against critical double-loss.

Abstract: Modern deep neural networks rely heavily on massive model weights and training samples, incurring substantial computational costs. Weight pruning and coreset selection are two emerging paradigms proposed to improve computational efficiency. In this paper, we first explore the interplay between redundant weights and training samples through a transparent analysis: redundant samples, particularly noisy ones, cause model weights to become unnecessarily overtuned to fit them, complicating the identification of irrelevant weights during pruning; conversely, irrelevant weights tend to overfit noisy data, undermining coreset selection effectiveness. To further investigate and harness this interplay in deep learning, we develop a Simultaneous Weight and Sample Tailoring mechanism (SWaST) that alternately performs weight pruning and coreset selection to establish a synergistic effect in training. During this investigation, we observe that when simultaneously removing a large number of weights and samples, a phenomenon we term critical double-loss can occur, where important weights and their supportive samples are mistakenly eliminated at the same time, leading to model instability and nearly irreversible degradation that cannot be recovered in subsequent training. Unlike classic machine learning models, this issue can arise in deep learning due to the lack of theoretical guarantees on the correctness of weight pruning and coreset selection, which explains why these paradigms are often developed independently. We mitigate this by integrating a state preservation mechanism into SWaST, enabling stable joint optimization. Extensive experiments reveal a strong synergy between pruning and coreset selection across varying prune rates and coreset sizes, delivering accuracy boosts of up to 17.83% alongside 10% to 90% FLOPs reductions.

Method: Using topological-dynamical arguments and algebraic properties of autonomous flows to prove impossibility theorems for single-step flows and construct universal approximation results through compositions of multiple flows, including dimension lifting techniques.

Result: Single-step autonomous flows are meagre and not universal, while compositions of at most K_d flows achieve O(n^{-1/d}) approximation rates for Lipschitz homeomorphisms, with dimension-free rates under smoothness assumptions, and universal approximation for continuous functions and probability measures.

Conclusion: Incremental generation through flow compositions is both necessary and sufficient for universal flow-based generation, providing theoretical justification for the empirical success of incremental denoising models.

Abstract: Incremental flow-based denoising models have reshaped generative modelling, but their empirical advantage still lacks a rigorous approximation-theoretic foundation. We show that incremental generation is necessary and sufficient for universal flow-based generation on the largest natural class of self-maps of $[0,1]^d$ compatible with denoising pipelines, namely the orientation-preserving homeomorphisms of $[0,1]^d$. All our guarantees are uniform on the underlying maps and hence imply approximation both samplewise and in distribution. Using a new topological-dynamical argument, we first prove an impossibility theorem: the class of all single-step autonomous flows, independently of the architecture, width, depth, or Lipschitz activation of the underlying neural network, is meagre and therefore not universal in the space of orientation-preserving homeomorphisms of $[0,1]^d$. By exploiting algebraic properties of autonomous flows, we conversely show that every orientation-preserving Lipschitz homeomorphism on $[0,1]^d$ can be approximated at rate $\mathcal{O}(n^{-1/d})$ by a composition of at most $K_d$ such flows, where $K_d$ depends only on the dimension. Under additional smoothness assumptions, the approximation rate can be made dimension-free, and $K_d$ can be chosen uniformly over the class being approximated. Finally, by linearly lifting the domain into one higher dimension, we obtain structured universal approximation results for continuous functions and for probability measures on $[0,1]^d$, the latter realized as pushforwards of empirical measures with vanishing $1$-Wasserstein error.

Method: PRISM decouples logit-matching and regularization objectives, supervising them with different teacher architectures: a primary model for logits and a stochastic subset for batch-normalization alignment. It also introduces scalable cross-class batch formation for fast parallel synthesis.

Result: On ImageNet-1K, PRISM consistently outperforms single-teacher methods (SRe2L) and multi-teacher variants (G-VBSM) at low- and mid-IPC regimes, with significantly richer intra-class diversity reflected by notable drop in cosine similarity between features.

Conclusion: PRISM effectively addresses the architectural bias problem in dataset distillation through multi-teacher supervision, achieving better performance and diversity while enabling scalable synthesis.

Abstract: Dataset distillation (DD) promises compact yet faithful synthetic data, but existing approaches often inherit the inductive bias of a single teacher model. As dataset size increases, this bias drives generation toward overly smooth, homogeneous samples, reducing intra-class diversity and limiting generalization. We present PRISM (PRIors from diverse Source Models), a framework that disentangles architectural priors during synthesis. PRISM decouples the logit-matching and regularization objectives, supervising them with different teacher architectures: a primary model for logits and a stochastic subset for batch-normalization (BN) alignment. On ImageNet-1K, PRISM consistently and reproducibly outperforms single-teacher methods (e.g., SRe2L) and recent multi-teacher variants (e.g., G-VBSM) at low- and mid-IPC regimes. The generated data also show significantly richer intra-class diversity, as reflected by a notable drop in cosine similarity between features. We further analyze teacher selection strategies (pre- vs. intra-distillation) and introduce a scalable cross-class batch formation scheme for fast parallel synthesis. Code will be released after the review period.

Method: Proposes a dual-pathway encoder to independently reconstruct missing node attributes and graph structure, preventing error propagation. Fuses pathways in a joint latent space where normals occupy a compact inner manifold and anomalies reside on an outer shell. Uses hard negative sampling by decoding latent codes outside the normal region to create realistic negative examples that sharpen decision boundaries.

Result: Experiments on seven public benchmarks show M$^2$V-UGAD consistently outperforms existing unsupervised GAD methods across varying missing rates.

Conclusion: The proposed framework effectively handles multiple missing values in graph data, mitigates cross-view interference and imputation bias, and improves anomaly detection performance in incomplete graph scenarios.

Abstract: Unsupervised graph anomaly detection (GAD) has received increasing attention in recent years, which aims to identify data anomalous patterns utilizing only unlabeled node information from graph-structured data. However, prevailing unsupervised GAD methods typically presuppose complete node attributes and structure information, a condition hardly satisfied in real-world scenarios owing to privacy, collection errors or dynamic node arrivals. Existing standard imputation schemes risk “repairing” rare anomalous nodes so that they appear normal, thereby introducing imputation bias into the detection process. In addition, when both node attributes and edges are missing simultaneously, estimation errors in one view can contaminate the other, causing cross-view interference that further undermines the detection performance. To overcome these challenges, we propose M$^2$V-UGAD, a multiple missing values-resistant unsupervised GAD framework on incomplete graphs. Specifically, a dual-pathway encoder is first proposed to independently reconstruct missing node attributes and graph structure, thereby preventing errors in one view from propagating to the other. The two pathways are then fused and regularized in a joint latent space so that normals occupy a compact inner manifold while anomalies reside on an outer shell. Lastly, to mitigate imputation bias, we sample latent codes just outside the normal region and decode them into realistic node features and subgraphs, providing hard negative examples that sharpen the decision boundary. Experiments on seven public benchmarks demonstrate that M$^2$V-UGAD consistently outperforms existing unsupervised GAD methods across varying missing rates.

Method: Two-stage pipeline: PPO pretraining followed by TD-ES refinement using bounded antithetic triangular perturbations and centered-rank finite-difference estimation for gradient-free updates.

Result: Across robotic manipulation tasks, TD-ES increased success rates by 26.5% relative to PPO alone and significantly reduced performance variance.

Conclusion: TD-ES provides a simple, compute-efficient path to reliable policy refinement that preserves early sample efficiency while enabling robust late-stage performance gains.

Abstract: Improving competent robot policies with on-policy RL is often hampered by noisy, low-signal gradients. We revisit Evolution Strategies (ES) as a policy-gradient proxy and localize exploration with bounded, antithetic triangular perturbations, suitable for policy refinement. We propose Triangular-Distribution ES (TD-ES) which pairs bounded triangular noise with a centered-rank finite-difference estimator to deliver stable, parallelizable, gradient-free updates. In a two-stage pipeline – PPO pretraining followed by TD-ES refinement – this preserves early sample efficiency while enabling robust late-stage gains. Across a suite of robotic manipulation tasks, TD-ES raises success rates by 26.5% relative to PPO and greatly reduces variance, offering a simple, compute-light path to reliable refinement.

Method: Proposes three key innovations: 1) Adaptive fused dual-domain seasonal MLP that categorizes seasonal signals into strong/weak components with adaptive channel fusion; 2) Energy invariant attention mechanism for adaptive feature channel focus while maintaining constant total signal energy; 3) Dynamic capacity adjustment mechanism that scales neuron count with square root of channel count.

Result: Extensive experiments across nine benchmark datasets demonstrate state-of-the-art performance in both prediction accuracy and computational efficiency.

Conclusion: MDMLP-EIA effectively addresses key limitations of existing MLP-based forecasting methods and achieves superior performance through its multi-domain dynamic MLP architecture with energy invariant attention.

Abstract: Time series forecasting is essential across diverse domains. While MLP-based methods have gained attention for achieving Transformer-comparable performance with fewer parameters and better robustness, they face critical limitations including loss of weak seasonal signals, capacity constraints in weight-sharing MLPs, and insufficient channel fusion in channel-independent strategies. To address these challenges, we propose MDMLP-EIA (Multi-domain Dynamic MLPs with Energy Invariant Attention) with three key innovations. First, we develop an adaptive fused dual-domain seasonal MLP that categorizes seasonal signals into strong and weak components. It employs an adaptive zero-initialized channel fusion strategy to minimize noise interference while effectively integrating predictions. Second, we introduce an energy invariant attention mechanism that adaptively focuses on different feature channels within trend and seasonal predictions across time steps. This mechanism maintains constant total signal energy to align with the decomposition-prediction-reconstruction framework and enhance robustness against disturbances. Third, we propose a dynamic capacity adjustment mechanism for channel-independent MLPs. This mechanism scales neuron count with the square root of channel count, ensuring sufficient capacity as channels increase. Extensive experiments across nine benchmark datasets demonstrate that MDMLP-EIA achieves state-of-the-art performance in both prediction accuracy and computational efficiency.

Method: Leverage large language models to schedule and plan multiple tools in a toolbox for EEG preprocessing, feature extraction, event detection, and other EEG-related tasks.

Result: EEGAgent successfully performs EEG basic information perception, spatiotemporal exploration, event detection, user interaction, and report generation, evaluated on public datasets.

Conclusion: EEGAgent supports flexible and interpretable EEG analysis, demonstrating potential for real-world clinical applications through its multi-task capabilities.

Abstract: Scalable and generalizable analysis of brain activity is essential for advancing both clinical diagnostics and cognitive research. Electroencephalography (EEG), a non-invasive modality with high temporal resolution, has been widely used for brain states analysis. However, most existing EEG models are usually tailored for individual specific tasks, limiting their utility in realistic scenarios where EEG analysis often involves multi-task and continuous reasoning. In this work, we introduce EEGAgent, a general-purpose framework that leverages large language models (LLMs) to schedule and plan multiple tools to automatically complete EEG-related tasks. EEGAgent is capable of performing the key functions: EEG basic information perception, spatiotemporal EEG exploration, EEG event detection, interaction with users, and EEG report generation. To realize these capabilities, we design a toolbox composed of different tools for EEG preprocessing, feature extraction, event detection, etc. These capabilities were evaluated on public datasets, and our EEGAgent can support flexible and interpretable EEG analysis, highlighting its potential for real-world clinical applications.

Method: Proves theoretically that dynamic thresholds outperform fixed ones, then introduces Dynamic Threshold Determination algorithm that enhances existing detectors with adaptive threshold adjustment.

Result: Extensive experiments on synthetic and real-world datasets (image and tabular) show substantial performance improvements over state-of-the-art drift detectors.

Conclusion: Dynamic threshold adaptation is provably better than fixed thresholds and significantly enhances drift detection performance across diverse datasets.

Abstract: Existing drift detection methods focus on designing sensitive test statistics. They treat the detection threshold as a fixed hyperparameter, set once to balance false alarms and late detections, and applied uniformly across all datasets and over time. However, maintaining model performance is the key objective from the perspective of machine learning, and we observe that model performance is highly sensitive to this threshold. This observation inspires us to investigate whether a dynamic threshold could be provably better. In this paper, we prove that a threshold that adapts over time can outperform any single fixed threshold. The main idea of the proof is that a dynamic strategy, constructed by combining the best threshold from each individual data segment, is guaranteed to outperform any single threshold that apply to all segments. Based on the theorem, we propose a Dynamic Threshold Determination algorithm. It enhances existing drift detection frameworks with a novel comparison phase to inform how the threshold should be adjusted. Extensive experiments on a wide range of synthetic and real-world datasets, including both image and tabular data, validate that our approach substantially enhances the performance of state-of-the-art drift detectors.

Method: Introduces MultiTab-Net with a novel multitask masked-attention mechanism that dynamically models feature-feature dependencies while mitigating task competition. Also creates MultiTab-Bench, a synthetic dataset generator for systematic multitask evaluation.

Result: MultiTab-Net consistently achieves higher multitask gain than existing MTL architectures and single-task transformers across diverse domains including recommendation systems, socioeconomic data, and physics datasets, handling various task counts, types, and feature modalities.

Conclusion: MultiTab-Net successfully bridges the gap between transformer architectures and multitask learning for tabular data, demonstrating superior performance and scalability while providing a systematic evaluation framework through MultiTab-Bench.

Abstract: Tabular data is the most abundant data type in the world, powering systems in finance, healthcare, e-commerce, and beyond. As tabular datasets grow and span multiple related targets, there is an increasing need to exploit shared task information for improved multitask generalization. Multitask learning (MTL) has emerged as a powerful way to improve generalization and efficiency, yet most existing work focuses narrowly on large-scale recommendation systems, leaving its potential in broader tabular domains largely underexplored. Also, existing MTL approaches for tabular data predominantly rely on multi-layer perceptron-based backbones, which struggle to capture complex feature interactions and often fail to scale when data is abundant, a limitation that transformer architectures have overcome in other domains. Motivated by this, we introduce MultiTab-Net, the first multitask transformer architecture specifically designed for large tabular data. MultiTab-Net employs a novel multitask masked-attention mechanism that dynamically models feature-feature dependencies while mitigating task competition. Through extensive experiments, we show that MultiTab-Net consistently achieves higher multitask gain than existing MTL architectures and single-task transformers across diverse domains including large-scale recommendation data, census-like socioeconomic data, and physics datasets, spanning a wide range of task counts, task types, and feature modalities. In addition, we contribute MultiTab-Bench, a generalized multitask synthetic dataset generator that enables systematic evaluation of multitask dynamics by tuning task count, task correlations, and relative task complexity. Our code is publicly available at https://github.com/Armanfard-Lab/MultiTab.

Method: Used AI Feynman symbolic regression with observational and inductive biases through data preprocessing and search space restriction on lunar ephemerides data. Proposed automated preprocessing for coordinate system selection.

Result: Successfully recovered the first-order analytical form of the Equation of the Centre, demonstrating that targeted domain knowledge enables symbolic regression to rediscover physical laws.

Conclusion: Domain knowledge embedding enables symbolic regression to find physical laws, but challenges remain in constraining symbolic regression to derive physics equations when using tailored biases.

Abstract: This work explores using the physics-inspired AI Feynman symbolic regression algorithm to automatically rediscover a fundamental equation in astronomy – the Equation of the Centre. Through the introduction of observational and inductive biases corresponding to the physical nature of the system through data preprocessing and search space restriction, AI Feynman was successful in recovering the first-order analytical form of this equation from lunar ephemerides data. However, this manual approach highlights a key limitation in its reliance on expert-driven coordinate system selection. We therefore propose an automated preprocessing extension to find the canonical coordinate system. Results demonstrate that targeted domain knowledge embedding enables symbolic regression to rediscover physical laws, but also highlight further challenges in constraining symbolic regression to derive physics equations when leveraging domain knowledge through tailored biases.

Method: The study investigates fundamental challenges of multimodal learning robustness in open-world settings, though specific methods are not detailed in the abstract.

Result: The abstract does not provide specific experimental results, focusing instead on identifying the problem space and research direction.

Conclusion: There is a critical need to bridge the gap between controlled experimental performance and practical deployment requirements for multimodal learning systems in open-world environments.

Abstract: The rapid evolution of machine learning has propelled neural networks to unprecedented success across diverse domains. In particular, multimodal learning has emerged as a transformative paradigm, leveraging complementary information from heterogeneous data streams (e.g., text, vision, audio) to advance contextual reasoning and intelligent decision-making. Despite these advancements, current neural network-based models often fall short in open-world environments characterized by inherent unpredictability, where unpredictable environmental composition dynamics, incomplete modality inputs, and spurious distributions relations critically undermine system reliability. While humans naturally adapt to such dynamic, ambiguous scenarios, artificial intelligence systems exhibit stark limitations in robustness, particularly when processing multimodal signals under real-world complexity. This study investigates the fundamental challenge of multimodal learning robustness in open-world settings, aiming to bridge the gap between controlled experimental performance and practical deployment requirements.

Method: Proposed learning paradigm that strongly incorporates empirical data and minimizes algorithmic decisions based on prior assumptions.

Result: The approach demonstrates generalization capabilities across various learning assumptions including similarity, clustering, Lipschitzness, and contrastive learning.

Conclusion: The method enables utilization of learning assumptions without requiring knowledge of their true parameters in advance.

Abstract: We address the general task of learning with a set of candidate models that is too large to have a uniform convergence of empirical estimates to true losses. While the common approach to such challenges is SRM (or regularization) based learning algorithms, we propose a novel learning paradigm that relies on stronger incorporation of empirical data and requires less algorithmic decisions to be based on prior assumptions. We analyze the generalization capabilities of our approach and demonstrate its merits in several common learning assumptions, including similarity of close points, clustering of the domain into highly label-homogeneous regions, Lipschitzness assumptions of the labeling rule, and contrastive learning assumptions. Our approach allows utilizing such assumptions without the need to know their true parameters a priori.

Method: Uses structured chain-of-thought prompts with LLMs to extract tuning hints from documents, then applies a hint-aware demonstration reinforcement learning algorithm (HA-DDPGfD) to integrate these hints into RL agent training.

Result: Experimental evaluations on MySQL and PostgreSQL show significant performance improvements and online tuning cost reduction compared to DB-BERT, GPTuner and CDBTune baselines, with superior adaptability to unknown workloads.

Conclusion: DemoTuner successfully leverages textual knowledge through LLM-assisted demonstration reinforcement learning, establishing the first demonstration RL approach for DBMS knob tuning with strong practical advantages.

Abstract: The performance of modern DBMSs such as MySQL and PostgreSQL heavily depends on the configuration of performance-critical knobs. Manual tuning these knobs is laborious and inefficient due to the complex and high-dimensional nature of the configuration space. Among the automated tuning methods, reinforcement learning (RL)-based methods have recently sought to improve the DBMS knobs tuning process from several different perspectives. However, they still encounter challenges with slow convergence speed during offline training. In this paper, we mainly focus on how to leverage the valuable tuning hints contained in various textual documents such as DBMS manuals and web forums to improve the offline training of RL-based methods. To this end, we propose an efficient DBMS knobs tuning framework named DemoTuner via a novel LLM-assisted demonstration reinforcement learning method. Specifically, to comprehensively and accurately mine tuning hints from documents, we design a structured chain of thought prompt to employ LLMs to conduct a condition-aware tuning hints extraction task. To effectively integrate the mined tuning hints into RL agent training, we propose a hint-aware demonstration reinforcement learning algorithm HA-DDPGfD in DemoTuner. As far as we know, DemoTuner is the first work to introduce the demonstration reinforcement learning algorithm for DBMS knobs tuning. Experimental evaluations conducted on MySQL and PostgreSQL across various workloads demonstrate the significant advantages of DemoTuner in both performance improvement and online tuning cost reduction over three representative baselines including DB-BERT, GPTuner and CDBTune. Additionally, DemoTuner also exhibits superior adaptability to application scenarios with unknown workloads.

Method: Developed Interference Kernel Classifier (IKC) using coherent aggregation based on Born rule principles. Introduced Coherent Gain and Interference Information diagnostics. Used controlled phase sweeps and applied to XOR task and real tabular datasets (Adult, Bank Marketing).

Result: IKC outperformed baselines on XOR task. On real datasets, competitive overall but slightly trailed capacity-rich baselines. Toggling from incoherent to coherent aggregation consistently improved performance metrics (negative log-likelihood, Brier score, calibration error) with positive Coherent Gain.

Conclusion: The quantum-inspired coherent aggregation framework provides effective control over interaction effects, with IKC demonstrating practical utility and consistent performance improvements through coherent processing.

Abstract: Classical approaches often treat interaction as engineered product terms or as emergent patterns in flexible models, offering little control over how synergy or antagonism arises. We take a quantum-inspired view: following the Born rule (probability as squared amplitude), \emph{coherent} aggregation sums complex amplitudes before squaring, creating an interference cross-term, whereas an \emph{incoherent} proxy sums squared magnitudes and removes it. In a minimal linear-amplitude model, this cross-term equals the standard potential-outcome interaction contrast (Δ_{\mathrm{INT}}) in a (2\times 2) factorial design, giving relative phase a direct, mechanism-level control over synergy versus antagonism. We instantiate this idea in a lightweight \emph{Interference Kernel Classifier} (IKC) and introduce two diagnostics: \emph{Coherent Gain} (log-likelihood gain of coherent over the incoherent proxy) and \emph{Interference Information} (the induced Kullback-Leibler gap). A controlled phase sweep recovers the identity. On a high-interaction synthetic task (XOR), IKC outperforms strong baselines under paired, budget-matched comparisons; on real tabular data (\emph{Adult} and \emph{Bank Marketing}) it is competitive overall but typically trails the most capacity-rich baseline in paired differences. Holding learned parameters fixed, toggling aggregation from incoherent to coherent consistently improves negative log-likelihood, Brier score, and expected calibration error, with positive Coherent Gain on both datasets.

Method: Proposes GraphSB framework with Structural Balance strategy: Structure Enhancement builds similarity-based edges to strengthen minority-class connectivity, and Relation Diffusion captures higher-order dependencies while amplifying minority-class signals.

Result: Extensive experiments show GraphSB significantly outperforms state-of-the-art methods, and Structural Balance can be integrated as plug-and-play module to increase accuracy by average 3.67%.

Conclusion: GraphSB effectively balances structural distribution before node synthesis, enabling more effective learning in GNNs for imbalanced node classification tasks.

Abstract: Imbalanced node classification is a critical challenge in graph learning, where most existing methods typically utilize Graph Neural Networks (GNNs) to learn node representations. These methods can be broadly categorized into the data-level and the algorithm-level. The former aims to synthesize minority-class nodes to mitigate quantity imbalance, while the latter tries to optimize the learning process to highlight minority classes. However, neither category addresses the inherently imbalanced graph structure, which is a fundamental factor that incurs majority-class dominance and minority-class assimilation in GNNs. Our theoretical analysis further supports this critical insight. Therefore, we propose GraphSB (Graph Structural Balance), a novel framework that incorporates Structural Balance as a key strategy to address the underlying imbalanced graph structure before node synthesis. Structural Balance performs a two-stage structure optimization: Structure Enhancement that adaptively builds similarity-based edges to strengthen connectivity of minority-class nodes, and Relation Diffusion that captures higher-order dependencies while amplifying signals from minority classes. Thus, GraphSB balances structural distribution before node synthesis, enabling more effective learning in GNNs. Extensive experiments demonstrate that GraphSB significantly outperforms the state-of-the-art methods. More importantly, the proposed Structural Balance can be seamlessly integrated into state-of-the-art methods as a simple plug-and-play module, increasing their accuracy by an average of 3.67%.

Method: Parameterizes the kernel using SVD with two lightweight networks learning left/right singular functions, a diagonal matrix for singular values, and Gram-matrix regularization for orthonormality.

Result: Achieves new state-of-the-art performance on five diverse benchmark equations, with particularly strong gains on PDEs with highly spatially variable solutions.

Conclusion: SVD-NO provides a highly expressive yet computationally efficient neural operator framework that outperforms existing methods, especially for challenging PDE problems.

Abstract: Neural operators have emerged as a promising paradigm for learning solution operators of partial differential equa- tions (PDEs) directly from data. Existing methods, such as those based on Fourier or graph techniques, make strong as- sumptions about the structure of the kernel integral opera- tor, assumptions which may limit expressivity. We present SVD-NO, a neural operator that explicitly parameterizes the kernel by its singular-value decomposition (SVD) and then carries out the integral directly in the low-rank basis. Two lightweight networks learn the left and right singular func- tions, a diagonal parameter matrix learns the singular values, and a Gram-matrix regularizer enforces orthonormality. As SVD-NO approximates the full kernel, it obtains a high de- gree of expressivity. Furthermore, due to its low-rank struc- ture the computational complexity of applying the operator remains reasonable, leading to a practical system. In exten- sive evaluations on five diverse benchmark equations, SVD- NO achieves a new state of the art. In particular, SVD-NO provides greater performance gains on PDEs whose solutions are highly spatially variable. The code of this work is publicly available at https://github.com/2noamk/SVDNO.git.

Method: Developed a temporal latent variable structural causal model with causal strength and adjacency coefficients, and used Variational Inference with prior knowledge incorporation for parameter learning.

Result: Experimental results show the proposed method achieves stability and accuracy in causal relationship inference.

Conclusion: The method effectively handles external interferences in causal inference by incorporating latent variables and prior knowledge through Variational Inference.

Abstract: Inferring causal relationships from observed data is an important task, yet it becomes challenging when the data is subject to various external interferences. Most of these interferences are the additional effects of external factors on observed variables. Since these external factors are often unknown, we introduce latent variables to represent these unobserved factors that affect the observed data. Specifically, to capture the causal strength and adjacency information, we propose a new temporal latent variable structural causal model, incorporating causal strength and adjacency coefficients that represent the causal relationships between variables. Considering that expert knowledge can provide information about unknown interferences in certain scenarios, we develop a method that facilitates the incorporation of prior knowledge into parameter learning based on Variational Inference, to guide the model estimation. Experimental results demonstrate the stability and accuracy of our proposed method.

Method: The paper proposes a method to handle prefetching failures in MoE models without specifying the exact technique, but focusing on avoiding both PCIe latency stalls and accuracy degradation when experts aren’t prefetched correctly.

Result: The paper presents results showing improved handling of prefetch failures, likely demonstrating reduced latency while maintaining model accuracy compared to existing approaches that either stall or drop experts.

Conclusion: The work successfully addresses the critical challenge of maintaining both high inference speed and model accuracy in MoE architectures when prefetching fails, providing a solution to the trade-off between latency and accuracy in expert offloading systems.

Abstract: Mixture-of-Experts (MoE) architectures scale language models by activating only a subset of specialized expert networks for each input token, thereby reducing the number of floating-point operations. However, the growing size of modern MoE models causes their full parameter sets to exceed GPU memory capacity; for example, Mixtral-8x7B has 45 billion parameters and requires 87 GB of memory even though only 14 billion parameters are used per token. Existing systems alleviate this limitation by offloading inactive experts to CPU memory, but transferring experts across the PCIe interconnect incurs significant latency (about 10 ms). Prefetching heuristics aim to hide this latency by predicting which experts are needed, but prefetch failures introduce significant stalls and amplify inference latency. In the event of a prefetch failure, prior work offers two primary solutions: either fetch the expert on demand, which incurs a long stall due to the PCIe bottleneck, or drop the expert from the computation, which significantly degrades model accuracy. The critical challenge, therefore, is to maintain both high inference speed and model accuracy when prefetching fails.

Method: Analyzed structural vocabulary to identify semantic redundancy, then developed a ‘synonym swap’ strategy that perturbs predicted structures by replacing tokens with their structural synonyms to generate conformational ensembles.

Result: The synonym swap method accurately recapitulates protein flexibility and performs competitively with state-of-the-art models for modeling protein dynamics, while being computationally lightweight.

Conclusion: Semantic redundancy in structural tokens is not a flaw but can be exploited for efficient protein dynamics modeling, providing fundamental insights into discrete protein structure representations.

Abstract: Protein structure tokenization converts 3D structures into discrete or vectorized representations, enabling the integration of structural and sequence data. Despite many recent works on structure tokenization, the properties of the underlying discrete representations are not well understood. In this work, we first demonstrate that the successful utilization of structural tokens in a language model for structure prediction depends on using rich, pre-trained sequence embeddings to bridge the semantic gap between the sequence and structural “language”. The analysis of the structural vocabulary itself then reveals significant semantic redundancy, where multiple distinct tokens correspond to nearly identical local geometries, acting as “structural synonyms”. This redundancy, rather than being a flaw, can be exploited with a simple “synonym swap” strategy to generate diverse conformational ensembles by perturbing a predicted structure with its structural synonyms. This computationally lightweight method accurately recapitulates protein flexibility, performing competitively with state-of-the-art models. Our study provides fundamental insights into the nature of discrete protein structure representations and introduces a powerful, near-instantaneous method for modeling protein dynamics. Source code is available in https://github.com/IDEA-XL/TokenMD.

Method: FAQNAS formulates HQNN design as multi-objective optimization balancing accuracy and FLOPs, explicitly incorporating FLOPs into the optimization objective to discover architectures with strong performance and minimal computational cost.

Result: Experiments on five benchmark datasets show quantum FLOPs dominate accuracy improvements while classical FLOPs remain largely fixed. Pareto-optimal solutions achieve competitive accuracy with significantly reduced computational cost compared to FLOPs-agnostic baselines.

Conclusion: FLOPs-awareness is established as a practical criterion for HQNN design in the NISQ era and as a scalable principle for future HQNN systems.

Abstract: Hybrid Quantum Neural Networks (HQNNs), which combine parameterized quantum circuits with classical neural layers, are emerging as promising models in the noisy intermediate-scale quantum (NISQ) era. While quantum circuits are not naturally measured in floating point operations (FLOPs), most HQNNs (in NISQ era) are still trained on classical simulators where FLOPs directly dictate runtime and scalability. Hence, FLOPs represent a practical and viable metric to measure the computational complexity of HQNNs. In this work, we introduce FAQNAS, a FLOPs-aware neural architecture search (NAS) framework that formulates HQNN design as a multi-objective optimization problem balancing accuracy and FLOPs. Unlike traditional approaches, FAQNAS explicitly incorporates FLOPs into the optimization objective, enabling the discovery of architectures that achieve strong performance while minimizing computational cost. Experiments on five benchmark datasets (MNIST, Digits, Wine, Breast Cancer, and Iris) show that quantum FLOPs dominate accuracy improvements, while classical FLOPs remain largely fixed. Pareto-optimal solutions reveal that competitive accuracy can often be achieved with significantly reduced computational cost compared to FLOPs-agnostic baselines. Our results establish FLOPs-awareness as a practical criterion for HQNN design in the NISQ era and as a scalable principle for future HQNN systems.

Method: Tree-based Stochastic Optimization (TSO) uses tree-based action representation to map the entire action space onto a tree structure, incorporates this into the loss function, and employs a sample-and-prune mechanism to avoid local optima.

Result: Extensive experiments show TSO outperforms baseline algorithms in solving Urban Network Security Games, demonstrating superior performance in finding Nash Equilibria.

Conclusion: TSO effectively bridges stochastic optimization with tree-based representation to address the challenges of massive action spaces in Urban Network Security Games, providing a superior solution over existing methods.

Abstract: Urban Network Security Games (UNSGs), which model the strategic allocation of limited security resources on city road networks, are critical for urban safety. However, finding a Nash Equilibrium (NE) in large-scale UNSGs is challenging due to their massive and combinatorial action spaces. One common approach to addressing these games is the Policy-Space Response Oracle (PSRO) framework, which requires computing best responses (BR) at each iteration. However, precisely computing exact BRs is impractical in large-scale games, and employing reinforcement learning to approximate BRs inevitably introduces errors, which limits the overall effectiveness of the PSRO methods. Recent advancements in leveraging non-convex stochastic optimization to approximate an NE offer a promising alternative to the burdensome BR computation. However, utilizing existing stochastic optimization techniques with an unbiased loss function for UNSGs remains challenging because the action spaces are too vast to be effectively represented by neural networks. To address these issues, we introduce Tree-based Stochastic Optimization (TSO), a framework that bridges the gap between the stochastic optimization paradigm for NE-finding and the demands of UNSGs. Specifically, we employ the tree-based action representation that maps the whole action space onto a tree structure, addressing the challenge faced by neural networks in representing actions when the action space cannot be enumerated. We then incorporate this representation into the loss function and theoretically demonstrate its equivalence to the unbiased loss function. To further enhance the quality of the converged solution, we introduce a sample-and-prune mechanism that reduces the risk of being trapped in suboptimal local optima. Extensive experimental results indicate the superiority of TSO over other baseline algorithms in addressing the UNSGs.

Method: Single-step attack that modifies original explanations while being undetectable to human eye, using only model predictions and explanations without requiring model weights or architecture access.

Result: Successfully generated attacks that dramatically change explanations from saliency maps, integrated gradients, and DeepLIFT SHAP for ResNet-18 and ViT-B16 on ImageNet, without altering predictive probabilities.

Conclusion: The low requirements of this attack reveal critical vulnerabilities in current XAI methods, raising concerns about their reliability in safety-critical applications.

Abstract: Post-hoc explainability methods are a subset of Machine Learning (ML) that aim to provide a reason for why a model behaves in a certain way. In this paper, we show a new black-box model-agnostic adversarial attack for post-hoc explainable Artificial Intelligence (XAI), particularly in the image domain. The goal of the attack is to modify the original explanations while being undetected by the human eye and maintain the same predicted class. In contrast to previous methods, we do not require any access to the model or its weights, but only to the model’s computed predictions and explanations. Additionally, the attack is accomplished in a single step while significantly changing the provided explanations, as demonstrated by empirical evaluation. The low requirements of our method expose a critical vulnerability in current explainability methods, raising concerns about their reliability in safety-critical applications. We systematically generate attacks based on the explanations generated by post-hoc explainability methods (saliency maps, integrated gradients, and DeepLIFT SHAP) for pretrained ResNet-18 and ViT-B16 on ImageNet. The results show that our attacks could lead to dramatically different explanations without changing the predictive probabilities. We validate the effectiveness of our attack, compute the induced change based on the explanation with mean absolute difference, and verify the closeness of the original image and the corrupted one with the Structural Similarity Index Measure (SSIM).

Method: Created OutSafe-Bench dataset spanning 4 modalities with systematic annotation across 9 risk categories. Introduced MCRS metric for overlapping risk assessment and FairScore framework using top-performing models as adaptive juries for bias mitigation.

Result: Evaluation of 9 state-of-the-art MLLMs revealed persistent and substantial safety vulnerabilities, highlighting critical safety gaps in current multimodal models.

Conclusion: There is a pressing need for robust safeguards in MLLMs, as current models exhibit significant safety weaknesses that require comprehensive evaluation frameworks like OutSafe-Bench.

Abstract: Since Multimodal Large Language Models (MLLMs) are increasingly being integrated into everyday tools and intelligent agents, growing concerns have arisen regarding their possible output of unsafe contents, ranging from toxic language and biased imagery to privacy violations and harmful misinformation. Current safety benchmarks remain highly limited in both modality coverage and performance evaluations, often neglecting the extensive landscape of content safety. In this work, we introduce OutSafe-Bench, the first most comprehensive content safety evaluation test suite designed for the multimodal era. OutSafe-Bench includes a large-scale dataset that spans four modalities, featuring over 18,000 bilingual (Chinese and English) text prompts, 4,500 images, 450 audio clips and 450 videos, all systematically annotated across nine critical content risk categories. In addition to the dataset, we introduce a Multidimensional Cross Risk Score (MCRS), a novel metric designed to model and assess overlapping and correlated content risks across different categories. To ensure fair and robust evaluation, we propose FairScore, an explainable automated multi-reviewer weighted aggregation framework. FairScore selects top-performing models as adaptive juries, thereby mitigating biases from single-model judgments and enhancing overall evaluation reliability. Our evaluation of nine state-of-the-art MLLMs reveals persistent and substantial safety vulnerabilities, underscoring the pressing need for robust safeguards in MLLMs.

Method: Empirical study across five most popular open-source T2I models using ten neutral profession-related prompts, generating 100 images per profession (5,000 total images) evaluated by diverse human assessors representing different races and genders.

Result: All five models encode and amplify pronounced societal biases: caregiving roles consistently feminized, high-status professions overwhelmingly represented by males and mostly White individuals, with model-specific patterns (QWEN-Image focuses on East Asians, Kandinsky dominates White individuals, SDXL has broader but still biased distributions).

Conclusion: Results provide critical insights for AI project managers to select equitable models and customized prompts aligned with responsible AI principles, with discussion of bias risks and proposed actionable strategies for bias mitigation in GenAI systems.

Abstract: Text-to-image (T2I) generative models are largely used in AI-powered real-world applications and value creation. However, their strategic deployment raises critical concerns for responsible AI management, particularly regarding the reproduction and amplification of race- and gender-related stereotypes that can undermine organizational ethics. In this work, we investigate whether such societal biases are systematically encoded within the pretrained latent spaces of state-of-the-art T2I models. We conduct an empirical study across the five most popular open-source models, using ten neutral, profession-related prompts to generate 100 images per profession, resulting in a dataset of 5,000 images evaluated by diverse human assessors representing different races and genders. We demonstrate that all five models encode and amplify pronounced societal skew: caregiving and nursing roles are consistently feminized, while high-status professions such as corporate CEO, politician, doctor, and lawyer are overwhelmingly represented by males and mostly White individuals. We further identify model-specific patterns, such as QWEN-Image’s near-exclusive focus on East Asian outputs, Kandinsky’s dominance of White individuals, and SDXL’s comparatively broader but still biased distributions. These results provide critical insights for AI project managers and practitioners, enabling them to select equitable AI models and customized prompts that generate images in alignment with the principles of responsible AI. We conclude by discussing the risks of these biases and proposing actionable strategies for bias mitigation in building responsible GenAI systems.

Method: Extends Matryoshka representation learning to transcoder architectures for hierarchical sparse feature learning, trains on intermediate representations from physical plausibility classifiers, and uses large multimodal models for feature interpretation.

Result: Achieves superior feature relevance and accuracy compared to existing approaches, identifies diverse physics-related failure modes without manual feature engineering, and establishes a benchmark for evaluating physical plausibility.

Conclusion: The framework provides valuable insights into how state-of-the-art generative models fail to follow physical constraints, enabling targeted improvements and paving the way for more physically plausible AI systems.

Abstract: Although recent generative models are remarkably capable of producing instruction-following and realistic outputs, they remain prone to notable physical plausibility failures. Though critical in applications, these physical plausibility errors often escape detection by existing evaluation methods. Furthermore, no framework exists for automatically identifying and interpreting specific physical error patterns in natural language, preventing targeted model improvements. We introduce Matryoshka Transcoders, a novel framework for the automatic discovery and interpretation of physical plausibility features in generative models. Our approach extends the Matryoshka representation learning paradigm to transcoder architectures, enabling hierarchical sparse feature learning at multiple granularity levels. By training on intermediate representations from a physical plausibility classifier and leveraging large multimodal models for interpretation, our method identifies diverse physics-related failure modes without manual feature engineering, achieving superior feature relevance and feature accuracy compared to existing approaches. We utilize the discovered visual patterns to establish a benchmark for evaluating physical plausibility in generative models. Our analysis of eight state-of-the-art generative models provides valuable insights into how these models fail to follow physical constraints, paving the way for further model improvements.

Method: AgentEvolver introduces three synergistic mechanisms: self-questioning for curiosity-driven task generation, self-navigating for improved exploration efficiency through experience reuse, and self-attributing for enhanced sample efficiency via differentiated rewards.

Result: Preliminary experiments show AgentEvolver achieves more efficient exploration, better sample utilization, and faster adaptation compared to traditional RL-based baselines.

Conclusion: AgentEvolver enables scalable, cost-effective, and continual improvement of agent capabilities by leveraging LLMs’ semantic understanding and reasoning in a unified self-evolving framework.

Abstract: Autonomous agents powered by large language models (LLMs) have the potential to significantly enhance human productivity by reasoning, using tools, and executing complex tasks in diverse environments. However, current approaches to developing such agents remain costly and inefficient, as they typically require manually constructed task datasets and reinforcement learning (RL) pipelines with extensive random exploration. These limitations lead to prohibitively high data-construction costs, low exploration efficiency, and poor sample utilization. To address these challenges, we present AgentEvolver, a self-evolving agent system that leverages the semantic understanding and reasoning capabilities of LLMs to drive autonomous agent learning. AgentEvolver introduces three synergistic mechanisms: (i) self-questioning, which enables curiosity-driven task generation in novel environments, reducing dependence on handcrafted datasets; (ii) self-navigating, which improves exploration efficiency through experience reuse and hybrid policy guidance; and (iii) self-attributing, which enhances sample efficiency by assigning differentiated rewards to trajectory states and actions based on their contribution. By integrating these mechanisms into a unified framework, AgentEvolver enables scalable, cost-effective, and continual improvement of agent capabilities. Preliminary experiments indicate that AgentEvolver achieves more efficient exploration, better sample utilization, and faster adaptation compared to traditional RL-based baselines.

Method: Introduces Residual-Informed Loss (RI-Loss) based on Hilbert-Schmidt Independence Criterion (HSIC) that enforces dependence between residual sequence and random time series to model noise structure. Provides theoretical guarantees with non-asymptotic HSIC bounds and optimal convergence rates.

Result: Empirical experiments across eight real-world benchmarks and five leading forecasting models demonstrate improved predictive performance, validating the effectiveness of the approach.

Conclusion: RI-Loss enables more robust, noise-aware representations for time series forecasting, overcoming fundamental limitations of MSE while providing rigorous theoretical guarantees.

Abstract: Time series forecasting relies on predicting future values from historical data, yet most state-of-the-art approaches-including transformer and multilayer perceptron-based models-optimize using Mean Squared Error (MSE), which has two fundamental weaknesses: its point-wise error computation fails to capture temporal relationships, and it does not account for inherent noise in the data. To overcome these limitations, we introduce the Residual-Informed Loss (RI-Loss), a novel objective function based on the Hilbert-Schmidt Independence Criterion (HSIC). RI-Loss explicitly models noise structure by enforcing dependence between the residual sequence and a random time series, enabling more robust, noise-aware representations. Theoretically, we derive the first non-asymptotic HSIC bound with explicit double-sample complexity terms, achieving optimal convergence rates through Bernstein-type concentration inequalities and Rademacher complexity analysis. This provides rigorous guarantees for RI-Loss optimization while precisely quantifying kernel space interactions. Empirically, experiments across eight real-world benchmarks and five leading forecasting models demonstrate improvements in predictive performance, validating the effectiveness of our approach. Code will be made publicly available to ensure reproducibility.

Method: Analyzed the effects of removing LayerNorm parameters in both Pre- and Post-LayerNorm transformer architectures across 13 models and 6 Vision/Language datasets, focusing on memorization patterns and learning stability.

Result: Removing LayerNorm in Pre-LayerNorm models worsens memorization and destabilizes learning, while in Post-LayerNorm models it reduces memorization by restoring genuine labels. Early layers’ LayerNorm is most critical.

Conclusion: LayerNorm plays distinct roles in different transformer architectures - enabling stable learning in Pre-LayerNorm models and controlling memorization in Post-LayerNorm models, with early layers being particularly influential.

Abstract: Layer Normalization (LayerNorm) is one of the fundamental components in transformers that stabilizes training and improves optimization. In recent times, Pre-LayerNorm transformers have become the preferred choice over Post-LayerNorm transformers due to their stable gradient flow. However, the impact of LayerNorm on learning and memorization across these architectures remains unclear. In this work, we investigate how LayerNorm influences memorization and learning for Pre- and Post-LayerNorm transformers. We identify that LayerNorm serves as a key factor for stable learning in Pre-LayerNorm transformers, while in Post-LayerNorm transformers, it impacts memorization. Our analysis reveals that eliminating LayerNorm parameters in Pre-LayerNorm models exacerbates memorization and destabilizes learning, while in Post-LayerNorm models, it effectively mitigates memorization by restoring genuine labels. We further precisely identify that early layers LayerNorm are the most critical over middle/later layers and their influence varies across Pre and Post LayerNorm models. We have validated it through 13 models across 6 Vision and Language datasets. These insights shed new light on the role of LayerNorm in shaping memorization and learning in transformers.

Method: Energy Preference Optimization (EPO) uses stochastic differential equation sampling with a novel energy-ranking mechanism based on list-wise preference optimization, and introduces a practical upper bound to approximate intractable probabilities in continuous-time generative models.

Result: EPO establishes new state-of-the-art performance on Tetrapeptides, ATLAS, and Fast-Folding benchmarks across nine evaluation metrics, generating diverse and physically realistic ensembles.

Conclusion: Energy-only preference signals can efficiently steer generative models toward thermodynamically consistent conformational ensembles, providing an alternative to long MD simulations and expanding applications in structural biology and drug discovery.

Abstract: Accurate exploration of protein conformational ensembles is essential for uncovering function but remains hard because molecular-dynamics (MD) simulations suffer from high computational costs and energy-barrier trapping. This paper presents Energy Preference Optimization (EPO), an online refinement algorithm that turns a pretrained protein ensemble generator into an energy-aware sampler without extra MD trajectories. Specifically, EPO leverages stochastic differential equation sampling to explore the conformational landscape and incorporates a novel energy-ranking mechanism based on list-wise preference optimization. Crucially, EPO introduces a practical upper bound to efficiently approximate the intractable probability of long sampling trajectories in continuous-time generative models, making it easily adaptable to existing pretrained generators. On Tetrapeptides, ATLAS, and Fast-Folding benchmarks, EPO successfully generates diverse and physically realistic ensembles, establishing a new state-of-the-art in nine evaluation metrics. These results demonstrate that energy-only preference signals can efficiently steer generative models toward thermodynamically consistent conformational ensembles, providing an alternative to long MD simulations and widening the applicability of learned potentials in structural biology and drug discovery.

Method: QME embeds states into quantum-inspired representations using classically simulable unitary embeddings. The method was tested with SAC and IQL algorithms on three datasets with only 100 samples each.

Result: QME achieved significant performance improvements: 116.2% for SAC and 117.6% for IQL on average across datasets. The embedded states showed low Δ-hyperbolicity, indicating improved state space geometry.

Conclusion: QME effectively enhances offline RL with limited samples by modifying state space geometry through quantum-inspired embeddings, providing a promising approach for sample-efficient RL.

Abstract: Reinforcement learning (RL) with limited samples is common in real-world applications. However, offline RL performance under this constraint is often suboptimal. We consider an alternative approach to dealing with limited samples by introducing the Quantum Metric Encoder (QME). In this methodology, instead of applying the RL framework directly on the original states and rewards, we embed the states into a more compact and meaningful representation, where the structure of the encoding is inspired by quantum circuits. For classical data, QME is a classically simulable, trainable unitary embedding and thus serves as a quantum-inspired module, on a classical device. For quantum data in the form of quantum states, QME can be implemented directly on quantum hardware, allowing for training without measurement or re-encoding. We evaluated QME on three datasets, each limited to 100 samples. We use Soft-Actor-Critic (SAC) and Implicit-Q-Learning (IQL), two well-known RL algorithms, to demonstrate the effectiveness of our approach. From the experimental results, we find that training offline RL agents on QME-embedded states with decoded rewards yields significantly better performance than training on the original states and rewards. On average across the three datasets, for maximum reward performance, we achieve a 116.2% improvement for SAC and 117.6% for IQL. We further investigate the $Δ$-hyperbolicity of our framework, a geometric property of the state space known to be important for the RL training efficacy. The QME-embedded states exhibit low $Δ$-hyperbolicity, suggesting that the improvement after embedding arises from the modified geometry of the state space induced by QME. Thus, the low $Δ$-hyperbolicity and the corresponding effectiveness of QME could provide valuable information for developing efficient offline RL methods under limited-sample conditions.

Method: Used vector quantization and gumbel-softmax vector quantization of HuBERT embeddings to create discrete acoustic token sequences, then performed pairwise distance analysis and sequence classification using k-Nearest Neighbour with Levenshtein distance.

Result: Token sequences from HuBERT embeddings successfully discriminated call-types and callers across four bioacoustics datasets, achieving reasonable classification performance for both call-type and caller identification.

Conclusion: Vector-quantized token sequences show promise as alternative feature representations that can effectively leverage sequential information in animal vocalizations for bioacoustics analysis.

Abstract: Animal vocalizations contain sequential structures that carry important communicative information, yet most computational bioacoustics studies average the extracted frame-level features across the temporal axis, discarding the order of the sub-units within a vocalization. This paper investigates whether discrete acoustic token sequences, derived through vector quantization and gumbel-softmax vector quantization of extracted self-supervised speech model representations can effectively capture and leverage temporal information. To that end, pairwise distance analysis of token sequences generated from HuBERT embeddings shows that they can discriminate call-types and callers across four bioacoustics datasets. Sequence classification experiments using $k$-Nearest Neighbour with Levenshtein distance show that the vector-quantized token sequences yield reasonable call-type and caller classification performances, and hold promise as alternative feature representations towards leveraging sequential information in animal vocalizations.

Method: Discretizes observed values into ordered intervals, uses parametric distribution for probability supervision, predicts probability distributions with linear model, and applies OCE loss between cumulative distributions to preserve ordinal relationships.

Result: OCE-TS consistently outperformed five baseline models (Autoformer, DLinear, iTransformer, TimeXer, TimeBridge) on seven public datasets using MSE and MAE metrics.

Conclusion: The proposed OCE-TS approach successfully addresses MSE limitations by providing uncertainty quantification and improved robustness while achieving superior forecasting performance.

Abstract: Time series forecasting is an important task that involves analyzing temporal dependencies and underlying patterns (such as trends, cyclicality, and seasonality) in historical data to predict future values or trends. Current deep learning-based forecasting models primarily employ Mean Squared Error (MSE) loss functions for regression modeling. Despite enabling direct value prediction, this method offers no uncertainty estimation and exhibits poor outlier robustness. To address these limitations, we propose OCE-TS, a novel ordinal classification approach for time series forecasting that replaces MSE with Ordinal Cross-Entropy (OCE) loss, preserving prediction order while quantifying uncertainty through probability output. Specifically, OCE-TS begins by discretizing observed values into ordered intervals and deriving their probabilities via a parametric distribution as supervision signals. Using a simple linear model, we then predict probability distributions for each timestep. The OCE loss is computed between the cumulative distributions of predicted and ground-truth probabilities, explicitly preserving ordinal relationships among forecasted values. Through theoretical analysis using influence functions, we establish that cross-entropy (CE) loss exhibits superior stability and outlier robustness compared to MSE loss. Empirically, we compared OCE-TS with five baseline models-Autoformer, DLinear, iTransformer, TimeXer, and TimeBridge-on seven public time series datasets. Using MSE and Mean Absolute Error (MAE) as evaluation metrics, the results demonstrate that OCE-TS consistently outperforms benchmark models. The code will be published.

Method: FNA models token interactions through Lévy diffusion governed by the fractional Laplacian, intrinsically realizing multiscale information processing. The framework is theoretically grounded in fractional diffusion equations and uses diffusion map algorithms for dimensionality reduction.

Result: FNA achieves competitive text-classification performance with single layer/head configurations, improves performance in image processing and neural machine translation, exhibits larger spectral gaps and shorter path lengths for enhanced computational efficiency.

Conclusion: FNA establishes a principled mechanism connecting self-attention, stochastic dynamics, and geometry, providing an interpretable, biologically grounded foundation for neuroscience-inspired AI with enhanced expressivity and computational efficiency.

Abstract: Attention mechanisms underpin the computational power of Transformer models, which have achieved remarkable success across diverse domains. Yet understanding and extending the principles underlying self-attention remains a key challenge for advancing artificial intelligence. Drawing inspiration from the multiscale dynamics of biological attention and from dynamical systems theory, we introduce Fractional Neural Attention (FNA), a principled, neuroscience-inspired framework for multiscale information processing. FNA models token interactions through Lévy diffusion governed by the fractional Laplacian, intrinsically realizing simultaneous short- and long-range dependencies across multiple scales. This mechanism yields greater expressivity and faster information mixing, advancing the foundational capacity of Transformers. Theoretically, we show that FNA’s dynamics are governed by the fractional diffusion equation, and that the resulting attention networks exhibit larger spectral gaps and shorter path lengths – mechanistic signatures of enhanced computational efficiency. Empirically, FNA achieves competitive text-classification performance even with a single layer and a single head; it also improves performance in image processing and neural machine translation. Finally, the diffusion map algorithm from geometric harmonics enables dimensionality reduction of FNA weights while preserving the intrinsic structure of embeddings and hidden states. Together, these results establish FNA as a principled mechanism connecting self-attention, stochastic dynamics, and geometry, providing an interpretable, biologically grounded foundation for powerful, neuroscience-inspired AI.

Method: Proposes VDT with Domain-Invariant Variational Align module for joint encoding of source/target domains, domain consistency constraint for semantic preservation, and test-time training with confidence-variance filtering for dynamic adaptation.

Result: Extensive experiments on NewsCLIPpings dataset show VDT outperforms state-of-the-art baselines under most domain adaptation settings.

Conclusion: VDT effectively addresses domain adaptation challenges in OOC misinformation detection through domain-invariant feature learning and test-time training strategies.

Abstract: Out-of-context misinformation (OOC) is a low-cost form of misinformation in news reports, which refers to place authentic images into out-of-context or fabricated image-text pairings. This problem has attracted significant attention from researchers in recent years. Current methods focus on assessing image-text consistency or generating explanations. However, these approaches assume that the training and test data are drawn from the same distribution. When encountering novel news domains, models tend to perform poorly due to the lack of prior knowledge. To address this challenge, we propose \textbf{VDT} to enhance the domain adaptation capability for OOC misinformation detection by learning domain-invariant features and test-time training mechanisms. Domain-Invariant Variational Align module is employed to jointly encodes source and target domain data to learn a separable distributional space domain-invariant features. For preserving semantic integrity, we utilize domain consistency constraint module to reconstruct the source and target domain latent distribution. During testing phase, we adopt the test-time training strategy and confidence-variance filtering module to dynamically updating the VAE encoder and classifier, facilitating the model’s adaptation to the target domain distribution. Extensive experiments conducted on the benchmark dataset NewsCLIPpings demonstrate that our method outperforms state-of-the-art baselines under most domain adaptation settings.

Method: FedCure uses three key rules: (1) preference rule for coalition formation to maximize collective benefits and establish stable partitions, (2) scheduling rule integrating virtual queue technique with Bayesian-estimated coalition dynamics, and (3) resource allocation rule optimizing client CPU frequencies based on coalition dynamics.

Result: FedCure improves accuracy by up to 5.1x compared to four state-of-the-art baselines, achieves the lowest coefficient of variation (0.0223) for per-round latency, and maintains long-term balance across diverse scenarios.

Conclusion: FedCure effectively mitigates participation bias in semi-asynchronous federated learning with non-IID data through its coalition-based approach and dynamic scheduling mechanisms, demonstrating superior performance and efficiency across real-world datasets.

Abstract: While semi-asynchronous federated learning (SAFL) combines the efficiency of synchronous training with the flexibility of asynchronous updates, it inherently suffers from participation bias, which is further exacerbated by non-IID data distributions. More importantly, hierarchical architecture shifts participation from individual clients to client groups, thereby further intensifying this issue. Despite notable advancements in SAFL research, most existing works still focus on conventional cloud-end architectures while largely overlooking the critical impact of non-IID data on scheduling across the cloud-edge-client hierarchy. To tackle these challenges, we propose FedCure, an innovative semi-asynchronous Federated learning framework that leverages coalition construction and participation-aware scheduling to mitigate participation bias with non-IID data. Specifically, FedCure operates through three key rules: (1) a preference rule that optimizes coalition formation by maximizing collective benefits and establishing theoretically stable partitions to reduce non-IID-induced performance degradation; (2) a scheduling rule that integrates the virtual queue technique with Bayesian-estimated coalition dynamics, mitigating efficiency loss while ensuring mean rate stability; and (3) a resource allocation rule that enhances computational efficiency by optimizing client CPU frequencies based on estimated coalition dynamics while satisfying delay requirements. Comprehensive experiments on four real-world datasets demonstrate that FedCure improves accuracy by up to 5.1x compared with four state-of-the-art baselines, while significantly enhancing efficiency with the lowest coefficient of variation 0.0223 for per-round latency and maintaining long-term balance across diverse scenarios.

Method: Systematic analysis of LLM robustness by decomposing graph serializations into node labeling, edge encoding, and syntax components. Evaluation on comprehensive benchmarking suite with novel spectral tasks to assess generalization.

Result: Larger non-fine-tuned models are more robust. Fine-tuning reduces sensitivity to node relabeling but increases sensitivity to structural and formatting variations. No consistent improvement on unseen tasks.

Conclusion: Fine-tuning has mixed effects on LLM graph reasoners - it improves robustness to node relabeling but may degrade robustness to other variations, while not enhancing generalization to unseen tasks.

Abstract: While promising, graph reasoners based on Large Language Models (LLMs) lack built-in invariance to symmetries in graph representations. Operating on sequential graph serializations, LLMs can produce different outputs under node reindexing, edge reordering, or formatting changes, raising robustness concerns. We systematically analyze these effects, studying how fine-tuning impacts encoding sensitivity as well generalization on unseen tasks. We propose a principled decomposition of graph serializations into node labeling, edge encoding, and syntax, and evaluate LLM robustness to variations of each of these factors on a comprehensive benchmarking suite. We also contribute a novel set of spectral tasks to further assess generalization abilities of fine-tuned reasoners. Results show that larger (non-fine-tuned) models are more robust. Fine-tuning reduces sensitivity to node relabeling but may increase it to variations in structure and format, while it does not consistently improve performance on unseen tasks.

Method: Proposes Heuristic Transformer (HT) that uses a VAE to learn a low-dimensional stochastic variable representing the posterior distribution over rewards, which is incorporated alongside in-context dataset and query states as prompt to the transformer policy.

Result: HT consistently surpasses comparable baselines in effectiveness and generalization across Darkroom, Miniworld, and MuJoCo environments.

Conclusion: The method presents a promising direction for bridging belief-based augmentations with transformer-based decision-making in in-context reinforcement learning.

Abstract: Transformers have demonstrated exceptional in-context learning (ICL) capabilities, enabling applications across natural language processing, computer vision, and sequential decision-making. In reinforcement learning, ICL reframes learning as a supervised problem, facilitating task adaptation without parameter updates. Building on prior work leveraging transformers for sequential decision-making, we propose Heuristic Transformer (HT), an in-context reinforcement learning (ICRL) approach that augments the in-context dataset with a belief distribution over rewards to achieve better decision-making. Using a variational auto-encoder (VAE), a low-dimensional stochastic variable is learned to represent the posterior distribution over rewards, which is incorporated alongside an in-context dataset and query states as prompt to the transformer policy. We assess the performance of HT across the Darkroom, Miniworld, and MuJoCo environments, showing that it consistently surpasses comparable baselines in terms of both effectiveness and generalization. Our method presents a promising direction to bridge the gap between belief-based augmentations and transformer-based decision-making.

Method: Built on Transformer architecture and trained on a large-scale industrial lithography simulation dataset with hundreds of thousands of cases, supporting end-to-end mask generation, lithography simulation, and rule violation detection.

Result: Experimental results show Unitho substantially surpasses academic baselines in performance and generalizability, enabling agile and high-fidelity lithography simulation.

Conclusion: Unitho facilitates the construction of robust data foundations for intelligent EDA by providing a unified solution for multiple lithography tasks with superior performance.

Abstract: Reliable, generalizable data foundations are critical for enabling large-scale models in computational lithography. However, essential tasks-mask generation, rule violation detection, and layout optimization-are often handled in isolation, hindered by scarce datasets and limited modeling approaches. To address these challenges, we introduce Unitho, a unified multi-task large vision model built upon the Transformer architecture. Trained on a large-scale industrial lithography simulation dataset with hundreds of thousands of cases, Unitho supports end-to-end mask generation, lithography simulation, and rule violation detection. By enabling agile and high-fidelity lithography simulation, Unitho further facilitates the construction of robust data foundations for intelligent EDA. Experimental results validate its effectiveness and generalizability, with performance substantially surpassing academic baselines.

Method: Developed Torch-Uncertainty, a PyTorch and Lightning-based framework that streamlines DNN training and evaluation with UQ techniques and metrics.

Result: Comprehensive experimental results benchmarking diverse UQ methods across classification, segmentation, and regression tasks.

Conclusion: Torch-Uncertainty bridges the gap by providing a unified tool for evaluating and integrating UQ methods in deep learning.

Abstract: Deep Neural Networks (DNNs) have demonstrated remarkable performance across various domains, including computer vision and natural language processing. However, they often struggle to accurately quantify the uncertainty of their predictions, limiting their broader adoption in critical real-world applications. Uncertainty Quantification (UQ) for Deep Learning seeks to address this challenge by providing methods to improve the reliability of uncertainty estimates. Although numerous techniques have been proposed, a unified tool offering a seamless workflow to evaluate and integrate these methods remains lacking. To bridge this gap, we introduce Torch-Uncertainty, a PyTorch and Lightning-based framework designed to streamline DNN training and evaluation with UQ techniques and metrics. In this paper, we outline the foundational principles of our library and present comprehensive experimental results that benchmark a diverse set of UQ methods across classification, segmentation, and regression tasks. Our library is available at https://github.com/ENSTA-U2IS-AI/Torch-Uncertainty

Method: Define prototypes as cluster centroids for similar individuals under same treatment, perform instance-to-prototype matching within groups, and use multi-prototype alignment strategy to bring matched prototypes closer across treatment arms in latent space.

Result: Extensive evaluations on benchmark datasets show PITE outperforms 13 state-of-the-art methods, achieving more accurate and robust ITE estimation.

Conclusion: PITE effectively reduces distribution shift through fine-grained prototype-level alignment while preserving local structures, providing meaningful constraints for improved ITE estimation.

Abstract: Estimating Individual Treatment Effects (ITE) from observational data is challenging due to confounding bias. Most studies tackle this bias by balancing distributions globally, but ignore individual heterogeneity and fail to capture the local structure that represents the natural clustering among individuals, which ultimately compromises ITE estimation. While instance-level alignment methods consider heterogeneity, they similarly overlook the local structure information. To address these issues, we propose an end-to-end Multi-\textbf{P}rototype alignment method for \textbf{ITE} estimation (\textbf{PITE}). PITE effectively captures local structure within groups and enforces cross-group alignment, thereby achieving robust ITE estimation. Specifically, we first define prototypes as cluster centroids based on similar individuals under the same treatment. To identify local similarity and the distribution consistency, we perform instance-to-prototype matching to assign individuals to the nearest prototype within groups, and design a multi-prototype alignment strategy to encourage the matched prototypes to be close across treatment arms in the latent space. PITE not only reduces distribution shift through fine-grained, prototype-level alignment, but also preserves the local structures of treated and control groups, which provides meaningful constraints for ITE estimation. Extensive evaluations on benchmark datasets demonstrate that PITE outperforms 13 state-of-the-art methods, achieving more accurate and robust ITE estimation.

Method: Three components: 1) Down-sampling to efficiently estimate gradient entropy, 2) Theoretical model linking compression rate with gradient entropy, 3) Window-based adjustment mechanism that dynamically adapts compression rates across pipeline stages.

Result: EDGC reduces communication latency by up to 46.45% and training time by 16.13% while preserving LLM accuracy, tested on 32-NVIDIA-V100 and 64-NVIDIA-H100 clusters for GPT2-2.5B and GPT2-12.1B models.

Conclusion: EDGC successfully addresses the challenge of accelerating LLM training through dynamic gradient compression that adapts to gradient entropy changes, achieving significant communication efficiency gains without performance loss.

Abstract: Training large language models (LLMs) poses significant challenges regarding computational resources and memory capacity. Although distributed training techniques help mitigate these issues, they still suffer from considerable communication overhead. Existing approaches primarily rely on static gradient compression to enhance communication efficiency; however, these methods neglect the dynamic nature of evolving gradients during training, leading to performance degradation. Accelerating LLM training via compression without sacrificing performance remains a challenge. In this paper, we propose an entropy-driven dynamic gradient compression framework called EDGC. The core concept is to adjust the compression rate during LLM training based on the evolving trends of gradient entropy, taking into account both compression efficiency and error. EDGC consists of three key components.First, it employs a down-sampling method to efficiently estimate gradient entropy, reducing computation overhead. Second, it establishes a theoretical model linking compression rate with gradient entropy, enabling more informed compression decisions. Lastly, a window-based adjustment mechanism dynamically adapts the compression rate across pipeline stages, improving communication efficiency and maintaining model performance. We implemented EDGC on a 32-NVIDIA-V100 cluster and a 64-NVIDIA-H100 cluster to train GPT2-2.5B and GPT2-12.1B, respectively. The results show that EDGC significantly reduces communication latency and training time by up to 46.45% and 16.13% while preserving LLM accuracy.

Method: Developed DeMABAR algorithm that uses robust mechanisms to handle adversarial corruption of reward observations and Byzantine agents that may arbitrarily select arms and communicate wrong information.

Result: Theoretical analysis shows DeMABAR limits individual regret to an additive term proportional to corruption budget, and can almost completely eliminate influence of attacks when adversary targets few agents. Numerical experiments confirm robustness.

Conclusion: DeMABAR provides effective robustness against adversarial corruption and Byzantine attacks in decentralized multi-agent bandit settings, with minimal impact on regret performance.

Abstract: Decentralized cooperative multi-agent multi-armed bandits (DeCMA2B) considers how multiple agents collaborate in a decentralized multi-armed bandit setting. Though this problem has been extensively studied in previous work, most existing methods remain susceptible to various adversarial attacks. In this paper, we first study DeCMA2B with adversarial corruption, where an adversary can corrupt reward observations of all agents with a limited corruption budget. We propose a robust algorithm, called DeMABAR, which ensures that each agent’s individual regret suffers only an additive term proportional to the corruption budget. Then we consider a more realistic scenario where the adversary can only attack a small number of agents. Our theoretical analysis shows that the DeMABAR algorithm can also almost completely eliminate the influence of adversarial attacks and is inherently robust in the Byzantine setting, where an unknown fraction of the agents can be Byzantine, i.e., may arbitrarily select arms and communicate wrong information. We also conduct numerical experiments to illustrate the robustness and effectiveness of the proposed method.

Method: Analysis using adjacency matrix representation to formulate gradient flow equations as converging matrix ODEs that are nilpotent, polynomial, isospectral with conservation laws.

Result: Loss landscape contains infinitely many global minima and saddle points (both strict and nonstrict), no local minima/maxima. Loss function is Lyapunov function with unbounded level sets as invariant critical sets. Critical values correspond to learned singular values.

Conclusion: Adjacency matrix representation reveals quotient space structure where each critical value appears once, enabling determination of stable/unstable manifolds even when Hessian fails, providing comprehensive understanding of deep linear network dynamics.

Abstract: The paper surveys recent progresses in understanding the dynamics and loss landscape of the gradient flow equations associated to deep linear neural networks, i.e., the gradient descent training dynamics (in the limit when the step size goes to 0) of deep neural networks missing the activation functions and subject to quadratic loss functions. When formulated in terms of the adjacency matrix of the neural network, as we do in the paper, these gradient flow equations form a class of converging matrix ODEs which is nilpotent, polynomial, isospectral, and with conservation laws. The loss landscape is described in detail. It is characterized by infinitely many global minima and saddle points, both strict and nonstrict, but lacks local minima and maxima. The loss function itself is a positive semidefinite Lyapunov function for the gradient flow, and its level sets are unbounded invariant sets of critical points, with critical values that correspond to the amount of singular values of the input-output data learnt by the gradient along a certain trajectory. The adjacency matrix representation we use in the paper allows to highlight the existence of a quotient space structure in which each critical value of the loss function is represented only once, while all other critical points with the same critical value belong to the fiber associated to the quotient space. It also allows to easily determine stable and unstable submanifolds at the saddle points, even when the Hessian fails to obtain them.

Method: Developed an efficient algorithm that uses the advice distribution Q’s parameters. The algorithm works when ||p-q||₁ < εd^(0.5-Ω(η)), without requiring this bound to be known beforehand.

Result: Achieves sample complexity Õ(d^(1-η)/ε²) for learning P within TV distance ε, which is better than the standard Ω(d/ε²) when η > 0.

Conclusion: Advice distributions can significantly reduce sample complexity for learning product distributions when the target and advice distributions are sufficiently close in ℓ₁ distance of their mean vectors.

Abstract: Given i.i.d.~samples from an unknown distribution $P$, the goal of distribution learning is to recover the parameters of a distribution that is close to $P$. When $P$ belongs to the class of product distributions on the Boolean hypercube ${0,1}^d$, it is known that $Ω(d/\varepsilon^2)$ samples are necessary to learn $P$ within total variation (TV) distance $\varepsilon$. We revisit this problem when the learner is also given as advice the parameters of a product distribution $Q$. We show that there is an efficient algorithm to learn $P$ within TV distance $\varepsilon$ that has sample complexity $\tilde{O}(d^{1-η}/\varepsilon^2)$, if $|\mathbf{p} - \mathbf{q}|_1 < \varepsilon d^{0.5 - Ω(η)}$. Here, $\mathbf{p}$ and $\mathbf{q}$ are the mean vectors of $P$ and $Q$ respectively, and no bound on $|\mathbf{p} - \mathbf{q}|_1$ is known to the algorithm a priori.

Method: RFF-KPKM uses random Fourier features to generate efficient low-dimensional feature maps, avoiding full kernel matrices. IP-RFF-MKPKM combines possibilistic and fuzzy membership for improved robustness and multiple kernel learning.

Result: Theoretical guarantees include: excess risk bound of O(√(k³/n)), strong consistency with membership values, and (1+ε) relative error bound with poly(ε⁻¹log k) RFF dimension. Experiments show superior efficiency and clustering accuracy on large datasets.

Conclusion: The proposed methods overcome KPKM’s scalability and robustness limitations through RFF approximation and enhanced membership learning, achieving state-of-the-art performance with strong theoretical foundations.

Abstract: Kernel power $k$-means (KPKM) leverages a family of means to mitigate local minima issues in kernel $k$-means. However, KPKM faces two key limitations: (1) the computational burden of the full kernel matrix restricts its use on extensive data, and (2) the lack of authentic centroid-sample assignment learning reduces its noise robustness. To overcome these challenges, we propose RFF-KPKM, introducing the first approximation theory for applying random Fourier features (RFF) to KPKM. RFF-KPKM employs RFF to generate efficient, low-dimensional feature maps, bypassing the need for the whole kernel matrix. Crucially, we are the first to establish strong theoretical guarantees for this combination: (1) an excess risk bound of $\mathcal{O}(\sqrt{k^3/n})$, (2) strong consistency with membership values, and (3) a $(1+\varepsilon)$ relative error bound achievable using the RFF of dimension $\mathrm{poly}(\varepsilon^{-1}\log k)$. Furthermore, to improve robustness and the ability to learn multiple kernels, we propose IP-RFF-MKPKM, an improved possibilistic RFF-based multiple kernel power $k$-means. IP-RFF-MKPKM ensures the scalability of MKPKM via RFF and refines cluster assignments by combining the merits of the possibilistic membership and fuzzy membership. Experiments on large-scale datasets demonstrate the superior efficiency and clustering accuracy of the proposed methods compared to the state-of-the-art alternatives.

Method: FedSTGD uses a federated nonlinear computation decomposition module to approximate graph operations, a graph node embedding augmentation module to mitigate decomposition effects, and a client-server collective learning protocol that breaks down dependency learning into parallelizable subtasks.

Result: Extensive experiments on four real-world datasets show FedSTGD outperforms state-of-the-art baselines in RMSE, MAE, and MAPE metrics, approaching centralized baseline performance. Ablation studies confirm each module’s contribution.

Conclusion: FedSTGD effectively addresses dynamic inter-client spatial dependencies in federated learning settings while maintaining data privacy, demonstrating robustness to hyperparameter variations and achieving near-centralized performance.

Abstract: Spatio-temporal graphs are powerful tools for modeling complex dependencies in traffic time series. However, the distributed nature of real-world traffic data across multiple stakeholders poses significant challenges in modeling and reconstructing inter-client spatial dependencies while adhering to data locality constraints. Existing methods primarily address static dependencies, overlooking their dynamic nature and resulting in suboptimal performance. In response, we propose Federated Spatio-Temporal Graph with Dynamic Inter-Client Dependencies (FedSTGD), a framework designed to model and reconstruct dynamic inter-client spatial dependencies in federated learning. FedSTGD incorporates a federated nonlinear computation decomposition module to approximate complex graph operations. This is complemented by a graph node embedding augmentation module, which alleviates performance degradation arising from the decomposition. These modules are coordinated through a client-server collective learning protocol, which decomposes dynamic inter-client spatial dependency learning tasks into lightweight, parallelizable subtasks. Extensive experiments on four real-world datasets demonstrate that FedSTGD achieves superior performance over state-of-the-art baselines in terms of RMSE, MAE, and MAPE, approaching that of centralized baselines. Ablation studies confirm the contribution of each module in addressing dynamic inter-client spatial dependencies, while sensitivity analysis highlights the robustness of FedSTGD to variations in hyperparameters.

Method: Trained neural networks with different learning rates and analyzed the resulting weight and bias fluctuations, correlating these patterns with final model accuracy.

Result: Established a clear link between learning rate values, parameter fluctuation patterns, and overall model performance, showing how learning rate mediates the exploration-exploitation balance.

Conclusion: Provides deeper insights into hyperparameter tuning and deep learning mechanics by revealing how learning rate influences neural fluctuations and optimization dynamics.

Abstract: Deep Neural Networks (DNNs) rely on inherent fluctuations in their internal parameters (weights and biases) to effectively navigate the complex optimization landscape and achieve robust performance. While these fluctuations are recognized as crucial for escaping local minima and improving generalization, their precise relationship with fundamental hyperparameters remains underexplored. A significant knowledge gap exists concerning how the learning rate, a critical parameter governing the training process, directly influences the dynamics of these neural fluctuations. This study systematically investigates the impact of varying learning rates on the magnitude and character of weight and bias fluctuations within a neural network. We trained a model using distinct learning rates and analyzed the corresponding parameter fluctuations in conjunction with the network’s final accuracy. Our findings aim to establish a clear link between the learning rate’s value, the resulting fluctuation patterns, and overall model performance. By doing so, we provide deeper insights into the optimization process, shedding light on how the learning rate mediates the crucial exploration-exploitation trade-off during training. This work contributes to a more nuanced understanding of hyperparameter tuning and the underlying mechanics of deep learning.

Method: Introduces ReCalX, a novel approach to recalibrate models for improved explanations while preserving original predictions. The method addresses perturbation-specific miscalibration through calibration techniques.

Result: Empirical evaluations across diverse models and datasets show ReCalX consistently reduces perturbation-specific miscalibration most effectively, enhances explanation robustness, and improves identification of globally important input features.

Conclusion: Model calibration under explainability-specific perturbations is crucial for reliable explanations. ReCalX successfully addresses this issue by recalibrating models to improve explanation quality without compromising original prediction performance.

Abstract: Perturbation-based explanations are widely utilized to enhance the transparency of machine-learning models in practice. However, their reliability is often compromised by the unknown model behavior under the specific perturbations used. This paper investigates the relationship between uncertainty calibration - the alignment of model confidence with actual accuracy - and perturbation-based explanations. We show that models systematically produce unreliable probability estimates when subjected to explainability-specific perturbations and theoretically prove that this directly undermines global and local explanation quality. To address this, we introduce ReCalX, a novel approach to recalibrate models for improved explanations while preserving their original predictions. Empirical evaluations across diverse models and datasets demonstrate that ReCalX consistently reduces perturbation-specific miscalibration most effectively while enhancing explanation robustness and the identification of globally important input features.

Method: Use Intrinsic Dimensionality (ID) as a measure of class imbalance that can be incorporated into various imbalance mitigation methods. Test across five datasets with diverse imbalance ratios.

Result: ID consistently outperforms cardinality-based re-weighting and re-sampling techniques. Combining ID with cardinality further improves performance.

Conclusion: Intrinsic Dimensionality is an effective, model-free measure for class imbalance that provides better performance than traditional cardinality-based approaches and can be enhanced by combining with cardinality.

Abstract: Imbalance in classification tasks is commonly quantified by the cardinalities of examples across classes. This, however, disregards the presence of redundant examples and inherent differences in the learning difficulties of classes. Alternatively, one can use complex measures such as training loss and uncertainty, which, however, depend on training a machine learning model. Our paper proposes using data Intrinsic Dimensionality (ID) as an easy-to-compute, model-free measure of imbalance that can be seamlessly incorporated into various imbalance mitigation methods. Our results across five different datasets with a diverse range of imbalance ratios show that ID consistently outperforms cardinality-based re-weighting and re-sampling techniques used in the literature. Moreover, we show that combining ID with cardinality can further improve performance. Code: https://github.com/cagries/IDIM.

Method: Panda uses negative data augmentation by dividing images into patches and randomly assembling them from a shared pool, disrupting semantic content while retaining corruption features. It subtracts mean features of negative samples from original image features to suppress corruption-related components.

Result: Panda delivers superior performance compared to PDA methods, significantly enhances various TTA methods when integrated, and achieves minimal computational overhead by sharing augmentations across batches.

Conclusion: Panda effectively mitigates prediction bias under distribution shifts, can be seamlessly integrated into existing TTA frameworks, and substantially improves VLM robustness against image corruptions with minimal computational cost.

Abstract: Pretrained VLMs exhibit strong zero-shot classification capabilities, but their predictions degrade significantly under common image corruptions. To improve robustness, many test-time adaptation (TTA) methods adopt positive data augmentation (PDA), which generates multiple views of each test sample to reduce prediction variance. However, these methods suffer from two key limitations. First, it introduces considerable computational overhead due to the large number of augmentations required per image. Second, it fails to mitigate prediction bias, where the model tends to predict certain classes disproportionately under corruption, as PDA operates on corrupted inputs and typically does not remove the corruption itself. To address these challenges, we propose Panda, a novel TTA method based on negative data augmentation (NDA). Unlike positive augmentations that preserve object semantics, Panda generates negative augmentations by disrupting semantic content. It divides images into patches and randomly assembles them from a shared patch pool. These negatively augmented images retain corruption-specific features while discarding object-relevant signals. We then subtract the mean feature of these negative samples from the original image feature, effectively suppressing corruption-related components while preserving class-relevant information. This mitigates prediction bias under distribution shifts. Panda allows augmentation to be shared across samples within a batch, resulting in minimal computational overhead. Panda can be seamlessly integrated into existing test-time adaptation frameworks and substantially improve their robustness. Our experiments indicate that Panda delivers superior performance compared to PDA methods, and a wide range of TTA methods exhibit significantly enhanced performance when integrated with Panda. Our code is available at https://github.com/ruxideng/Panda .

Method: Kinematic-Informed Neural Networks (KINN) with loss function that penalizes errors between predictions and supervised labels, plus errors between next point prediction and previous point plus velocity prediction.

Result: Tested on multiple AR ANN architectures with 15 years of Dow Jones data, showing statistically meaningful improvement across normalization-sensitive activation functions prone to spurious behavior in OOD conditions.

Conclusion: The architecture addresses normalization issues in auto-regressive models by weakly enforcing data neighborhood proximity preservation during ANN transformation.

Abstract: We propose loss function week enforcement of the velocity relations between time-series points in the Kinematic-Informed artificial Neural Networks (KINN) for long-term stock prediction. Problems of the series volatility, Out-of-Distribution (OOD) test data, and outliers in training data are addressed by (Artificial Neural Networks) ANN’s learning not only future points prediction but also by learning velocity relations between the points, such a way as avoiding unrealistic spurious predictions. The presented loss function penalizes not only errors between predictions and supervised label data, but also errors between the next point prediction and the previous point plus velocity prediction. The loss function is tested on the multiple popular and exotic AR ANN architectures, and around fifteen years of Dow Jones function demonstrated statistically meaningful improvement across the normalization-sensitive activation functions prone to spurious behaviour in the OOD data conditions. Results show that such architecture addresses the issue of the normalization in the auto-regressive models that break the data topology by weakly enforcing the data neighbourhood proximity (relation) preservation during the ANN transformation.

Method: Proposed Holonorm with residual connections and nonlinearity, which maps vectors into the open unit ball to prevent exploding activations. It’s a generalized form of softsign function suitable for high-dimensional tensors and vectors.

Result: Holonorm preserves signal properties (orthogonality, direction, invertibility) and improves stability in deep Transformer models by preventing exploding activations.

Conclusion: Holonorm serves as an effective normalization function that is more reliable than Tanh-based approaches, with intuitive interpretability as a percentage metric between 0 and 1.

Abstract: Normalization is a key point in transformer training . In Dynamic Tanh (DyT), the author demonstrated that Tanh can be used as an alternative layer normalization (LN) and confirmed the effectiveness of the idea. But Tanh itself faces orthogonality, linearity and distortion problems. Due to that, his proposition cannot be reliable. So we propose a Holonorm (hn) which has residual connections and nonlinearity. Holonorm is suitable for replacing Tanh in the context of normalization. Although the HoloNorm expression could be similar to the softsign function in dimension one, softsign is a componentwise function which is not good for tensors and vectors of great dimension. Holonorm preserves the orthogonality, the direction, the invertibility of the signal. Holonorm is also a suitable metric, maps all vectors into the open unit ball. This prevents exploding activations and improves stability in deep Transformer models. In this work, we have meticulously examined the normalization in transformers and say that Holonorm, a generalized form of softsign function suited as a normalization function first.Second, defined between 0 and 1 hn serves as a percentage, and $1 - \text{Holonorm}$ is its complement, making it better understandable in evaluating a model.

Method: Frames dataset compression as a minimum set cover problem over atom-centered environments, identifying the smallest subset of structures that contains maximum information from the original dataset.

Result: Outperforms other subsampling methods across 64 datasets, consistently retains outliers, preserves diversity, reproduces long-tail force distributions even at high compression rates, and reduces error for out-of-distribution data.

Conclusion: The MSC approach provides efficient dataset compression for improved MLIP training at lower cost, with implementation available in the open-source QUESTS package.

Abstract: Machine learning interatomic potentials (MLIPs) balance high accuracy and lower costs compared to density functional theory calculations, but their performance often depends on the size and diversity of training datasets. Large datasets improve model accuracy and generalization but are computationally expensive to produce and train on, while smaller datasets risk discarding rare but important atomic environments and compromising MLIP accuracy/reliability. Here, we develop an information-theoretical framework to quantify the efficiency of dataset compression methods and propose an algorithm that maximizes this efficiency. By framing atomistic dataset compression as an instance of the minimum set cover (MSC) problem over atom-centered environments, our method identifies the smallest subset of structures that contains as much information as possible from the original dataset while pruning redundant information. The approach is extensively demonstrated on the GAP-20 and TM23 datasets, and validated on 64 varied datasets from the ColabFit repository. Across all cases, MSC consistently retains outliers, preserves dataset diversity, and reproduces the long-tail distributions of forces even at high compression rates, outperforming other subsampling methods. Furthermore, MLIPs trained on MSC-compressed datasets exhibit reduced error for out-of-distribution data even in low-data regimes. We explain these results using an outlier analysis and show that such quantitative conclusions could not be achieved with conventional dimensionality reduction methods. The algorithm is implemented in the open-source QUESTS package and can be used for several tasks in atomistic modeling, from data subsampling, outlier detection, and training improved MLIPs at a lower cost.

Method: Two-stage deep learning with U-Net models: one for precipitation detection and another for quantitative estimation, trained on GPM CORRA v07 data and pre-trained on IMERG-Final retrievals.

Result: Achieves quasi-global coverage and superior performance compared to existing regional and global precipitation baselines.

Conclusion: Oya offers a promising pathway to improved precipitation monitoring and forecasting, particularly for hydrological applications in the Global South.

Abstract: Accurate precipitation estimation is critical for hydrological applications, especially in the Global South where ground-based observation networks are sparse and forecasting skill is limited. Existing satellite-based precipitation products often rely on the longwave infrared channel alone or are calibrated with data that can introduce significant errors, particularly at sub-daily timescales. This study introduces Oya, a novel real-time precipitation retrieval algorithm utilizing the full spectrum of visible and infrared (VIS-IR) observations from geostationary (GEO) satellites. Oya employs a two-stage deep learning approach, combining two U-Net models: one for precipitation detection and another for quantitative precipitation estimation (QPE), to address the inherent data imbalance between rain and no-rain events. The models are trained using high-resolution GPM Combined Radar-Radiometer Algorithm (CORRA) v07 data as ground truth and pre-trained on IMERG-Final retrievals to enhance robustness and mitigate overfitting due to the limited temporal sampling of CORRA. By leveraging multiple GEO satellites, Oya achieves quasi-global coverage and demonstrates superior performance compared to existing competitive regional and global precipitation baselines, offering a promising pathway to improved precipitation monitoring and forecasting.

Method: Belief Net formulates the HMM’s forward filter as a structured neural network where learnable weights are explicitly the logits of the initial distribution, transition matrix, and emission matrix. It uses a decoder-only architecture trained end-to-end with standard autoregressive next-observation prediction loss.

Result: On synthetic HMM data, Belief Net achieves superior convergence speed compared to Baum-Welch and successfully recovers parameters in both undercomplete and overcomplete settings where spectral methods fail. Comparisons with Transformer-based models are also presented on real-world language data.

Conclusion: Belief Net provides an efficient, interpretable alternative to classical HMM learning methods, combining the benefits of gradient-based optimization with explicit parameter interpretability while handling challenging settings where existing methods struggle.

Abstract: Hidden Markov Models (HMMs) are fundamental for modeling sequential data, yet learning their parameters from observations remains challenging. Classical methods like the Baum-Welch (EM) algorithm are computationally intensive and prone to local optima, while modern spectral algorithms offer provable guarantees but may produce probability outputs outside valid ranges. This work introduces Belief Net, a novel framework that learns HMM parameters through gradient-based optimization by formulating the HMM’s forward filter as a structured neural network. Unlike black-box Transformer models, Belief Net’s learnable weights are explicitly the logits of the initial distribution, transition matrix, and emission matrix, ensuring full interpretability. The model processes observation sequences using a decoder-only architecture and is trained end-to-end with standard autoregressive next-observation prediction loss. On synthetic HMM data, Belief Net achieves superior convergence speed compared to Baum-Welch, successfully recovering parameters in both undercomplete and overcomplete settings where spectral methods fail. Comparisons with Transformer-based models are also presented on real-world language data.

Method: Integrates strict Top-K LISTA and convex FISTA-based LISTAConv into discriminative LC-KSVD2 model, enabling co-evolution between sparse encoder and dictionary under supervised/unsupervised regimes.

Result: Achieves 95.6% on CIFAR-10, 86.3% on CIFAR-100, and 88.5% on TinyImageNet with faster convergence and lower memory cost (<4GB GPU).

Conclusion: The LC-KSVD2 + LISTA/LISTAConv pipeline offers an interpretable and computationally efficient alternative for modern deep architectures.

Abstract: We present a semi-unified sparse dictionary learning framework that bridges the gap between classical sparse models and modern deep architectures. Specifically, the method integrates strict Top-$K$ LISTA and its convex FISTA-based variant (LISTAConv) into the discriminative LC-KSVD2 model, enabling co-evolution between the sparse encoder and the dictionary under supervised or unsupervised regimes. This unified design retains the interpretability of traditional sparse coding while benefiting from efficient, differentiable training. We further establish a PALM-style convergence analysis for the convex variant, ensuring theoretical stability under block alternation. Experimentally, our method achieves 95.6% on CIFAR-10, 86.3% on CIFAR-100, and 88.5% on TinyImageNet with faster convergence and lower memory cost ($<$4GB GPU). The results confirm that the proposed LC-KSVD2 + LISTA/LISTAConv pipeline offers an interpretable and computationally efficient alternative for modern deep architectures.

Method: The authors show that the convex hull of an l0-ball is the intersection of its bounding box and an asymmetrically scaled l1-like polytope. They develop a linear bound propagation that precisely computes bounds over this convex hull.

Result: The proposed bound propagation scales the state-of-the-art l0 verifier by 1.24x-7.07x on challenging robustness benchmarks, with a geometric mean improvement of 3.16x.

Conclusion: The method enables efficient verification of neural network robustness against few-pixel attacks by leveraging the geometric properties of l0-ball convex hulls and developing tight linear bound propagations.

Abstract: Few-pixel attacks mislead a classifier by modifying a few pixels of an image. Their perturbation space is an $\ell_0$-ball, which is not convex, unlike $\ell_p$-balls for $p\geq1$. However, existing local robustness verifiers typically scale by relying on linear bound propagation, which captures convex perturbation spaces. We show that the convex hull of an $\ell_0$-ball is the intersection of its bounding box and an asymmetrically scaled $\ell_1$-like polytope. The volumes of the convex hull and this polytope are nearly equal as the input dimension increases. We then show a linear bound propagation that precisely computes bounds over the convex hull and is significantly tighter than bound propagations over the bounding box or our $\ell_1$-like polytope. This bound propagation scales the state-of-the-art $\ell_0$ verifier on its most challenging robustness benchmarks by 1.24x-7.07x, with a geometric mean of 3.16.

Method: Using Epistemic Neural Networks (ENNs) to obtain scalable joint predictive distributions of binding affinity, with investigation into prior networks and pretraining on synthetic data to improve Batch Bayesian Optimization performance.

Result: The method rediscovered known potent EGFR inhibitors in up to 5x fewer iterations on semi-synthetic benchmarks, and potent inhibitors from real-world small-molecule libraries in up to 10x fewer iterations.

Conclusion: The approach offers a promising solution for large-scale drug discovery applications by significantly reducing the number of iterations needed to discover potent inhibitors.

Abstract: Batched synthesis and testing of molecular designs is the key bottleneck of drug development. There has been great interest in leveraging biomolecular foundation models as surrogates to accelerate this process. In this work, we show how to obtain scalable probabilistic surrogates of binding affinity for use in Batch Bayesian Optimization (Batch BO). This demands parallel acquisition functions that hedge between designs and the ability to rapidly sample from a joint predictive density to approximate them. Through the framework of Epistemic Neural Networks (ENNs), we obtain scalable joint predictive distributions of binding affinity on top of representations taken from large structure-informed models. Key to this work is an investigation into the importance of prior networks in ENNs and how to pretrain them on synthetic data to improve downstream performance in Batch BO. Their utility is demonstrated by rediscovering known potent EGFR inhibitors on a semi-synthetic benchmark in up to 5x fewer iterations, as well as potent inhibitors from a real-world small-molecule library in up to 10x fewer iterations, offering a promising solution for large-scale drug discovery applications.

Method: Two parameterized families of bandit algorithms are proposed: one includes the optimal randomized algorithm from prior work, and the other provides algorithms that guarantee best-arm identification on well-behaved instances while maintaining worst-case guarantees on poorly-behaved instances. The sample complexity of learning near-optimal algorithms from each family using offline data is bounded.

Result: The first family achieves stronger guarantees with optimal dependence on k when arm reward curves satisfy additional concavity properties. The second family provides data-dependent guarantees without needing to verify whether the assumptions are satisfied.

Conclusion: By taking a statistical learning perspective and using parameterized algorithm families, the work achieves improved performance guarantees for improving multi-armed bandits while maintaining robustness across different instance types.

Abstract: The improving multi-armed bandits problem is a formal model for allocating effort under uncertainty, motivated by scenarios such as investing research effort into new technologies, performing clinical trials, and hyperparameter selection from learning curves. Each pull of an arm provides reward that increases monotonically with diminishing returns. A growing line of work has designed algorithms for improving bandits, albeit with somewhat pessimistic worst-case guarantees. Indeed, strong lower bounds of $Ω(k)$ and $Ω(\sqrt{k})$ multiplicative approximation factors are known for both deterministic and randomized algorithms (respectively) relative to the optimal arm, where $k$ is the number of bandit arms. In this work, we propose two new parameterized families of bandit algorithms and bound the sample complexity of learning the near-optimal algorithm from each family using offline data. The first family we define includes the optimal randomized algorithm from prior work. We show that an appropriately chosen algorithm from this family can achieve stronger guarantees, with optimal dependence on $k$, when the arm reward curves satisfy additional properties related to the strength of concavity. Our second family contains algorithms that both guarantee best-arm identification on well-behaved instances and revert to worst case guarantees on poorly-behaved instances. Taking a statistical learning perspective on the bandit rewards optimization problem, we achieve stronger data-dependent guarantees without the need for actually verifying whether the assumptions are satisfied.

Method: Treats black-box LLM as environment, uses M-Pilot as policy to provide intermediate guidance through prompts, trains M-Pilot to align black-box outputs with preferences during iterative interaction.

Result: Empirical evaluations show M-Pilot effectively enhances black-box LLM capabilities in complex, long-horizon tasks.

Conclusion: M-Pilot enables controllable generation and self-improvement for black-box LLMs through iterative guidance decomposition.

Abstract: Despite the impressive generative abilities of black-box large language models (LLMs), their inherent opacity hinders further advancements in capabilities such as reasoning, planning, and personalization. Existing works aim to enhance LLM capabilities via domain-specific adaptation, which require additional training on accessible model parameters, an infeasible option for black-box LLMs. To address this challenge, we introduce Matryoshka Pilot (M-Pilot), a lightweight white-box LLM controller that guides a large-scale black-box LLM generator by decomposing complex tasks into a series of intermediate outputs. Specifically, we consider the black-box LLM as an environment, with M-Pilot serving as a policy to provide intermediate guidance through prompts for driving the black-box LLM. M-Pilot is trained to pivot the outputs of the black-box LLM aligning with preferences during iterative interaction, which enables controllable multi-turn generation and self-improvement in optimizing intermediate guidance. Empirical evaluations on diverse tasks demonstrate that our method effectively enhances the capabilities of black-box LLMs in complex, long-horizon tasks. Our code is publicly available at: https://github.com/lichangh20/Matryoshka.

Method: Theoretical analysis comparing transformers with leftmost-hard attention versus rightmost-hard attention, examining their computational equivalence to Linear Temporal Logic fragments and soft attention models.

Result: Leftmost-hard attention transformers correspond to a strictly weaker fragment of LTL compared to rightmost-hard attention models, and are equivalent to soft attention transformers.

Conclusion: The directionality of hard attention (leftmost vs rightmost) significantly impacts transformer expressivity, with leftmost-hard attention being strictly weaker and better approximating real-world transformers through its equivalence to soft attention.

Abstract: Understanding the expressive power of transformers has recently attracted attention, as it offers insights into their abilities and limitations. Many studies analyze unique hard attention transformers, where attention selects a single position that maximizes the attention scores. When multiple positions achieve the maximum score, either the rightmost or the leftmost of those is chosen. In this paper, we highlight the importance of this seeming triviality. Recently, finite-precision transformers with both leftmost- and rightmost-hard attention were shown to be equivalent to Linear Temporal Logic (LTL). We show that this no longer holds with only leftmost-hard attention – in that case, they correspond to a \emph{strictly weaker} fragment of LTL. Furthermore, we show that models with leftmost-hard attention are equivalent to \emph{soft} attention, suggesting they may better approximate real-world transformers than right-attention models. These findings refine the landscape of transformer expressivity and underscore the role of attention directionality.

Method: Performs PCA on activations and module-output gradients to identify subspaces with common representations, then collapses these subspaces before computing unlearning updates (Collapse of Irrelevant Representations).

Result: Achieved over 30x greater reduction in post-attack accuracy than Circuit Breakers baseline, disrupted general performance 30x less, and used less than 3 GPU-seconds per fact when unlearning bio- and cyber-hazardous facts from Llama-3.1-8B.

Conclusion: By disentangling harmful and benign capabilities at the representation level, CIR enables robust and non-disruptive unlearning.

Abstract: Current unlearning and safety training methods consistently fail to remove dangerous knowledge from language models. We identify the root cause - unlearning targets representations which are too general - and develop a highly selective technique that unlearns robustly while preserving general performance. Our method performs PCA on activations and module-output gradients to identify subspaces containing common representations, then collapses these subspaces before computing unlearning updates, a technique we term Collapse of Irrelevant Representations (CIR). This avoids unlearning general knowledge and targets only representations specific to the facts being unlearned. When unlearning bio- and cyber-hazardous facts from Llama-3.1-8B, we achieve over 30x greater reduction in post-attack accuracy than the best baseline (Circuit Breakers), while disrupting general performance 30x less, and using less than 3 GPU-seconds per fact. Thus, by disentangling harmful and benign capabilities at the level of representations, CIR enables robust and non-disruptive unlearning.

Method: Propose Markovian Thinking paradigm with Delethink environment that structures reasoning into fixed-size chunks, using RL to learn state carryover between chunks for seamless continuation.

Result: 1.5B model achieves 24K token reasoning with 8K chunks, matching 24K LongCoT-RL performance while reducing compute from 27 to 7 H100-months at 96K thinking length.

Conclusion: Redesigning the thinking environment enables efficient long reasoning without quadratic overhead, opening path to scalable reasoning LLMs with linear compute scaling.

Abstract: Reinforcement learning (RL) has recently become a strong recipe for training reasoning LLMs that produce long chains of thought (LongCoT). Yet the standard RL “thinking environment”, where the state is the prompt plus all prior reasoning tokens, makes the state unbounded and forces attention-based policies to pay quadratic compute as thoughts lengthen. We revisit the environment itself. We propose Markovian Thinking, a paradigm in which the policy advances reasoning while conditioning on a constant-size state, decoupling thinking length from context size. As an immediate consequence this yields linear compute with constant memory. We instantiate this idea with Delethink, an RL environment that structures reasoning into fixed-size chunks. Within each chunk, the model thinks as usual; at the boundary, the environment resets the context and reinitializes the prompt with a short carryover. Through RL, the policy learns to write a textual state near the end of each chunk sufficient for seamless continuation of reasoning after reset. Trained in this environment, an R1-Distill 1.5B model reasons in 8K-token chunks yet thinks up to 24K tokens, matching or surpassing LongCoT-RL trained with a 24K budget. With test-time scaling, Delethink continues to improve where LongCoT plateaus. The effect of linear compute is substantial: we empirically estimate at 96K average thinking length LongCoT-RL costs 27 H100-months vs. 7 for Delethink. Analysis at RL initialization shows off-the-shelf reasoning models (1.5B-120B) often sample Markovian traces zero-shot across diverse benchmarks, providing positive samples that make RL effective at scale. Our results show that redesigning the thinking environment is a powerful lever: it enables very long reasoning without quadratic overhead and opens a path toward efficient, scalable reasoning LLMs.

Method: GRACE measures distributional properties of student gradients without needing teacher logits, internals, or test data, connecting to information-theoretic principles of gradient-based algorithm stability.

Result: On GSM8K and MATH benchmarks, GRACE shows strong correlation (up to 86% Spearman) with distilled student performance and improves results by up to 7.4% over naive teacher selection.

Conclusion: GRACE efficiently identifies compatible teachers and provides fine-grained guidance on distillation design choices including temperature settings, size constraints, and model family selection.

Abstract: Knowledge distillation is an efficient strategy to use data generated by large “teacher” language models to train smaller capable “student” models, but selecting the optimal teacher for a specific student-task combination requires expensive trial-and-error. We propose a lightweight score called GRACE to quantify how effective a teacher will be for post-training a student model. GRACE measures distributional properties of the student’s gradients without access to a verifier, teacher logits, teacher internals, or test data. From an information-theoretic perspective, GRACE connects to leave-one-out stability of gradient-based algorithms, which controls the generalization performance of the distilled students. On GSM8K and MATH, GRACE correlates strongly (up to 86% Spearman correlation) with the performance of the distilled LLaMA and OLMo students. In particular, training a student using the GRACE-selected teacher can improve the performance by up to 7.4% over naively using the best-performing teacher. Further, GRACE can provide guidance on crucial design choices in distillation, including (1) the best temperature to use when generating from the teacher, (2) the best teacher to use given a size constraint, and (3) the best teacher to use within a specific model family. Altogether, our findings demonstrate that GRACE can efficiently and effectively identify a strongly compatible teacher for a given student and provide fine-grained guidance on how to perform distillation.

Method: PerturbODE incorporates biologically informative neural ordinary differential equations (neural ODEs) to model cell state trajectories under perturbations and derives the causal gene regulatory network from the neural ODE’s parameters.

Result: PerturbODE demonstrates efficacy in trajectory prediction and gene regulatory network inference across both simulated and real over-expression datasets.

Conclusion: The proposed PerturbODE framework successfully addresses the limitations of existing methods by leveraging neural ODEs to capture dynamic biological processes and infer causal gene regulatory networks from large-scale perturbation data.

Abstract: Modern high-throughput biological datasets with thousands of perturbations provide the opportunity for large-scale discovery of causal graphs that represent the regulatory interactions between genes. Differentiable causal graphical models have been proposed to infer a gene regulatory network (GRN) from large scale interventional datasets, capturing the causal gene regulatory relationships from genetic perturbations. However, existing models are limited in their expressivity and scalability while failing to address the dynamic nature of biological processes such as cellular differentiation. We propose PerturbODE, a novel framework that incorporates biologically informative neural ordinary differential equations (neural ODEs) to model cell state trajectories under perturbations and derive the causal GRN from the neural ODE’s parameters. We demonstrate PerturbODE’s efficacy in trajectory prediction and GRN inference across simulated and real over-expression datasets.

Method: A novel method that efficiently constructs an optimal behavior basis, which is proven to be more expressive than Convex Coverage Sets and requires fewer base policies.

Result: The method significantly reduces the number of base policies needed for optimality, handles non-linear tasks, and outperforms state-of-the-art approaches, especially in complex domains.

Conclusion: The proposed optimal behavior basis enables zero-shot identification of optimal solutions for linear tasks and extends to certain non-linear tasks, demonstrating superior performance over existing methods.

Abstract: Multi-task reinforcement learning aims to quickly identify solutions for new tasks with minimal or no additional interaction with the environment. Generalized Policy Improvement (GPI) addresses this by combining a set of base policies to produce a new one that is at least as good – though not necessarily optimal – as any individual base policy. Optimality can be ensured, particularly in the linear-reward case, via techniques that compute a Convex Coverage Set (CCS). However, these are computationally expensive and do not scale to complex domains. The Option Keyboard (OK) improves upon GPI by producing policies that are at least as good – and often better. It achieves this through a learned meta-policy that dynamically combines base policies. However, its performance critically depends on the choice of base policies. This raises a key question: is there an optimal set of base policies – an optimal behavior basis – that enables zero-shot identification of optimal solutions for any linear tasks? We solve this open problem by introducing a novel method that efficiently constructs such an optimal behavior basis. We show that it significantly reduces the number of base policies needed to ensure optimality in new tasks. We also prove that it is strictly more expressive than a CCS, enabling particular classes of non-linear tasks to be solved optimally. We empirically evaluate our technique in challenging domains and show that it outperforms state-of-the-art approaches, increasingly so as task complexity increases.

Method: Reimplemented 18 studies (2013-2023) using balanced dataset of 124,000 apps in contemporary Android environment, comparing static vs dynamic features and model complexity.

Result: Feature-based approaches achieve >98% accuracy despite larger Android feature sets; static features nearly match dynamic ones; simpler models often outperform complex ones; API calls and opcodes are most productive static features; network traffic is most predictive dynamic feature.

Conclusion: Simple, fast ML approaches remain effective for malware detection, challenging the literature’s focus on slower, more expensive models.

Abstract: We report the findings of a reimplementation of 18 foundational studies in feature-based machine learning for Android malware detection, published during the period 2013-2023. These studies are reevaluated on a level playing field using a contemporary Android environment and a balanced dataset of 124,000 applications. Our findings show that feature-based approaches can still achieve detection accuracies beyond 98%, despite a considerable increase in the size of the underlying Android feature sets. We observe that features derived through dynamic analysis yield only a small benefit over those derived from static analysis, and that simpler models often out-perform more complex models. We also find that API calls and opcodes are the most productive static features within our evaluation context, network traffic is the most predictive dynamic feature, and that ensemble models provide an efficient means of combining models trained on static and dynamic features. Together, these findings suggest that simple, fast machine learning approaches can still be an effective basis for malware detection, despite the increasing focus on slower, more expensive machine learning models in the literature.

Method: OODTE applies differential testing by running original and optimized models on user-defined inputs, iteratively isolating problematic optimization passes when discrepancies occur.

Result: Evaluation of 130 models showed 9.2% caused crashes or invalid models, 30% of classification models and 16.6% of vision models had output discrepancies, and 15 issues (14 new) were found affecting 9 optimization passes.

Conclusion: OODTE provides an effective framework for validating AI model optimizers, revealing significant reliability issues in the widely-used ONNX Optimizer that need addressing.

Abstract: With over 760 stars on GitHub and being part of the official ONNX repository, the ONNX Optimizer is the default tool for applying graph-based optimizations to ONNX models. Despite its widespread use, its ability to maintain model accuracy during optimization has not been thoroughly investigated. In this work, we present OODTE, a utility designed to automatically and comprehensively evaluate the correctness of the ONNX Optimizer. OODTE adopts a straightforward yet powerful differential testing and evaluation methodology, which can be readily adapted for use with other compiler optimizers. Specifically, OODTE takes a collection of ONNX models, applies optimizations, and executes both the original and optimized versions across a user-defined input set, automatically capturing any issues encountered during optimization. When discrepancies in accuracy arise, OODTE iteratively isolates the responsible optimization pass by repeating the process at a finer granularity. We applied OODTE to 130 well-known models from the official ONNX Model Hub, spanning diverse tasks including classification, object detection, semantic segmentation, text summarization, question answering, and sentiment analysis. Our evaluation revealed that 9.2% of the model instances either caused the optimizer to crash or led to the generation of invalid models using default optimization strategies. Additionally, 30% of classification models and 16.6% of object detection and segmentation models exhibited differing outputs across original and optimized versions, whereas models focused on text-related tasks were generally robust to optimization. OODTE uncovered 15 issues-14 previously unknown-affecting 9 of 47 optimization passes and the optimizer overall. All issues were reported to the ONNX Optimizer team. OODTE offers a simple but effective framework for validating AI model optimizers, applicable beyond the ONNX ecosystem.

Method: Provides efficient implementations of state-of-the-art global and regional feature effects methods under unified API, integrating with scikit-learn and PyTorch.

Result: Package offers modular, extensible design with comprehensive documentation and tutorials for interpreting ML models on tabular data.

Conclusion: Effector is an open-source tool that addresses limitations of global feature effects by incorporating regional analysis for more accurate model interpretation.

Abstract: Effector is a Python package for interpreting machine learning (ML) models that are trained on tabular data through global and regional feature effects. Global effects, like Partial Dependence Plot (PDP) and Accumulated Local Effects (ALE), are widely used for explaining tabular ML models due to their simplicity – each feature’s average influence on the prediction is summarized by a single 1D plot. However, when features are interacting, global effects can be misleading. Regional effects address this by partitioning the input space into disjoint subregions with minimal interactions within each and computing a separate regional effect per subspace. Regional effects are then visualized by a set of 1D plots per feature. Effector provides efficient implementations of state-of-the-art global and regional feature effects methods under a unified API. The package integrates seamlessly with major ML libraries like scikit-learn and PyTorch. It is designed to be modular and extensible, and comes with comprehensive documentation and tutorials. Effector is an open-source project publicly available on Github at https://github.com/givasile/effector.

Method: Combines PPO with PGD for adversarial state calculation and Lipschitz-regularized critic to enhance policy smoothness and robustness.

Result: Achieves better performance and smoother action predictions under environmental perturbations on control tasks and real-world robotic locomotion.

Conclusion: PPO-PGDLC effectively improves policy robustness through the integration of adversarial training and critic regularization.

Abstract: Uncertainties in transition dynamics pose a critical challenge in reinforcement learning (RL), often resulting in performance degradation of trained policies when deployed on hardware. Many robust RL approaches follow two strategies: enforcing smoothness in actor or actor-critic modules with Lipschitz regularization, or learning robust Bellman operators. However, the first strategy does not investigate the impact of critic-only Lipschitz regularization on policy robustness, while the second lacks comprehensive validation in real-world scenarios. Building on this gap and prior work, we propose PPO-PGDLC, an algorithm based on Proximal Policy Optimization (PPO) that integrates Projected Gradient Descent (PGD) with a Lipschitz-regularized critic (LC). The PGD component calculates the adversarial state within an uncertainty set to approximate the robust Bellman operator, and the Lipschitz-regularized critic further improves the smoothness of learned policies. Experimental results on two classic control tasks and one real-world robotic locomotion task demonstrates that, compared to several baseline algorithms, PPO-PGDLC achieves better performance and predicts smoother actions under environmental perturbations.

Method: Apply exponentially weighted average (EWA) or randomized weighted majority (RWM) forecasters on samples using log loss, using regret bounds to derive sample complexity bounds.

Result: New sample-optimal polynomial time algorithm for learning trees of unknown structure, and first polynomial sample/time algorithm for learning Bayes nets over chordal skeletons.

Conclusion: Online learning framework provides computationally efficient methods for learning Bayes nets with improved sample complexity bounds.

Abstract: This work establishes a novel link between the problem of PAC-learning high-dimensional graphical models and the task of (efficient) counting and sampling of graph structures, using an online learning framework. We observe that if we apply the exponentially weighted average (EWA) or randomized weighted majority (RWM) forecasters on a sequence of samples from a distribution P using the log loss function, the average regret incurred by the forecaster’s predictions can be used to bound the expected KL divergence between P and the predictions. Known regret bounds for EWA and RWM then yield new sample complexity bounds for learning Bayes nets. Moreover, these algorithms can be made computationally efficient for several interesting classes of Bayes nets. Specifically, we give a new sample-optimal and polynomial time learning algorithm with respect to trees of unknown structure and the first polynomial sample and time algorithm for learning with respect to Bayes nets over a given chordal skeleton.

Method: Leverages pre-trained vision-language models (BLIP for caption generation and stable diffusion for sample generation) with dedicated Task Heads and a Task Router that learns to assign inputs using generated data.

Result: Improves average task accuracy by up to 54% and achieves 63 times better memory efficiency compared to four recent continual learning baselines on three benchmark datasets.

Conclusion: CLTS introduces a novel perspective by integrating generative text-image augmentation for scalable continual learning, demonstrating improved retention and adaptability.

Abstract: Continual learning (CL) enables models to adapt to evolving data streams without catastrophic forgetting, a fundamental requirement for real-world AI systems. However, the current methods often depend on large replay buffers or heavily annotated datasets which are impractical due to storage, privacy, and cost constraints. We propose CLTS (Continual Learning via Text-Image Synergy), a novel class-incremental framework that mitigates forgetting without storing real task data. CLTS leverages pre-trained vision-language models, BLIP (Bootstrapping Language-Image Pre-training) for caption generation and stable diffusion for sample generation. Each task is handled by a dedicated Task Head, while a Task Router learns to assign inputs to the correct Task Head using the generated data. On three benchmark datasets, CLTS improves average task accuracy by up to 54% and achieves 63 times better memory efficiency compared to four recent continual learning baselines, demonstrating improved retention and adaptability. CLTS introduces a novel perspective by integrating generative text-image augmentation for scalable continual learning.

Method: Extensive exploration of MIA vulnerability across various decentralized learning architectures by varying graph structure, dynamics, aggregation strategies, datasets, and data distributions, with theoretical analysis of mixing properties.

Result: Vulnerability to MIA is strongly correlated with local model mixing strategies and global communication graph mixing properties. Enhanced mixing properties combined with Differential Privacy significantly reduce MIA risk.

Conclusion: Decentralized learning systems should be designed with strong mixing properties to reduce MIA vulnerability by default, and mixing enhancement works synergistically with other privacy-preserving techniques.

Abstract: The primary promise of decentralized learning is to allow users to engage in the training of machine learning models in a collaborative manner while keeping their data on their premises and without relying on any central entity. However, this paradigm necessitates the exchange of model parameters or gradients between peers. Such exchanges can be exploited to infer sensitive information about training data, which is achieved through privacy attacks (e.g., Membership Inference Attacks – MIA). In order to devise effective defense mechanisms, it is important to understand the factors that increase/reduce the vulnerability of a given decentralized learning architecture to MIA. In this study, we extensively explore the vulnerability to MIA of various decentralized learning architectures by varying the graph structure (e.g., number of neighbors), the graph dynamics, and the aggregation strategy, across diverse datasets and data distributions. Our key finding, which to the best of our knowledge we are the first to report, is that the vulnerability to MIA is heavily correlated to (i) the local model mixing strategy performed by each node upon reception of models from neighboring nodes and (ii) the global mixing properties of the communication graph. We illustrate these results experimentally using four datasets and by theoretically analyzing the mixing properties of various decentralized architectures. We also empirically show that enhancing mixing properties is highly beneficial when combined with other privacy-preserving techniques such as Differential Privacy. Our paper draws a set of lessons learned for devising decentralized learning systems that reduce by design the vulnerability to MIA.

Method: Three key components: 1) Neighbor Hidden Style Enhancement with style transfer for temporal-then-spatial virtual node representation, 2) Virtual-Component Contrastive Learning to enrich node representations, 3) Similarity-Based Graph Denoising Strategy to reduce noisy connections.

Result: Extensive experiments demonstrate that DarkFarseer significantly outperforms existing ISK methods.

Conclusion: DarkFarseer effectively addresses limitations in current ISK methods by improving virtual sensor representation and handling graph connectivity issues, enabling better sensor network coverage.

Abstract: With the rapid growth of the Internet of Things and Cyber-Physical Systems, widespread sensor deployment has become essential. However, the high costs of building sensor networks limit their scale and coverage, making fine-grained deployment challenging. Inductive Spatio-Temporal Kriging (ISK) addresses this issue by introducing virtual sensors. Based on graph neural networks (GNNs) extracting the relationships between physical and virtual sensors, ISK can infer the measurements of virtual sensors from physical sensors. However, current ISK methods rely on conventional message-passing mechanisms and network architectures, without effectively extracting spatio-temporal features of physical sensors and focusing on representing virtual sensors. Additionally, existing graph construction methods face issues of sparse and noisy connections, destroying ISK performance. To address these issues, we propose DarkFarseer, a novel ISK framework with three key components. First, we propose the Neighbor Hidden Style Enhancement module with a style transfer strategy to enhance the representation of virtual nodes in a temporal-then-spatial manner to better extract the spatial relationships between physical and virtual nodes. Second, we propose Virtual-Component Contrastive Learning, which aims to enrich the node representation by establishing the association between the patterns of virtual nodes and the regional patterns within graph components. Lastly, we design a Similarity-Based Graph Denoising Strategy, which reduces the connectivity strength of noisy connections around virtual nodes and their neighbors based on their temporal information and regional spatial patterns. Extensive experiments demonstrate that DarkFarseer significantly outperforms existing ISK methods.

Method: Developed PISA algorithm based on preconditioned inexact stochastic ADMM that converges under Lipschitz continuity assumption alone, supports parallel computing, and incorporates various preconditions including second-order information, second moment, and orthogonalized momentum.

Result: PISA’s variants SISA and NSISA demonstrated superior performance across diverse deep learning models (vision, LLMs, RL, GANs, RNNs) compared to state-of-the-art optimizers.

Conclusion: PISA provides an effective solution for training foundation models with data heterogeneity, offering strong theoretical guarantees and practical performance improvements over existing methods.

Abstract: The recent advancement of foundation models (FMs) has brought about a paradigm shift, revolutionizing various sectors worldwide. The popular optimizers used to train these models are stochastic gradient descent-based algorithms, which face inherent limitations, such as slow convergence and stringent assumptions for convergence. In particular, data heterogeneity arising from distributed settings poses significant challenges to their theoretical and numerical performance. This paper develops an algorithm, PISA (Preconditioned Inexact Stochastic Alternating Direction Method of Multipliers). Grounded in rigorous theoretical guarantees, the algorithm converges under the sole assumption of Lipschitz continuity of the gradient on a bounded region, thereby removing the need for other conditions commonly imposed by stochastic methods. This capability enables the proposed algorithm to tackle the challenge of data heterogeneity effectively. Moreover, the algorithmic architecture enables scalable parallel computing and supports various preconditions, such as second-order information, second moment, and orthogonalized momentum by Newton-Schulz iterations. Incorporating the latter two preconditions in PISA yields two computationally efficient variants: SISA and NSISA. Comprehensive experimental evaluations for training or fine-tuning diverse deep models, including vision models, large language models, reinforcement learning models, generative adversarial networks, and recurrent neural networks, demonstrate superior numerical performance of SISA and NSISA compared to various state-of-the-art optimizers.

Method: Proposes xLSTMAD with full encoder-decoder xLSTM architecture. Two variants: forecasting (xLSTMAD-F) generates future values, reconstruction (xLSTMAD-R) reconstructs input sequences. Uses MSE and SoftDTW loss functions for local and global alignment respectively.

Result: Outperforms 23 popular anomaly detection baselines on TSB-AD-M benchmark spanning 17 real-world datasets, achieving state-of-the-art accuracy with challenging metrics like VUS-PR.

Conclusion: xLSTM demonstrates powerful modeling capabilities for anomaly detection, paving the way for new developments in this area. The work successfully establishes xLSTM as a viable architecture for anomaly detection tasks.

Abstract: The recently proposed xLSTM is a powerful model that leverages expressive multiplicative gating and residual connections, providing the temporal capacity needed for long-horizon forecasting and representation learning. This architecture has demonstrated success in time series forecasting, lossless compression, and even large-scale language modeling tasks, where its linear memory footprint and fast inference make it a viable alternative to Transformers. Despite its growing popularity, no prior work has explored xLSTM for anomaly detection. In this work, we fill this gap by proposing xLSTMAD, the first anomaly detection method that integrates a full encoder-decoder xLSTM architecture, purpose-built for multivariate time series data. Our encoder processes input sequences to capture historical context, while the decoder is devised in two separate variants of the method. In the forecasting approach, the decoder iteratively generates forecasted future values xLSTMAD-F, while the reconstruction approach reconstructs the input time series from its encoded counterpart xLSTMAD-R. We investigate the performance of two loss functions: Mean Squared Error (MSE), and Soft Dynamic Time Warping (SoftDTW) to consider local reconstruction fidelity and global sequence alignment, respectively. We evaluate our method on the comprehensive TSB-AD-M benchmark, which spans 17 real-world datasets, using state-of-the-art challenging metrics such as VUS-PR. In our results, xLSTM showcases state-of-the-art accuracy, outperforming 23 popular anomaly detection baselines. Our paper is the first work revealing the powerful modeling capabilities of xLSTM for anomaly detection, paving the way for exciting new developments on this subject. Our code is available at: https://github.com/Nyderx/xlstmad

Method: Progressively Refined Differentiable Physics (PRDP) starts with coarse physics, adaptively refines it during training, and stops at the level sufficient for training accuracy, applicable to both unrolled and implicit differentiation of iterative linear solvers.

Result: PRDP achieves significant compute savings (62% training time reduction for Navier-Stokes emulation) while maintaining full network accuracy across various learning scenarios including inverse problems, neural emulators, and neural-hybrid solvers.

Conclusion: Full training accuracy is achievable with physics significantly coarser than fully converged solvers, and PRDP provides an effective adaptive refinement strategy for efficient differentiable physics training.

Abstract: The physics solvers employed for neural network training are primarily iterative, and hence, differentiating through them introduces a severe computational burden as iterations grow large. Inspired by works in bilevel optimization, we show that full accuracy of the network is achievable through physics significantly coarser than fully converged solvers. We propose Progressively Refined Differentiable Physics (PRDP), an approach that identifies the level of physics refinement sufficient for full training accuracy. By beginning with coarse physics, adaptively refining it during training, and stopping refinement at the level adequate for training, it enables significant compute savings without sacrificing network accuracy. Our focus is on differentiating iterative linear solvers for sparsely discretized differential operators, which are fundamental to scientific computing. PRDP is applicable to both unrolled and implicit differentiation. We validate its performance on a variety of learning scenarios involving differentiable physics solvers such as inverse problems, autoregressive neural emulators, and correction-based neural-hybrid solvers. In the challenging example of emulating the Navier-Stokes equations, we reduce training time by 62%.

Method: Proposes overlap-aware meta-learning attention: 1) Combines structural and feature similarities with weighted factors, 2) Partitions nodes by overlap levels into tasks, 3) Uses multi-task Meta-Weight-Net to determine weights, 4) Jointly trains internal MWN with HGNN losses and external model with weighted factors.

Result: Experiments on six real-world datasets show OMA-HGNN outperforms nine state-of-the-art methods in node classification, demonstrating superior node representation learning.

Conclusion: OMA-HGNN effectively addresses limitations of existing HGNNs by integrating multiple attention mechanisms and handling node overlap diversity through meta-learning, achieving better performance.

Abstract: Although hypergraph neural networks (HGNNs) have emerged as a powerful framework for analyzing complex datasets, their practical performance often remains limited. On one hand, existing networks typically employ a single type of attention mechanism, focusing on either structural or feature similarities during message passing. On the other hand, assuming that all nodes in current hypergraph models have the same level of overlap may lead to suboptimal generalization. To overcome these limitations, we propose a novel framework, overlap-aware meta-learning attention for hypergraph neural networks (OMA-HGNN). First, we introduce a hypergraph attention mechanism that integrates both structural and feature similarities. Specifically, we linearly combine their respective losses with weighted factors for the HGNN model. Second, we partition nodes into different tasks based on their diverse overlap levels and develop a multi-task Meta-Weight-Net (MWN) to determine the corresponding weighted factors. Third, we jointly train the internal MWN model with the losses from the external HGNN model and train the external model with the weighted factors from the internal model. To evaluate the effectiveness of OMA-HGNN, we conducted experiments on six real-world datasets and benchmarked its perfor-mance against nine state-of-the-art methods for node classification. The results demonstrate that OMA-HGNN excels in learning superior node representations and outperforms these baselines.

Method: Introduced a framework for calculating importance metrics in pruning groups that considers application-specific constraints. Used two approaches: grid search exploration and gradient descent optimization to find optimal pruning coefficients.

Result: The method effectively maintained application-relevant performance while achieving significant model size reduction in both MNIST autoencoder and TDMPC agent use cases.

Conclusion: The proposed framework successfully addresses the unique requirements of neural network controller compression by balancing model size reduction with preservation of critical performance characteristics.

Abstract: Deep neural networks (DNNs) offer significant flexibility and robust performance. This makes them ideal for building not only system models but also advanced neural network controllers (NNCs). However, their high complexity and computational needs often limit their use. Various model compression strategies have been developed over the past few decades to address these issues. These strategies are effective for general DNNs but do not directly apply to NNCs. NNCs need both size reduction and the retention of key application-specific performance features. In structured pruning, which removes groups of related elements, standard importance metrics often fail to protect these critical characteristics. In this paper, we introduce a novel framework for calculating importance metrics in pruning groups. This framework not only shrinks the model size but also considers various application-specific constraints. To find the best pruning coefficient for each group, we evaluate two approaches. The first approach involves simple exploration through grid search. The second utilizes gradient descent optimization, aiming to balance compression and task performance. We test our method in two use cases: one on an MNIST autoencoder and the other on a Temporal Difference Model Predictive Control (TDMPC) agent. Results show that the method effectively maintains application-relevant performance while achieving a significant reduction in model size.

Method: Uses a Cartesian tensor network trained to predict orbital positions and coefficients with a symmetry-breaking mechanism that preserves rotation equivariance of charge density, inspired by Gaussian Splatting techniques.

Result: Achieves state-of-the-art balance between computational efficiency and predictive accuracy, and reduces self-consistent field iterations by 50.72% on average when initializing DFT calculations.

Conclusion: ELECTRA successfully enables data-driven prediction of floating orbital positions, overcoming traditional barriers and significantly accelerating quantum chemistry computations while maintaining accuracy.

Abstract: We present the Electronic Tensor Reconstruction Algorithm (ELECTRA) - an equivariant model for predicting electronic charge densities using floating orbitals. Floating orbitals are a long-standing concept in the quantum chemistry community that promises more compact and accurate representations by placing orbitals freely in space, as opposed to centering all orbitals at the position of atoms. Finding the ideal placement of these orbitals requires extensive domain knowledge, though, which thus far has prevented widespread adoption. We solve this in a data-driven manner by training a Cartesian tensor network to predict the orbital positions along with orbital coefficients. This is made possible through a symmetry-breaking mechanism that is used to learn position displacements with lower symmetry than the input molecule while preserving the rotation equivariance of the charge density itself. Inspired by recent successes of Gaussian Splatting in representing densities in space, we are using Gaussian orbitals and predicting their weights and covariance matrices. Our method achieves a state-of-the-art balance between computational efficiency and predictive accuracy on established benchmarks. Furthermore, ELECTRA is able to lower the compute time required to arrive at converged DFT solutions - initializing calculations using our predicted densities yields an average 50.72 % reduction in self-consistent field (SCF) iterations on unseen molecules.

Method: APCFR+ uses adaptive asymmetric step sizes between implicit and explicit regret updates; SAPCFR+ simplifies this with fixed asymmetry requiring minimal code changes.

Result: Both APCFR+ and SAPCFR+ outperform PCFR+ in most games, achieve comparable convergence rates, and can generalize to other CFR algorithms like DCFR.

Conclusion: Asymmetric step sizes effectively enhance PCFR+ robustness, with SAPCFR+ providing a practical simplified alternative that maintains performance.

Abstract: Counterfactual Regret Minimization (CFR) algorithms are widely used to compute a Nash equilibrium (NE) in two-player zero-sum imperfect-information extensive-form games (IIGs). Among them, Predictive CFR$^+$ (PCFR$^+$) is particularly powerful, achieving an exceptionally fast empirical convergence rate via the prediction in many games.However, the empirical convergence rate of PCFR$^+$ would significantly degrade if the prediction is inaccurate, leading to unstable performance on certain IIGs. To enhance the robustness of PCFR$^+$, we propose Asymmetric PCFR$^+$ (APCFR$^+$), which employs an adaptive asymmetry of step sizes between the updates of implicit and explicit accumulated counterfactual regrets to mitigate the impact of the prediction inaccuracy on convergence. We present a theoretical analysis demonstrating why APCFR$^+$ can enhance the robustness. To the best of our knowledge, we are the first to propose the asymmetry of step sizes, a simple yet novel technique that effectively improves the robustness of PCFR$^+$. Then, to reduce the difficulty of implementing APCFR$^+$ caused by the adaptive asymmetry, we propose a simplified version of APCFR$^+$ called Simple APCFR$^+$ (SAPCFR$^+$), which uses a fixed asymmetry of step sizes to enable only a single-line modification compared to original PCFR$^+$.Experimental results on five standard IIG benchmarks and two heads-up no-limit Texas Hold’ em (HUNL) Subagems show that (i) both APCFR$^+$ and SAPCFR$^+$ outperform PCFR$^+$ in most of the tested games, (ii) SAPCFR$^+$ achieves a comparable empirical convergence rate with APCFR$^+$,and (iii) our approach can be generalized to improve other CFR algorithms, e.g., Discount CFR (DCFR).

Method: XNQ synergizes unsupervised node embeddings from randomized recursive Graph Neural Networks with an Expectation-Maximization algorithm that provides quantification-aware adjustment to calibrated node classifier probabilities.

Result: XNQ consistently and significantly improves on existing network quantification methods, setting new state-of-the-art performance. It provides 10x-100x training speed-up over other graph learning methods.

Conclusion: XNQ successfully addresses the challenges of network quantification by combining efficient unsupervised embeddings with robust quantification-aware probability adjustment, achieving superior performance and scalability.

Abstract: Network quantification (NQ) is the problem of estimating the proportions of nodes belonging to each class in subsets of unlabelled graph nodes. When prior probability shift is at play, this task cannot be effectively addressed by first classifying the nodes and then counting the class predictions. In addition, unlike non-relational quantification, NQ demands enhanced flexibility in order to capture a broad range of connectivity patterns, resilience to the challenge of heterophily, and scalability to large networks. In order to meet these stringent requirements, we introduce XNQ, a novel method that synergizes the flexibility and efficiency of the unsupervised node embeddings computed by randomized recursive Graph Neural Networks, with an Expectation-Maximization algorithm that provides a robust quantification-aware adjustment to the output probabilities of a calibrated node classifier. In an extensive evaluation, in which we also validate the design choices underpinning XNQ through comprehensive ablation experiments, we find that XNQ consistently and significantly improves on the best network quantification methods to date, thereby setting the new state of the art for this challenging task. XNQ also provides a training speed-up of up to 10x-100x over other methods based on graph learning.

Method: Proposed Epistemic Physics-Informed Neural Networks (E-PINNs) that use a small network (epinet) as an add-on to pre-trained PINNs with minimal computational overhead.

Result: E-PINNs achieve calibrated coverage with competitive sharpness at substantially lower cost than B-PINNs using HMC, and show better calibration than Dropout-PINNs in various test cases.

Conclusion: E-PINNs provide a favorable accuracy-efficiency trade-off for uncertainty quantification in PINNs, offering efficient epistemic uncertainty estimation with good calibration properties.

Abstract: Physics-informed neural networks (PINNs) have demonstrated promise as a framework for solving forward and inverse problems involving partial differential equations. Despite recent progress in the field, it remains challenging to quantify uncertainty in these networks. While techniques such as Bayesian PINNs (B-PINNs) provide a principled approach to capturing epistemic uncertainty through Bayesian inference, they can be computationally expensive for large-scale applications. In this work, we propose Epistemic Physics-Informed Neural Networks (E-PINNs), a framework that uses a small network, the epinet, to efficiently quantify epistemic uncertainty in PINNs. The proposed approach works as an add-on to existing, pre-trained PINNs with a small computational overhead. We demonstrate the applicability of the proposed framework in various test cases and compare the results with B-PINNs using Hamiltonian Monte Carlo (HMC) posterior estimation and dropout-equipped PINNs (Dropout-PINNs). In our experiments, E-PINNs achieve calibrated coverage with competitive sharpness at substantially lower cost. We demonstrate that when B-PINNs produce narrower bands, they under-cover in our tests. E-PINNs also show better calibration than Dropout-PINNs in these examples, indicating a favorable accuracy-efficiency trade-off.

Method: Built from production microservices data where nodes represent services with performance metrics and edges encode dependencies, with expert-annotated incident windows.

Result: Provides baseline results for forecasting models, time-series foundation models, and anomaly detectors on this realistic microservices dataset.

Conclusion: ChronoGraph offers a comprehensive benchmark for structure-aware forecasting and incident-aware evaluation in microservice environments.

Abstract: We present ChronoGraph, a graph-structured multivariate time series forecasting dataset built from real-world production microservices. Each node is a service that emits a multivariate stream of system-level performance metrics, capturing CPU, memory, and network usage patterns, while directed edges encode dependencies between services. The primary task is forecasting future values of these signals at the service level. In addition, ChronoGraph provides expert-annotated incident windows as anomaly labels, enabling evaluation of anomaly detection methods and assessment of forecast robustness during operational disruptions. Compared to existing benchmarks from industrial control systems or traffic and air-quality domains, ChronoGraph uniquely combines (i) multivariate time series, (ii) an explicit, machine-readable dependency graph, and (iii) anomaly labels aligned with real incidents. We report baseline results spanning forecasting models, pretrained time-series foundation models, and standard anomaly detectors. ChronoGraph offers a realistic benchmark for studying structure-aware forecasting and incident-aware evaluation in microservice systems.

Method: Two-stage model-based classification: 1) Identify artificially occurring zeroes using statistical modeling and information criteria, 2) Classify demand types (regular/intermittent, intermittent smooth/lumpy, fractional/count). Different demand types require different features.

Result: Empirical results show the framework increases forecasting accuracy and reduces inventory costs compared to methods applied directly without the generated features and two-stage classification.

Conclusion: The proposed two-stage classification framework effectively handles zero values and different demand characteristics, leading to improved forecasting performance and better inventory management decisions.

Abstract: Effective demand forecasting is critical for inventory management, production planning, and decision making across industries. Selecting the appropriate model and suitable features to efficiently capture patterns in the data is one of the main challenges in demand forecasting. In reality, this becomes even more complicated when the recorded sales have zeroes, which can happen naturally or due to some anomalies, such as stockouts and recording errors. Mistreating the zeroes can lead to the application of inappropriate forecasting methods, and thus leading to poor decision making. Furthermore, the demand itself can have different fundamental characteristics, and being able to distinguish one type from another might bring substantial benefits in terms of accuracy and thus decision making. We propose a two-stage model-based classification framework that in the first step, identifies artificially occurring zeroes, and in the second, classifies demand to one of the possible types: regular/intermittent, intermittent smooth/lumpy, fractional/count. The framework relies on statistical modelling and information criteria. We argue that different types of demand need different features, and show empirically that they tend to increase the accuracy of the forecasting methods and reduce inventory costs compared to those applied directly to the dataset without the generated features and the two-stage framework.

Method: Proposed FlashKAT with restructured kernel design that minimizes accesses to slow memory and reduces usage of atomic adds during gradient accumulation.

Result: FlashKAT achieves 86.5x training speedup compared to state-of-the-art KAT while reducing rounding errors in gradient computation.

Conclusion: Memory bottlenecks, not just FLOPs, are critical for KAN-based models’ performance, and FlashKAT’s kernel restructuring effectively addresses these issues.

Abstract: The Kolmogorov-Arnold Network (KAN) has been gaining popularity as an alternative to the multi-layer perceptron (MLP) with its increased expressiveness and interpretability. Even so, the KAN suffers from being orders of magnitude slower due to its increased computational cost and training instability, limiting its applicability to larger-scale tasks. Recently, the Kolmogorov-Arnold Transformer (KAT) has been proposed, which can achieve FLOPs similar to the traditional Transformer with MLPs by leveraging Group-Rational KAN (GR-KAN). Unfortunately, despite the comparable FLOPs, our testing reveals that the KAT is still 123x slower in training speeds, indicating that there are other performance bottlenecks beyond FLOPs. In this paper, we conduct a series of experiments to understand the root cause of the slowdown in KAT. We uncover that the slowdown can be isolated to memory stalls, linked more specifically to inefficient gradient accumulations in the backward pass of GR-KAN. To address this memory bottleneck, we propose FlashKAT, which minimizes accesses to slow memory and the usage of atomic adds through a restructured kernel. Evaluations demonstrate that FlashKAT can achieve a training speedup of 86.5x compared with the state-of-the-art KAT, while reducing rounding errors in the computation of the gradients.

Method: Approximate Koopman Eigenfunctions (KEFs) with real positive eigenvalues that vanish at separatrices, using Deep Neural Networks and optimization methods to efficiently locate these boundaries in complex systems.

Result: Successfully demonstrated on synthetic benchmarks, ecological network models, and high-dimensional recurrent neural networks trained on neuroscience tasks or fit to real neural data. The method enables optimal perturbation design to shift systems across separatrices.

Conclusion: The framework provides an effective approach for characterizing separatrices in high-dimensional dynamical systems, with practical applications in neuroscience for predicting outcomes of interventions like optogenetic stimulation.

Abstract: Many natural systems, including neural circuits involved in decision making, are modeled as high-dimensional dynamical systems with multiple stable states. While existing analytical tools primarily describe behavior near stable equilibria, characterizing separatrices–the manifolds that delineate boundaries between different basins of attraction–remains challenging, particularly in high-dimensional settings. Here, we introduce a numerical framework leveraging Koopman Theory combined with Deep Neural Networks to effectively characterize separatrices. Specifically, we approximate Koopman Eigenfunctions (KEFs) associated with real positive eigenvalues, which vanish precisely at the separatrices. Utilizing these scalar KEFs, optimization methods efficiently locate separatrices even in complex systems. We demonstrate our approach on synthetic benchmarks, ecological network models, and high-dimensional recurrent neural networks trained on either neuroscience-inspired tasks or fit to real neural data. Moreover, we illustrate the practical utility of our method by designing optimal perturbations that can shift systems across separatrices, enabling predictions relevant to optogenetic stimulation experiments in neuroscience.

Method: Proposed H2T2 - an online two-threshold hierarchical inference policy that continuously adapts thresholds on local model’s confidence scores to determine local prediction vs. offloading to remote model.

Result: H2T2 achieves sublinear regret, outperforms naive and single-threshold policies, sometimes surpassing offline optima, and shows robustness to distribution shifts and mismatched classifiers.

Conclusion: H2T2 provides an effective model-agnostic online learning solution for hierarchical inference systems that requires no training and learns during inference with limited feedback.

Abstract: We focus on a binary classification problem in an edge intelligence system where false negatives are more costly than false positives. The system has a compact, locally deployed model, which is supplemented by a larger, remote model, which is accessible via the network by incurring an offloading cost. For each sample, our system first uses the locally deployed model for inference. Based on the output of the local model, the sample may be offloaded to the remote model. This work aims to understand the fundamental trade-off between classification accuracy and the offloading costs within such a hierarchical inference (HI) system. To optimise this system, we propose an online learning framework that continuously adapts a pair of thresholds on the local model’s confidence scores. These thresholds determine the prediction of the local model and whether a sample is classified locally or offloaded to the remote model. We present a closed-form solution for the setting where the local model is calibrated. For the more general case of uncalibrated models, we introduce H2T2, an online two-threshold hierarchical inference policy, and prove it achieves sublinear regret. H2T2 is model-agnostic, requires no training, and learns during the inference phase using limited feedback. Simulations on real-world datasets show that H2T2 consistently outperforms naive and single-threshold HI policies, sometimes even surpassing offline optima. The policy also demonstrates robustness to distribution shifts and adapts effectively to mismatched classifiers.

Method: Introduces ‘Markov entanglement’ measure for multi-agent MDPs, drawing inspiration from quantum entanglement measurement. Shows value decomposition is possible when transition matrix isn’t entangled.

Result: Proves that widely-used index policies are weakly entangled with sublinear O(√N) decomposition error scaling. Provides method to efficiently estimate Markov entanglement in practice.

Conclusion: Markov entanglement provides theoretical foundation for value decomposition, explains its effectiveness, and offers practical tools to assess decomposition quality in multi-agent systems.

Abstract: Value decomposition has long been a fundamental technique in multi-agent dynamic programming and reinforcement learning (RL). Specifically, the value function of a global state $(s_1,s_2,\ldots,s_N)$ is often approximated as the sum of local functions: $V(s_1,s_2,\ldots,s_N)\approx\sum_{i=1}^N V_i(s_i)$. This approach traces back to the index policy in restless multi-armed bandit problems and has found various applications in modern RL systems. However, the theoretical justification for why this decomposition works so effectively remains underexplored. In this paper, we uncover the underlying mathematical structure that enables value decomposition. We demonstrate that a multi-agent Markov decision process (MDP) permits value decomposition if and only if its transition matrix is not “entangled” – a concept analogous to quantum entanglement in quantum physics. Drawing inspiration from how physicists measure quantum entanglement, we introduce how to measure the “Markov entanglement” for multi-agent MDPs and show that this measure can be used to bound the decomposition error in general multi-agent MDPs. Using the concept of Markov entanglement, we proved that a widely-used class of index policies is weakly entangled and enjoys a sublinear $\mathcal O(\sqrt{N})$ scale of decomposition error for $N$-agent systems. Finally, we show how Markov entanglement can be efficiently estimated in practice, providing practitioners with an empirical proxy for the quality of value decomposition.

Method: CARGO uses two pretrained causal Transformers as domain-specific foundation models to infer one-shot causal graphs per sequence in parallel, then aggregates them using adaptive frequency fusion to reconstruct global Markov boundaries of labels in a two-stage approach.

Result: CARGO demonstrated effective structured reasoning on a challenging real-world automotive fault prediction dataset with over 29,100 unique event types and 474 imbalanced labels.

Conclusion: The two-stage approach enables efficient probabilistic reasoning at scale while bypassing the intractable cost of full-dataset conditional independence testing for causal discovery in high-dimensional event sequences.

Abstract: Understanding causality in event sequences where outcome labels such as diseases or system failures arise from preceding events like symptoms or error codes is critical. Yet remains an unsolved challenge across domains like healthcare or vehicle diagnostics. We introduce CARGO, a scalable multi-label causal discovery method for sparse, high-dimensional event sequences comprising of thousands of unique event types. Using two pretrained causal Transformers as domain-specific foundation models for event sequences. CARGO infers in parallel, per sequence one-shot causal graphs and aggregates them using an adaptive frequency fusion to reconstruct the global Markov boundaries of labels. This two-stage approach enables efficient probabilistic reasoning at scale while bypassing the intractable cost of full-dataset conditional independence testing. Our results on a challenging real-world automotive fault prediction dataset with over 29,100 unique event types and 474 imbalanced labels demonstrate CARGO’s ability to perform structured reasoning.

Method: Reinterpret activation sparsity as dynamic structured weight sparsity, extend Optimal Brain Compression with activation-aware calibration, use output residuals from dense model as correction terms, and optimize for efficient GPU execution.

Result: Outperforms state-of-the-art structured pruning methods by up to 9.17% accuracy at 1.39× speedup compared to baseline dense model on LLaMA-2 and LLaMA-3.

Conclusion: DuoGPT provides an effective approach for deploying LLMs by leveraging both weight and activation sparsity while preserving model accuracy through careful calibration and correction techniques.

Abstract: Large language models (LLMs) deliver strong performance but are difficult to deploy due to high memory and compute costs. While pruning reduces these demands, most methods ignore activation sparsity observed at runtime. We reinterpret activation sparsity as dynamic structured weight sparsity and propose DuoGPT, a unified framework that constructs dual-sparse (spMspV) workloads by combining unstructured weight pruning with activation sparsity. To preserve accuracy, we extend the Optimal Brain Compression (OBC) framework with activation-aware calibration and introduce output residuals from the dense model as correction terms. We further optimize the solution for efficient GPU execution, enabling scalability to billion-parameter LLMs. Evaluations on LLaMA-2 and LLaMA-3 show that DuoGPT outperforms state-of-the-art structured pruning methods by up to 9.17% accuracy at an iso-speedup of 1.39$\times$ compared to the baseline dense model. Code is available at Github.

Method: Uses two pretrained Transformers as density estimators to infer per-sequence Markov Boundaries, enabling efficient parallel causal discovery without global conditional independence testing.

Result: On a real-world automotive dataset with 29,100 events and 474 labels, OSCAR recovers interpretable causal structures in minutes, while classical methods fail to scale.

Conclusion: OSCAR enables practical scientific diagnostics at production scale by providing efficient causal discovery for large-scale event sequences.

Abstract: Understanding causality in event sequences with thousands of sparse event types is critical in domains such as healthcare, cybersecurity, or vehicle diagnostics, yet current methods fail to scale. We present OSCAR, a one-shot causal autoregressive method that infers per-sequence Markov Boundaries using two pretrained Transformers as density estimators. This enables efficient, parallel causal discovery without costly global CI testing. On a real-world automotive dataset with 29,100 events and 474 labels, OSCAR recovers interpretable causal structures in minutes, while classical methods fail to scale, enabling practical scientific diagnostics at production scale.

Method: Introduces a unified framework that captures four symmetry classes: permutations, semi-permutations, orthogonal transformations, and general invertible maps, broadening valid reparameterizations beyond previous limited approaches.

Result: Enables discovery of low- and zero-barrier linear interpolation paths between independently trained Vision Transformers and GPT-2 models for the first time. Extends to multi-model and width-heterogeneous settings, allowing alignment across architectures of different sizes.

Conclusion: Reveals deeper structure in loss landscapes and underscores the importance of symmetry-aware analysis for understanding model space geometry, with implications for generalization and optimization in deep learning.

Abstract: Understanding the geometry of neural network loss landscapes is a central question in deep learning, with implications for generalization and optimization. A striking phenomenon is linear mode connectivity (LMC), where independently trained models can be connected by low- or zero-loss paths despite appearing to lie in separate loss basins. However, this is often obscured by symmetries in parameter space – such as neuron permutations – which make functionally equivalent models appear dissimilar. Prior work has predominantly focused on neuron reordering through permutations, but such approaches are limited in scope and fail to capture the richer symmetries exhibited by modern architectures such as Transformers. In this work, we introduce a unified framework that captures four symmetry classes – permutations, semi-permutations, orthogonal transformations, and general invertible maps – broadening the set of valid reparameterizations and subsuming many previous approaches as special cases. Crucially, this generalization enables, for the first time, the discovery of low- and zero-barrier linear interpolation paths between independently trained Vision Transformers and GPT-2 models. Furthermore, our framework extends beyond pairwise alignment to multi-model and width-heterogeneous settings, enabling alignment across architectures of different sizes. These results reveal deeper structure in the loss landscape and underscore the importance of symmetry-aware analysis for understanding model space geometry.

Method: Depth-wise Operation Pruning (DOP) searches for strategies to prune redundant operations by minimizing divergence from original model output while satisfying computational constraints, using depth-wise pruning and additive approximation for efficient optimization.

Result: DOP achieves 86% TFLOPS reduction and 83% latency reduction on LLaVA-Next-7B with only 1% performance loss, establishing new SOTA across 6 MLLMs and 13 benchmarks against 12 baselines.

Conclusion: DOP provides fine-grained computation control through operation pruning, demonstrating strong generalization capabilities and significant efficiency improvements for MLLMs.

Abstract: Token reduction accelerates Multimodal Large Language Models (MLLMs) by reducing excessive tokens, but overlooks structural redundancy differences, where critical and redundant modules process identical token loads. For fine-grained computation control, we define an ``operation" as the computation for a module to process a group of tokens and introduce the operation pruning framework to enable modules to selectively process tokens. Built on this framework, we propose Depth-wise Operation Pruning (DOP), a data-driven method that searches for strategies to prune redundant operations and save computational budget for critical modules to process more tokens than uniform allocation by minimizing divergence from the original model’s output probability distribution on a small validation set while satisfying computational constraints. For efficient optimization, DOP applies depth-wise pruning to reduce policy space and uses an additive approximation to minimize required validation runs. Depth-wise pruning partitions operations by module type and token group, and prunes operations in deeper layers before those in shallower layers within each module-group pair. The additive approximation obtains individual divergences by independently varying each policy parameter, and then sums them to approximate the joint divergence of simultaneously changing all policy parameters, reducing required validation runs from exponential to linear with respect to the number of policy parameters. Comprehensive evaluations show that DOP establishes new state-of-the-art performance across 6 MLLMs and 13 benchmarks against 12 baselines. On LLaVA-Next-7B, DOP achieves 86% TFLOPS reduction and 83% latency reduction on real GPU with only 1% performance loss. Our extensive ablation studies further demonstrate DOP’s data and time efficiency as well as strong generalization capabilities.

Method: Novel dimensional classification approach decoupling sowing and irrigation practices, using Sentinel-1 satellite imagery and ground-truth data from 1,400 fields across Punjab obtained through farmer training partnerships.

Result: Achieved F1 scores of 0.8 for DSR and 0.74 for AWD classification. Model applied to 3 million fields revealed spatial heterogeneity in adoption. District-level adoption rates correlated well with government estimates (Spearman’s ρ=0.69, Rank Biased Overlap=0.77).

Conclusion: Provides policymakers with a scalable tool to track sustainable rice farming adoption, enabling targeted interventions for water conservation and climate mitigation at regional scale.

Abstract: Rice cultivation supplies half the world’s population with staple food, while also being a major driver of freshwater depletion–consuming roughly a quarter of global freshwater–and accounting for approx. 48% of greenhouse gas emissions from croplands. In regions like Punjab, India, where groundwater levels are plummeting at 41.6 cm/year, adopting water-saving rice farming practices is critical. Direct-Seeded Rice (DSR) and Alternate Wetting and Drying (AWD) can cut irrigation water use by 20-40% without hurting yields, yet lack of spatial data on adoption impedes effective adaptation policy and climate action. We present a machine learning framework to bridge this data gap by monitoring sustainable rice farming at scale. In collaboration with agronomy experts and a large-scale farmer training program, we obtained ground-truth data from 1,400 fields across Punjab. Leveraging this partnership, we developed a novel dimensional classification approach that decouples sowing and irrigation practices, achieving F1 scores of 0.8 and 0.74 respectively, solely employing Sentinel-1 satellite imagery. Explainability analysis reveals that DSR classification is robust while AWD classification depends primarily on planting schedule differences, as Sentinel-1’s 12-day revisit frequency cannot capture the higher frequency irrigation cycles characteristic of AWD practices. Applying this model across 3 million fields reveals spatial heterogeneity in adoption at the state level, highlighting gaps and opportunities for policy targeting. Our district-level adoption rates correlate well with government estimates (Spearman’s $ρ$=0.69 and Rank Biased Overlap=0.77). This study provides policymakers and sustainability programs a powerful tool to track practice adoption, inform targeted interventions, and drive data-driven policies for water conservation and climate mitigation at regional scale.

Method: Proposes HyperEvent which captures relative structural patterns in event sequences through intuitive encoding, combines interpretable features with lightweight transformer classifier, and reframes link prediction as event structure recognition.

Result: Achieves competitive results across multiple benchmarks, often matching the performance of more complex models despite its simplicity.

Conclusion: Effective modeling can be achieved through simple structural encoding, providing a clear baseline for evaluating future advancements in dynamic graph representation learning.

Abstract: Learning representations for continuous-time dynamic graphs is critical for dynamic link prediction. While recent methods have become increasingly complex, the field lacks a strong and informative baseline to reliably gauge progress. This paper proposes HyperEvent, a simple approach that captures relative structural patterns in event sequences through an intuitive encoding mechanism. As a straightforward baseline, HyperEvent leverages relative structural encoding to identify meaningful event sequences without complex parameterization. By combining these interpretable features with a lightweight transformer classifier, HyperEvent reframes link prediction as event structure recognition. Despite its simplicity, HyperEvent achieves competitive results across multiple benchmarks, often matching the performance of more complex models. This work demonstrates that effective modeling can be achieved through simple structural encoding, providing a clear reference point for evaluating future advancements.

Method: Uses TabularARGN autoregressive framework with differential privacy guarantees, fairness-aware generation, and supports diverse data types including multi-table and sequential datasets.

Result: Delivers competitive performance with improvements in speed and usability, deployed as both cloud service and local software with rapid adoption.

Conclusion: The SDK effectively addresses real-world data bottlenecks and promotes data democratization through practical synthetic data generation.

Abstract: Machine learning development critically depends on access to high-quality data. However, increasing restrictions due to privacy, proprietary interests, and ethical concerns have created significant barriers to data accessibility. Synthetic data offers a viable solution by enabling safe, broad data usage without compromising sensitive information. This paper presents the MOSTLY AI Synthetic Data Software Development Kit (SDK), an open-source toolkit designed specifically for synthesizing high-quality tabular data. The SDK integrates robust features such as differential privacy guarantees, fairness-aware data generation, and automated quality assurance into a flexible and accessible Python interface. Leveraging the TabularARGN autoregressive framework, the SDK supports diverse data types and complex multi-table and sequential datasets, delivering competitive performance with notable improvements in speed and usability. Currently deployed both as a cloud service and locally installable software, the SDK has seen rapid adoption, highlighting its practicality in addressing real-world data bottlenecks and promoting widespread data democratization.

Method: Develops a slicing technique based on appropriate lower bounds, hierarchical optimal transport, and quadrature rules for 1D OT problems. Avoids the underlying quadratic program of original GW/FGW.

Result: The novel sliced FGW significantly reduces computational effort while remaining invariant to isometric transformations and allowing comparison of arbitrary geometries. It defines a pseudo-metric that bounds FGW from below and shows interpolation properties between sliced Wasserstein and GW.

Conclusion: The proposed sliced distance is numerically more robust and reliable than original GW/FGW, particularly effective for shape retrieval and graph isomorphism testing applications.

Abstract: The Gromov–Wasserstein (GW) distance and its fused extension (FGW) are powerful tools for comparing heterogeneous data. Their computation is, however, challenging since both distances are based on non-convex, quadratic optimal transport (OT) problems. Leveraging 1D OT, a sliced version of GW has been proposed to lower the computational burden. Unfortunately, this sliced version is restricted to Euclidean geometry and loses invariance to isometries, strongly limiting its application in practice. To overcome these issues, we propose a novel slicing technique for GW as well as for FGW that is based on an appropriate lower bound, hierarchical OT, and suitable quadrature rules for the underlying 1D OT problems. Our novel sliced FGW significantly reduces the numerical effort while remaining invariant to isometric transformations and allowing the comparison of arbitrary geometries. We show that our new distance actually defines a pseudo-metric for structured spaces that bounds FGW from below and study its interpolation properties between sliced Wasserstein and GW. Since we avoid the underlying quadratic program, our sliced distance is numerically more robust and reliable than the original GW and FGW distance; especially in the context of shape retrieval and graph isomorphism testing.

Method: Introduce a simple framework to analyze discretization, initialization, and score estimation errors, emphasizing spatial regularity of learned score functions and controlling score error relative to true reverse processes.

Result: First Wasserstein convergence bound for Heun sampler and improved results for Euler sampler of probability flow ODE; sharpened initialization error bounds using smoothed Wasserstein distances.

Conclusion: The analysis highlights the importance of spatial regularity in score functions and proper error control in denoising score matching for effective convergence guarantees in diffusion models.

Abstract: We provide new convergence guarantees in Wasserstein distance for diffusion-based generative models, covering both stochastic (DDPM-like) and deterministic (DDIM-like) sampling methods. We introduce a simple framework to analyze discretization, initialization, and score estimation errors. Notably, we derive the first Wasserstein convergence bound for the Heun sampler and improve existing results for the Euler sampler of the probability flow ODE. Our analysis emphasizes the importance of spatial regularity of the learned score function and argues for controlling the score error with respect to the true reverse process, in line with denoising score matching. We also incorporate recent results on smoothed Wasserstein distances to sharpen initialization error bounds.

Method: Introduces Succeed or Learn Slowly (SoLS) algorithm with modified off-policy actor-critic approach: direct policy updates for positive samples with high returns, and conservative regularized updates for negative samples. Augmented with Successful Transition Replay (STR) to prioritize learning from successful interactions.

Result: Evaluated on AndroidWorld benchmark, SoLS significantly outperforms existing methods (at least 17% relative increase) including prompt-engineering and RL approaches, while requiring substantially fewer computational resources than GPT-4o-based methods with 5-60x faster inference.

Conclusion: SoLS effectively addresses policy degradation issues in foundation model fine-tuning for RL tasks, demonstrating superior sample efficiency and performance in mobile app control tasks while being computationally efficient.

Abstract: Reinforcement learning (RL) using foundation models for policy approximations in multi-turn tasks remains challenging. We identify two main limitations related to sparse reward settings and policy gradient updates, based on which we formulate a key insight: updates from positive samples with high returns typically do not require policy regularisation, whereas updates from negative samples, reflecting undesirable behaviour, can harm model performance. This paper introduces Succeed or Learn Slowly (SoLS), a novel off-policy RL algorithm evaluated on mobile app control tasks. SoLS improves sample efficiency when fine-tuning foundation models for user interface navigation via a modified off-policy actor-critic approach, applying direct policy updates for positive samples and conservative, regularised updates for negative ones to prevent model degradation. We augment SoLS with Successful Transition Replay (STR), which prioritises learning from successful interactions, further improving sample efficiency. We evaluate SoLS on the AndroidWorld benchmark, where it significantly outperforms existing methods (at least 17% relative increase), including prompt-engineering and RL approaches, while requiring substantially fewer computational resources than GPT-4o-based methods with 5-60x faster inference.

Method: Developed Logistic-KAN (generalization of logistic regression) and Kolmogorov-Arnold Additive Model (KAAM) - function-based architectures using KANs that provide symbolic representations, built-in patient insights, visualizations, and nearest-patient retrieval.

Result: Models match or outperform standard baselines across multiple health datasets while remaining fully interpretable, offering transparent symbolic formulas for clinical predictions.

Conclusion: KANs represent a promising step toward trustworthy AI in medicine, providing models that clinicians can understand, audit, and act upon with built-in interpretability.

Abstract: Why should a clinician trust an Artificial Intelligence (AI) prediction? Despite the increasing accuracy of machine learning methods in medicine, the lack of transparency continues to hinder their adoption in clinical practice. In this work, we explore Kolmogorov-Arnold Networks (KANs) for clinical classification tasks on tabular data. In contrast to traditional neural networks, KANs are function-based architectures that offer intrinsic interpretability through transparent, symbolic representations. We introduce \emph{Logistic-KAN}, a flexible generalization of logistic regression, and \emph{Kolmogorov-Arnold Additive Model (KAAM)}, a simplified additive variant that delivers transparent, symbolic formulas. Unlike ``black-box’’ models that require post-hoc explainability tools, our models support built-in patient-level insights, intuitive visualizations, and nearest-patient retrieval. Across multiple health datasets, our models match or outperform standard baselines, while remaining fully interpretable. These results position KANs as a promising step toward trustworthy AI that clinicians can understand, audit, and act upon. We release the code for reproducibility in \codeurl.

Method: Analyzed the geometry of iterative updates in recurrent-depth transformers, measuring loop step dynamics across training. Proposed an early-exit mechanism using second-order differences in step-size rather than KL-divergence or first-order approaches.

Result: Found that loop steps become smaller and more orthogonal during training, indicating better local modeling of fine structure. The second-order difference exit strategy demonstrated superior performance, stability, and time-efficiency compared to KL-divergence and first-order methods.

Conclusion: The geometric analysis reveals meaningful two-scale dynamics in recurrent-depth transformers, and the proposed second-order early-exit mechanism provides a more effective way to manage computational resources during inference.

Abstract: Recurrent-depth transformers scale test-time compute by iterating latent computations before emitting tokens. We study the geometry of these iterates and argue for a simple, two-scale operational picture: (i) within a looped block, updates act as small-scale refinements; (ii) across consecutive blocks, states undergo a larger-scale drift. Across training, our measurements show that loop steps become smaller and increasingly orthogonal to one another, indicating better local modeling of fine structure rather than merely pushing in a single direction. These dynamics motivate an early-exit mechanism based on the model’s second-order difference in step-size, which we show is superior in terms of performance, stability and time-efficiency, when compared to the KL-divergence exit strategy of Geiping et al. and its naive first-order counterpart.

Method: Uses a histogram-based algorithm to automatically update domain discretization in a data-driven manner, informed by previous layer output ranges.

Result: Matches or exceeds performance of prior KANs and MLPs on scientific equation learning, image classification, control Lyapunov function learning, and OOD detection tasks.

Conclusion: AdaptKAN successfully automates domain grid updates while maintaining KAN benefits, with the histogram algorithm also useful for OOD detection.

Abstract: Kolmogorov-Arnold Networks (KANs) are a class of neural networks that have received increased attention in recent literature. In contrast to MLPs, KANs leverage parameterized, trainable activation functions and offer several benefits including improved interpretability and higher accuracy on learning symbolic equations. However, the original KAN architecture requires adjustments to the domain discretization of the network (called the “domain grid”) during training, creating extra overhead for the user in the training process. Typical KAN layers are not designed with the ability to autonomously update their domains in a data-driven manner informed by the changing output ranges of previous layers. As an added benefit, this histogram algorithm may also be applied towards detecting out-of-distribution (OOD) inputs in a variety of settings. We demonstrate that AdaptKAN exceeds or matches the performance of prior KAN architectures and MLPs on four different tasks: learning scientific equations from the Feynman dataset, image classification from frozen features, learning a control Lyapunov function, and detecting OOD inputs on the OpenOOD v1.5 benchmark.

Method: Developed theory for permutation metrics’ dependence on superposition, trained sparse autoencoders to disentangle superposition in toy models, and tested alignment metrics on DNN-DNN and DNN-brain comparisons in visual domain.

Result: Alignment scores typically increased when base neurons were replaced with sparse overcomplete latent codes, showing similar improvements for both DNN-DNN and DNN-brain linear regression alignment.

Conclusion: Superposition disentanglement is necessary for mapping metrics to accurately measure the true representational alignment between neural networks.

Abstract: The superposition hypothesis states that single neurons may participate in representing multiple features in order for the neural network to represent more features than it has neurons. In neuroscience and AI, representational alignment metrics measure the extent to which different deep neural networks (DNNs) or brains represent similar information. In this work, we explore a critical question: does superposition interact with alignment metrics in any undesirable way? We hypothesize that models which represent the same features in different superposition arrangements, i.e., their neurons have different linear combinations of the features, will interfere with predictive mapping metrics (semi-matching, soft-matching, linear regression), producing lower alignment than expected. We develop a theory for how permutation metrics are dependent on superposition arrangements. This is tested by training sparse autoencoders (SAEs) to disentangle superposition in toy models, where alignment scores are shown to typically increase when a model’s base neurons are replaced with its sparse overcomplete latent codes. We find similar increases for DNN-DNN and DNN-brain linear regression alignment in the visual domain. Our results suggest that superposition disentanglement is necessary for mapping metrics to uncover the true representational alignment between neural networks.

Method: Proposes PhyKAN - a physics-informed Kolmogorov-Arnold Network that integrates local KAN branch with global branch containing Green’s function priors, derived from EFIE to maintain physical consistency.

Result: Achieves sub-0.01 reconstruction errors across canonical geometries and accurate unsupervised radar cross section predictions.

Conclusion: PhyKAN provides an interpretable, physics-consistent bridge between classical electromagnetic solvers and modern neural network approaches.

Abstract: The Method of Moments (MoM) is constrained by the usage of static, geometry-defined basis functions, such as the Rao-Wilton-Glisson (RWG) basis. This letter reframes electromagnetic modeling around a learnable basis representation rather than solving for the coefficients over a fixed basis. We first show that the RWG basis is essentially a static and piecewise-linear realization of the Kolmogorov-Arnold representation theorem. Inspired by this insight, we propose PhyKAN, a physics-informed Kolmogorov-Arnold Network (KAN) that generalizes RWG into a learnable and adaptive basis family. Derived from the EFIE, PhyKAN integrates a local KAN branch with a global branch embedded with Green’s function priors to preserve physical consistency. It is demonstrated that, across canonical geometries, PhyKAN achieves sub-0.01 reconstruction errors as well as accurate, unsupervised radar cross section predictions, offering an interpretable, physics-consistent bridge between classical solvers and modern neural network models for electromagnetic modeling.

Method: SRS module with Selective Patching and Dynamic Reassembly to adaptively select and shuffle patches from contextual time series, plus MLP head for forecasting.

Result: Achieves state-of-the-art performance on real-world datasets from multiple domains; also enhances existing patch-based models as plug-and-play module.

Conclusion: SRS enables flexible representation spaces that better exploit contextual information, improving forecasting performance in patch-based time series models.

Abstract: Time Series Forecasting has made significant progress with the help of Patching technique, which partitions time series into multiple patches to effectively retain contextual semantic information into a representation space beneficial for modeling long-term dependencies. However, conventional patching partitions a time series into adjacent patches, which causes a fixed representation space, thus resulting in insufficiently expressful representations. In this paper, we pioneer the exploration of constructing a selective representation space to flexibly include the most informative patches for forecasting. Specifically, we propose the Selective Representation Space (SRS) module, which utilizes the learnable Selective Patching and Dynamic Reassembly techniques to adaptively select and shuffle the patches from the contextual time series, aiming at fully exploiting the information of contextual time series to enhance the forecasting performance of patch-based models. To demonstrate the effectiveness of SRS module, we propose a simple yet effective SRSNet consisting of SRS and an MLP head, which achieves state-of-the-art performance on real-world datasets from multiple domains. Furthermore, as a novel plugin-and-play module, SRS can also enhance the performance of existing patch-based models. The resources are available at https://github.com/decisionintelligence/SRSNet.

Method: Introduces BATIS framework using Bayesian deep learning with iterative updating of prior predictions using limited observational data, benchmarking various uncertainty quantification approaches on eBird citizen science data.

Result: Bayesian deep learning approaches significantly improve reliability of species distribution models in data-scarce locations, enabling better ecological understanding and conservation efforts.

Conclusion: Bayesian deep learning can greatly enhance species distribution model reliability by effectively handling both aleatoric and epistemic uncertainty, particularly benefiting conservation in data-limited regions.

Abstract: Species distribution models (SDMs), which aim to predict species occurrence based on environmental variables, are widely used to monitor and respond to biodiversity change. Recent deep learning advances for SDMs have been shown to perform well on complex and heterogeneous datasets, but their effectiveness remains limited by spatial biases in the data. In this paper, we revisit deep SDMs from a Bayesian perspective and introduce BATIS, a novel and practical framework wherein prior predictions are updated iteratively using limited observational data. Models must appropriately capture both aleatoric and epistemic uncertainty to effectively combine fine-grained local insights with broader ecological patterns. We benchmark an extensive set of uncertainty quantification approaches on a novel dataset including citizen science observations from the eBird platform. Our empirical study shows how Bayesian deep learning approaches can greatly improve the reliability of SDMs in data-scarce locations, which can contribute to ecological understanding and conservation efforts.

Method: Proposes ScaleDL framework that integrates nonlinear layer-wise modeling with GNN-based cross-layer interaction mechanisms, and employs D-optimal method to reduce data collection costs.

Result: Experiments on five popular DNN models show ScaleDL achieves 6 times lower MRE and 5 times lower RMSE compared to baseline models, demonstrating enhanced prediction accuracy and generalizability.

Conclusion: ScaleDL effectively balances accuracy, generalizability, and data collection costs for DNN runtime prediction, making it a practical solution for optimizing distributed computing resource allocation.

Abstract: Deep neural networks (DNNs) form the cornerstone of modern AI services, supporting a wide range of applications, including autonomous driving, chatbots, and recommendation systems. As models increase in size and complexity, DNN workloads such as training and inference tasks impose unprecedented demands on distributed computing resources, making accurate runtime prediction essential for optimizing development and resource allocation. Traditional methods rely on additive computational unit models, limiting their accuracy and generalizability. In contrast, graph-enhanced modeling improves performance but significantly increases data collection costs. Therefore, there is a critical need for a method that strikes a balance between accuracy, generalizability, and data collection costs. To address these challenges, we propose ScaleDL, a novel runtime prediction framework that combines nonlinear layer-wise modeling with graph neural network (GNN)-based cross-layer interaction mechanism, enabling accurate DNN runtime prediction and hierarchical generalizability across different network architectures. Additionally, we employ the D-optimal method to reduce data collection costs. Experiments on the workloads of five popular DNN models demonstrate that ScaleDL enhances runtime prediction accuracy and generalizability, achieving 6 times lower MRE and 5 times lower RMSE compared to baseline models.

Method: Uses bi-level optimization with multi-reward reinforcement learning, including Pattern Balanced Rewarding Policy with pattern inversion and distribution consistency rewards, and Constrained Group Relative Reinforcement Learning with dynamic barrier constraints.

Result: Extensive experiments demonstrate the effectiveness of CREAT in achieving targeted pollution with minimal detectability.

Conclusion: CREAT successfully addresses limitations of previous PPA methods by synergizing constrained reinforcement learning to balance attack effectiveness and stealthiness.

Abstract: Sequential Recommenders, which exploit dynamic user intents through interaction sequences, is vulnerable to adversarial attacks. While existing attacks primarily rely on data poisoning, they require large-scale user access or fake profiles thus lacking practicality. In this paper, we focus on the Profile Pollution Attack that subtly contaminates partial user interactions to induce targeted mispredictions. Previous PPA methods suffer from two limitations, i.e., i) over-reliance on sequence horizon impact restricts fine-grained perturbations on item transitions, and ii) holistic modifications cause detectable distribution shifts. To address these challenges, we propose a constrained reinforcement driven attack CREAT that synergizes a bi-level optimization framework with multi-reward reinforcement learning to balance adversarial efficacy and stealthiness. We first develop a Pattern Balanced Rewarding Policy, which integrates pattern inversion rewards to invert critical patterns and distribution consistency rewards to minimize detectable shifts via unbalanced co-optimal transport. Then we employ a Constrained Group Relative Reinforcement Learning paradigm, enabling step-wise perturbations through dynamic barrier constraints and group-shared experience replay, achieving targeted pollution with minimal detectability. Extensive experiments demonstrate the effectiveness of CREAT.

Method: Developed REACT-LLM benchmark with comprehensive evaluation of LLMs combined with causal features from causal discovery algorithms, comparing against traditional ML methods.

Result: LLMs perform reasonably but don’t outperform traditional ML models; causal feature integration offers limited gains due to strict assumptions of causal discovery methods in complex clinical data.

Conclusion: While direct integration shows limited improvement, the benchmark reveals promising synergy potential between LLMs and causal learning in clinical prognostics.

Abstract: Large Language Models (LLMs) and causal learning each hold strong potential for clinical decision making (CDM). However, their synergy remains poorly understood, largely due to the lack of systematic benchmarks evaluating their integration in clinical risk prediction. In real-world healthcare, identifying features with causal influence on outcomes is crucial for actionable and trustworthy predictions. While recent work highlights LLMs’ emerging causal reasoning abilities, there lacks comprehensive benchmarks to assess their causal learning and performance informed by causal features in clinical risk prediction. To address this, we introduce REACT-LLM, a benchmark designed to evaluate whether combining LLMs with causal features can enhance clinical prognostic performance and potentially outperform traditional machine learning (ML) methods. Unlike existing LLM-clinical benchmarks that often focus on a limited set of outcomes, REACT-LLM evaluates 7 clinical outcomes across 2 real-world datasets, comparing 15 prominent LLMs, 6 traditional ML models, and 3 causal discovery (CD) algorithms. Our findings indicate that while LLMs perform reasonably in clinical prognostics, they have not yet outperformed traditional ML models. Integrating causal features derived from CD algorithms into LLMs offers limited performance gains, primarily due to the strict assumptions of many CD methods, which are often violated in complex clinical data. While the direct integration yields limited improvement, our benchmark reveals a more promising synergy.

Method: Two-step approach: first obtain model backbone without direct queries to victim model, then strategically use fixed query limit to extract most informative data.

Result: Experiments on eight real-world datasets show the attack is effective even under very restricted query limits and with defenses against model extraction in place.

Conclusion: The findings highlight the need for robust defenses against GNN model extraction threats, as current methods are vulnerable even under severe query limitations.

Abstract: Current graph neural network (GNN) model-stealing methods rely heavily on queries to the victim model, assuming no hard query limits. However, in reality, the number of allowed queries can be severely limited. In this paper, we demonstrate how an adversary can extract a GNN with very limited interactions with the model. Our approach first enables the adversary to obtain the model backbone without making direct queries to the victim model and then to strategically utilize a fixed query limit to extract the most informative data. The experiments on eight real-world datasets demonstrate the effectiveness of the attack, even under a very restricted query limit and under defense against model extraction in place. Our findings underscore the need for robust defenses against GNN model extraction threats.

Method: Proposes decision-focused continual learning with Fisher information based regularization to preserve parameters critical to prior tasks, and develops differentiable convex surrogate for stable gradient backpropagation.

Result: Experiments at Jurong Port show superior decision performance and generalization over existing methods with reduced computational cost, enabling sustainable long-term learning across varying task streams.

Conclusion: The framework successfully enables learning decision-aligned forecasting models across evolving scheduling tasks while maintaining computational efficiency and improving generalization.

Abstract: Power-logistics scheduling in modern seaports typically follow a predict-then-optimize pipeline. To enhance the decision quality of forecasts, decision-focused learning has been proposed, which aligns the training of forecasting models with downstream decision outcomes. However, this end-to-end design inherently restricts the value of forecasting models to only a specific task structure, and thus generalize poorly to evolving tasks induced by varying seaport vessel arrivals. We address this gap with a decision-focused continual learning framework that adapts online to a stream of scheduling tasks. Specifically, we introduce Fisher information based regularization to enhance cross-task generalization by preserving parameters critical to prior tasks. A differentiable convex surrogate is also developed to stabilize gradient backpropagation. The proposed approach enables learning a decision-aligned forecasting model across a varying tasks stream with a sustainable long-term computational burden. Experiments calibrated to the Jurong Port demonstrate superior decision performance and generalization over existing methods with reduced computational cost.

Method: Dual-branch framework with: 1) Temporal Stabilizing Fusion (TFS) using non-stationary mix of experts filter to suppress temporal non-stationary patterns while preserving long-term dependencies; 2) Frequency Wave Modeling (FWM) applying frequency differencing to highlight components with spectral shifts.

Result: Extensive experiments on real-world benchmarks show DTAF outperforms state-of-the-art baselines with significant improvements in forecasting accuracy under non-stationary conditions.

Conclusion: DTAF effectively handles non-stationarity in both temporal and frequency domains, generating robust forecasts that adapt to distribution shifts and spectral variability in time series data.

Abstract: Time series forecasting is critical for decision-making across dynamic domains such as energy, finance, transportation, and cloud computing. However, real-world time series often exhibit non-stationarity, including temporal distribution shifts and spectral variability, which pose significant challenges for long-term time series forecasting. In this paper, we propose DTAF, a dual-branch framework that addresses non-stationarity in both the temporal and frequency domains. For the temporal domain, the Temporal Stabilizing Fusion (TFS) module employs a non-stationary mix of experts (MOE) filter to disentangle and suppress temporal non-stationary patterns while preserving long-term dependencies. For the frequency domain, the Frequency Wave Modeling (FWM) module applies frequency differencing to dynamically highlight components with significant spectral shifts. By fusing the complementary outputs of TFS and FWM, DTAF generates robust forecasts that adapt to both temporal and frequency domain non-stationarity. Extensive experiments on real-world benchmarks demonstrate that DTAF outperforms state-of-the-art baselines, yielding significant improvements in forecasting accuracy under non-stationary conditions. All codes are available at https://github.com/PandaJunk/DTAF.

Method: Use normalizing flows to estimate the true conditional distribution of latent space given labels, and leverage the variance of the Gaussian decoder as a learnable parameter to address blurry image issues.

Result: Achieved 5% reduction in FID score and 7.7% increase in log likelihood compared to previous CVAE methods, demonstrating improved image quality and diversity.

Conclusion: CVAEs can still be competitive when properly modeling conditional distributions using normalizing flows, offering a viable alternative to diffusion models for conditional image generation tasks.

Abstract: Variational Autoencoders and Generative Adversarial Networks remained the state-of-the-art (SOTA) generative models until 2022. Now they are superseded by diffusion based models. Efforts to improve traditional models have stagnated as a result. In old-school fashion, we explore image generation with conditional Variational Autoencoders (CVAE) to incorporate desired attributes within the images. VAEs are known to produce blurry images with less diversity, we refer a method that solve this issue by leveraging the variance of the gaussian decoder as a learnable parameter during training. Previous works on CVAEs assumed that the conditional distribution of the latent space given the labels is equal to the prior distribution, which is not the case in reality. We show that estimating it using normalizing flows results in better image generation than existing methods by reducing the FID by 5% and increasing log likelihood by 7.7% than the previous case.

Method: Performs hierarchical agglomerative clustering with integrated cluster evaluation. At each agglomeration step, creates structure-aware data representation via self-supervised learning guided by current clustering, then measures consensus between original and self-supervised representations using nearest neighbor consensus score.

Result: Extensive experiments show SCMax outperforms existing clustering approaches designed for scenarios with unknown number of clusters.

Conclusion: The framework successfully addresses the parameter-free clustering challenge by using self-supervised consensus maximization to automatically determine optimal clustering structure.

Abstract: Clustering is a fundamental task in unsupervised learning, but most existing methods heavily rely on hyperparameters such as the number of clusters or other sensitive settings, limiting their applicability in real-world scenarios. To address this long-standing challenge, we propose a novel and fully parameter-free clustering framework via Self-supervised Consensus Maximization, named SCMax. Our framework performs hierarchical agglomerative clustering and cluster evaluation in a single, integrated process. At each step of agglomeration, it creates a new, structure-aware data representation through a self-supervised learning task guided by the current clustering structure. We then introduce a nearest neighbor consensus score, which measures the agreement between the nearest neighbor-based merge decisions suggested by the original representation and the self-supervised one. The moment at which consensus maximization occurs can serve as a criterion for determining the optimal number of clusters. Extensive experiments on multiple datasets demonstrate that the proposed framework outperforms existing clustering approaches designed for scenarios with an unknown number of clusters.

Method: Train centralized embedding model for trajectory representations, then decentralized models to approximate it. Retrieve relevant trajectories as context, use memory mechanism balancing online/offline data, and apply hybrid utility score with individual/team returns.

Result: Extensive experiments on LBF and SMAC benchmarks show MAICC enables faster adaptation to unseen tasks compared to existing methods.

Conclusion: MAICC effectively enhances coordination through fast adaptation in cooperative MARL by leveraging decentralized memory retrieval and balanced credit assignment.

Abstract: Large transformer models, trained on diverse datasets, have demonstrated impressive few-shot performance on previously unseen tasks without requiring parameter updates. This capability has also been explored in Reinforcement Learning (RL), where agents interact with the environment to retrieve context and maximize cumulative rewards, showcasing strong adaptability in complex settings. However, in cooperative Multi-Agent Reinforcement Learning (MARL), where agents must coordinate toward a shared goal, decentralized policy deployment can lead to mismatches in task alignment and reward assignment, limiting the efficiency of policy adaptation. To address this challenge, we introduce Multi-agent In-context Coordination via Decentralized Memory Retrieval (MAICC), a novel approach designed to enhance coordination by fast adaptation. Our method involves training a centralized embedding model to capture fine-grained trajectory representations, followed by decentralized models that approximate the centralized one to obtain team-level task information. Based on the learned embeddings, relevant trajectories are retrieved as context, which, combined with the agents’ current sub-trajectories, inform decision-making. During decentralized execution, we introduce a novel memory mechanism that effectively balances test-time online data with offline memory. Based on the constructed memory, we propose a hybrid utility score that incorporates both individual- and team-level returns, ensuring credit assignment across agents. Extensive experiments on cooperative MARL benchmarks, including Level-Based Foraging (LBF) and SMAC (v1/v2), show that MAICC enables faster adaptation to unseen tasks compared to existing methods. Code is available at https://github.com/LAMDA-RL/MAICC.

Method: Uses an orchestrator, tool-calling agent, and general chat agent with tool integration defined through structured specifications and domain-informed instructions. Avoids training and data generation by adopting behavior modeling and documentation.

Result: Enables scalable adaptation to private tools and evolving contexts without continual retraining. Supports lightweight deployment, API specifications as retrieval resources, and synthetic dialogue generation for evaluation.

Conclusion: Provides a sustainable method for aligning agent behavior with domain expertise in private conversational ecosystems without the need for training or data generation.

Abstract: The rise of agentic systems that combine orchestration, tool use, and conversational capabilities, has been more visible by the recent advent of large language models (LLMs). While open-domain frameworks exist, applying them in private domains remains difficult due to heterogeneous tool formats, domain-specific jargon, restricted accessibility of APIs, and complex governance. Conventional solutions, such as fine-tuning on synthetic dialogue data, are burdensome and brittle under domain shifts, and risk degrading general performance. In this light, we introduce a framework for private-domain multi-agent conversational systems that avoids training and data generation by adopting behavior modeling and documentation. Our design simply assumes an orchestrator, a tool-calling agent, and a general chat agent, with tool integration defined through structured specifications and domain-informed instructions. This approach enables scalable adaptation to private tools and evolving contexts without continual retraining. The framework supports practical use cases, including lightweight deployment of multi-agent systems, leveraging API specifications as retrieval resources, and generating synthetic dialogue for evaluation – providing a sustainable method for aligning agent behavior with domain expertise in private conversational ecosystems.

Method: Designed CP-WBFT, a confidence probe-based weighted Byzantine Fault Tolerant consensus mechanism that uses probe-based weighted information flow transmission to leverage LLMs’ intrinsic reflective and discriminative capabilities.

Result: CP-WBFT achieves superior performance across diverse network topologies under extreme Byzantine conditions (85.7% fault rate), surpassing traditional methods with remarkable accuracy on various topologies and maintaining strong reliability in mathematical reasoning and safety assessment tasks.

Conclusion: LLM-based agents demonstrate stronger skepticism and better Byzantine fault tolerance than traditional agents, and the proposed CP-WBFT mechanism effectively enhances MAS reliability across different network topologies.

Abstract: Ensuring the reliability of agent architectures and effectively identifying problematic agents when failures occur are crucial challenges in multi-agent systems (MAS). Advances in large language models (LLMs) have established LLM-based agents as a major branch of MAS, enabling major breakthroughs in complex problem solving and world modeling. However, the reliability implications of this shift remain largely unexplored. i.e., whether substituting traditional agents with LLM-based agents can effectively enhance the reliability of MAS. In this work, we investigate and quantify the reliability of LLM-based agents from the perspective of Byzantine fault tolerance. We observe that LLM-based agents demonstrate stronger skepticism when processing erroneous message flows, a characteristic that enables them to outperform traditional agents across different topological structures. Motivated by the results of the pilot experiment, we design CP-WBFT, a confidence probe-based weighted Byzantine Fault Tolerant consensus mechanism to enhance the stability of MAS with different topologies. It capitalizes on the intrinsic reflective and discriminative capabilities of LLMs by employing a probe-based, weighted information flow transmission method to improve the reliability of LLM-based agents. Extensive experiments demonstrate that CP-WBFT achieves superior performance across diverse network topologies under extreme Byzantine conditions (85.7% fault rate). Notably, our approach surpasses traditional methods by attaining remarkable accuracy on various topologies and maintaining strong reliability in both mathematical reasoning and safety assessment tasks.

Method: Modified PIBT with multi-action operations, combined with graph-guidance technique and large neighborhood search optimization.

Result: Achieves state-of-the-art performance in online LMAPF-T setting while preserving PIBT’s hallmark efficiency.

Conclusion: The enhanced PIBT successfully addresses orientation and rotation limitations while maintaining computational efficiency, making it suitable for complex MAPF scenarios requiring multi-action operations.

Abstract: PIBT is a rule-based Multi-Agent Path Finding (MAPF) solver, widely used as a low-level planner or action sampler in many state-of-the-art approaches. Its primary advantage lies in its exceptional speed, enabling action selection for thousands of agents within milliseconds by considering only the immediate next timestep. However, this short-horizon design leads to poor performance in scenarios where agents have orientation and must perform time-consuming rotation actions. In this work, we present an enhanced version of PIBT that addresses this limitation by incorporating multi-action operations. We detail the modifications introduced to improve PIBT’s performance while preserving its hallmark efficiency. Furthermore, we demonstrate how our method, when combined with graph-guidance technique and large neighborhood search optimization, achieves state-of-the-art performance in the online LMAPF-T setting.

Method: Implemented three watermarking techniques (spatial domain LSB, frequency domain DFT, hybrid LSB+DFT) in MATLAB framework, tested against JPEG compression, Gaussian noise, and salt-and-pepper noise attacks.

Result: LSB has superior imperceptibility but is fragile; DFT offers robustness but sacrifices visual quality; hybrid LSB+DFT provides optimal balance with high visual fidelity and superior resilience to all attacks.

Conclusion: Hybrid LSB+DFT watermarking technique is the most effective approach, maintaining visual quality while providing strong robustness against common image processing attacks.

Abstract: The proliferation of digital media necessitates robust methods for copyright protection and content authentication. This paper presents a comprehensive comparative study of digital image watermarking techniques implemented using the spatial domain (Least Significant Bit - LSB), the frequency domain (Discrete Fourier Transform - DFT), and a novel hybrid (LSB+DFT) approach. The core objective is to evaluate the trade-offs between imperceptibility (measured by Peak Signal-to-Noise Ratio - PSNR) and robustness (measured by Normalized Correlation - NC and Bit Error Rate - BER). We implemented these three techniques within a unified MATLAB-based experimental framework. The watermarked images were subjected to a battery of common image processing attacks, including JPEG compression, Gaussian noise, and salt-and-pepper noise, at varying intensities. Experimental results generated from standard image datasets (USC-SIPI) demonstrate that while LSB provides superior imperceptibility, it is extremely fragile. The DFT method offers significant robustness at the cost of visual quality. The proposed hybrid LSB+DFT technique, which leverages redundant embedding and a fallback extraction mechanism, is shown to provide the optimal balance, maintaining high visual fidelity while exhibiting superior resilience to all tested attacks.

Method: Two-stage approach: 1) Intra-Modality Denoising Stage extracts denoised modality-specific and shared representations from complete data; 2) Inter-Modality Complementation Stage uses these representations to compensate for missing modalities.

Result: Extensive evaluations on MOSI, MOSEI, and IEMOCAP datasets show TMDC achieves superior performance and establishes new state-of-the-art results compared to existing methods.

Conclusion: TMDC effectively addresses both missing modalities and noisy signals through its two-stage framework, demonstrating improved robustness and accuracy in multimodal sentiment analysis.

Abstract: Multimodal Sentiment Analysis (MSA) aims to infer human sentiment by integrating information from multiple modalities such as text, audio, and video. In real-world scenarios, however, the presence of missing modalities and noisy signals significantly hinders the robustness and accuracy of existing models. While prior works have made progress on these issues, they are typically addressed in isolation, limiting overall effectiveness in practical settings. To jointly mitigate the challenges posed by missing and noisy modalities, we propose a framework called Two-stage Modality Denoising and Complementation (TMDC). TMDC comprises two sequential training stages. In the Intra-Modality Denoising Stage, denoised modality-specific and modality-shared representations are extracted from complete data using dedicated denoising modules, reducing the impact of noise and enhancing representational robustness. In the Inter-Modality Complementation Stage, these representations are leveraged to compensate for missing modalities, thereby enriching the available information and further improving robustness. Extensive evaluations on MOSI, MOSEI, and IEMOCAP demonstrate that TMDC consistently achieves superior performance compared to existing methods, establishing new state-of-the-art results.

Method: Empirical analysis of speaker information distribution, followed by Time-Layer Adaptive Speaker Alignment (TLA-SA) loss that adapts to temporal and hierarchical variations.

Result: TLA-SA significantly improves speaker similarity on both research- and industrial-scale datasets and generalizes across diverse model architectures.

Conclusion: The proposed adaptive speaker alignment method effectively enhances speaker consistency in FM-based TTS systems without compromising synthesis quality.

Abstract: Flow-Matching (FM)-based zero-shot text-to-speech (TTS) systems exhibit high-quality speech synthesis and robust generalization capabilities. However, the speaker representation ability of such systems remains underexplored, primarily due to the lack of explicit speaker-specific supervision in the FM framework. To this end, we conduct an empirical analysis of speaker information distribution and reveal its non-uniform allocation across time steps and network layers, underscoring the need for adaptive speaker alignment. Accordingly, we propose Time-Layer Adaptive Speaker Alignment (TLA-SA), a loss that enhances speaker consistency by jointly leveraging temporal and hierarchical variations in speaker information. Experimental results show that TLA-SA significantly improves speaker similarity compared to baseline systems on both research- and industrial-scale datasets and generalizes effectively across diverse model architectures, including decoder-only language models (LM) and FM-based TTS systems free of LM.

Method: Time-varying beamformer weights estimated from multichannel signals using SI-SDR loss minimization, guided by continuously tracked relative transfer functions (RTFs) of the target speaker.

Result: RTF-guided models produce smoother, spatially consistent beampatterns that accurately track target direction; models without guidance fail to maintain spatial focus; estimated RTFs closely match oracle performance.

Conclusion: RTF guidance is crucial for maintaining spatial focus in dynamic environments, and the proposed tracking scheme effectively enables accurate beamforming for hearing aid applications.

Abstract: In this work, a deep beamforming framework for speech enhancement in dynamic acoustic environments is studied. The time-varying beamformer weights are estimated from the noisy multichannel signals by minimizing an SI-SDR loss. The estimation is guided by the continuously tracked relative transfer functions (RTFs) of the moving target speaker. The spatial behavior of the network is evaluated through both narrowband and wideband beampatterns under three settings: (i) oracle guidance using true RTFs, (ii) estimated RTFs obtained by a subspace tracking method, and (iii) without the RTF guidance. Results show that RTF-guided models produce smoother, spatially consistent beampatterns that accurately track the target’s direction of arrival. In contrast, the model fails to maintain a clear spatial focus when guidance is absent. Using the estimated RTFs as guidance closely matches the oracle RTF behavior, confirming the effectiveness of the tracking scheme. The model also outputs a binaural signal to preserve the speaker’s spatial cues, which promotes hearing aid and hearables applications.

Method: Curated MF-Skills dataset with rich captions and QA pairs; fine-tuned enhanced Audio Flamingo 3 backbone; post-training with MF-Think chain-of-thought dataset and GRPO-based reinforcement learning with custom rewards.

Result: Achieves state-of-the-art results across 10+ benchmarks for music understanding and reasoning, establishing itself as a generalist and musically intelligent model.

Conclusion: Sets new standard for advanced music understanding, moving from surface-level recognition to human-like perception of songs, providing benchmark and foundation for next-generation music models.

Abstract: We introduce Music Flamingo, a novel large audio-language model designed to advance music (including song) understanding in foundational audio models. While audio-language research has progressed rapidly, music remains challenging due to its dynamic, layered, and information-dense nature. Progress has been further limited by the difficulty of scaling open audio understanding models, primarily because of the scarcity of high-quality music data and annotations. As a result, prior models are restricted to producing short, high-level captions, answering only surface-level questions, and showing limited generalization across diverse musical cultures. To address these challenges, we curate MF-Skills, a large-scale dataset labeled through a multi-stage pipeline that yields rich captions and question-answer pairs covering harmony, structure, timbre, lyrics, and cultural context. We fine-tune an enhanced Audio Flamingo 3 backbone on MF-Skills and further strengthen multiple skills relevant to music understanding. To improve the model’s reasoning abilities, we introduce a post-training recipe: we first cold-start with MF-Think, a novel chain-of-thought dataset grounded in music theory, followed by GRPO-based reinforcement learning with custom rewards. Music Flamingo achieves state-of-the-art results across 10+ benchmarks for music understanding and reasoning, establishing itself as a generalist and musically intelligent audio-language model. Beyond strong empirical results, Music Flamingo sets a new standard for advanced music understanding by demonstrating how models can move from surface-level recognition toward layered, human-like perception of songs. We believe this work provides both a benchmark and a foundation for the community to build the next generation of models that engage with music as meaningfully as humans do.

Method: Builds on existing NCM framework with quasi-linear solution instead of exhaustive search, introduces novel DoA estimation across all frequency bins for improved reverberation robustness.

Result: Outperforms MUSIC in mid- to high-angle scenarios with lower angular errors, better noise rejection, interference canceling, and superior signal enhancement through beamforming.

Conclusion: The proposed joint estimation framework provides significant improvements over classical methods in both estimation accuracy and practical beamforming performance.

Abstract: We propose a joint estimation method for the Direction-of-Arrival (DoA) and the Noise Covariance Matrix (NCM) tailored for beamforming applications. Building upon an existing NCM framework, our approach simplifies the estimation procedure by deriving an quasi-linear solution, instead of the traditional exhaustive search. Additionally, we introduce a novel DoA estimation technique that operates across all frequency bins, improving robustness in reverberant environments. Simulation results demonstrate that our method outperforms classical techniques, such as MUSIC, in mid- to high-angle scenarios, achieving lower angular errors and superior signal enhancement through beamforming. The proposed framework was also fared against other techniques for signal enhancement, having better noise rejection and interference canceling capabilities. These improvements are validated using both theoretical and empirical performance metrics.

Method: Uses temporal-aware density peak clustering for adaptive speech segmentation and implicit duration coding that embeds both content and temporal span into single tokens.

Result: Achieves superior reconstruction naturalness with up to 23% fewer tokens than 40Hz fixed-frame-rate baseline, and improves word error rates and naturalness in zero-shot TTS synthesis.

Conclusion: First work demonstrating that a fully dynamic variable-frame-rate acoustic speech tokenizer can be seamlessly integrated into downstream speech language models.

Abstract: Existing speech tokenizers typically assign a fixed number of tokens per second, regardless of the varying information density or temporal fluctuations in the speech signal. This uniform token allocation mismatches the intrinsic structure of speech, where information is distributed unevenly over time. To address this, we propose VARSTok, a VAriable-frame-Rate Speech Tokenizer that adapts token allocation based on local feature similarity. VARSTok introduces two key innovations: (1) a temporal-aware density peak clustering algorithm that adaptively segments speech into variable-length units, and (2) a novel implicit duration coding scheme that embeds both content and temporal span into a single token index, eliminating the need for auxiliary duration predictors. Extensive experiments show that VARSTok significantly outperforms strong fixed-rate baselines. Notably, it achieves superior reconstruction naturalness while using up to 23% fewer tokens than a 40 Hz fixed-frame-rate baseline. VARSTok further yields lower word error rates and improved naturalness in zero-shot text-to-speech synthesis. To the best of our knowledge, this is the first work to demonstrate that a fully dynamic, variable-frame-rate acoustic speech tokenizer can be seamlessly integrated into downstream speech language models.

Method: Proposes a phase synthesizer combining frequency shifting and phase modulation in DFT filter banks, extended with variable delay lines inspired by vibrato and chorus effects.

Result: Demonstrated improvements in system stability and speech quality (measured by PESQ) in speech in-car communication using adaptive frequency-domain Kalman filter.

Conclusion: The unified phase synthesizer framework effectively decorrelates signals, enhancing system performance and speech quality in acoustic feedback scenarios.

Abstract: Undesired acoustic feedback is a known issue in communication systems, such as speech in-car communication, public address systems, or hearing aids. Without additional precautions, there is a high risk that the adaptive filter - intended to cancel the feedback path - also suppresses parts of the desired signal. One solution is to decorrelate the loudspeaker and microphone signals. In this work, we combine the two decorrelation approaches frequency shifting and phase modulation in a unified framework: a so-called \textit{phase synthesizer}, implemented in a discrete Fourier transform (DFT) filter bank. Furthermore, we extend the phase modulation technique using variable delay lines, as known from vibrato and chorus effects. We demonstrate the benefits of the proposed phase synthesizer using an example from speech in-car communication, employing an adaptive frequency-domain Kalman filter. Improvements in system stability, speech quality measured by perceptual evaluation of speech quality (PESQ) are presented.

Method: Used WHIS simulator to emulate hearing loss profiles in young normal-hearing listeners, conducted speech-in-noise experiments with IRM enhancement, and employed GESI objective intelligibility measure to predict performance.

Result: Older adults achieved higher speech intelligibility scores than average young listeners, with substantial individual variability among older adults suggesting differences in higher-level processing efficiency.

Conclusion: WHIS and GESI enable contrastive experiments between age groups independent of hearing level, providing a framework to investigate individual differences in higher-level processing in older adults.

Abstract: This paper introduces a novel approach to disentangle the effects of peripheral hearing loss (HL) and higher-level processes on speech intelligibility (SI). We conducted an SI experiment with 15 young normal-hearing (YNH) listeners using stimuli processed by the WHIS simulator to emulate the hearing loss profile of a specific older adult (OA) from a previous study involving 14 OA participants. Speech-in-noise materials were presented either with ideal ratio mask (IRM) enhancement or in an unprocessed form. Results showed that the target OA achieved higher SI scores than the average YNH listener, suggesting that the OA’s higher-level processes may perform more effectively than those of younger listeners. To examine the characteristics of the remaining OAs, we employed the GESI objective intelligibility measure to predict SI performance. GESI provided reasonably accurate predictions for both YNH and OA listeners. Using parameters estimated from the YNH experiment, we predicted SI scores for the 14 OA participants. The results revealed substantial variability: several OAs achieved higher SI scores than the average YNH listener, while one OA scored lower. These differences likely reflect individual variations in the efficiency of higher-level processing. Overall, these findings demonstrate that WHIS and GESI enable contrastive experiments between YNH and OA listeners, independent of hearing level, and offer a framework for investigating the role of higher-level processes in older adults on an individual basis.

Method: Proposes NDF approach that computes a single-channel complex mask from microphone array signals, applied to a reference microphone to approximate a virtual directional microphone with desired directivity pattern.

Result: NDF achieves frequency-invariant directivity patterns even above spatial aliasing frequency, approximates diverse higher-order patterns, enables pattern steering, generalizes to unseen conditions, and outperforms conventional beamforming.

Conclusion: The neural directional filtering approach successfully overcomes limitations of traditional beamformers for compact arrays, enabling flexible and high-performance directional sound capture.

Abstract: Beamforming with desired directivity patterns using compact microphone arrays is essential in many audio applications. Directivity patterns achievable using traditional beamformers depend on the number of microphones and the array aperture. Generally, their effectiveness degrades for compact arrays. To overcome these limitations, we propose a neural directional filtering (NDF) approach that leverages deep neural networks to enable sound capture with a predefined directivity pattern. The NDF computes a single-channel complex mask from the microphone array signals, which is then applied to a reference microphone to produce an output that approximates a virtual directional microphone with the desired directivity pattern. We introduce training strategies and propose data-dependent metrics to evaluate the directivity pattern and directivity factor. We show that the proposed method: i) achieves a frequency-invariant directivity pattern even above the spatial aliasing frequency, ii) can approximate diverse and higher-order patterns, iii) can steer the pattern in different directions, and iv) generalizes to unseen conditions. Lastly, experimental comparisons demonstrate superior performance over conventional beamforming and parametric approaches.

Method: Uses view-specific ConvNeXt encoders for frontal and lateral chest radiographs, fused with clinical indications, history, and vital signs using a transformer fusion module to generate context-aware representations.

Result: Exceeds state-of-the-art performance on both MIMIC-CXR and CXR-LT benchmarks, demonstrating improved chest X-ray classification accuracy.

Conclusion: Multimodal alignment of imaging studies with clinical data better reflects clinical reasoning and advances chest X-ray classification performance.

Abstract: Radiology is essential to modern healthcare, yet rising demand and staffing shortages continue to pose major challenges. Recent advances in artificial intelligence have the potential to support radiologists and help address these challenges. Given its widespread use and clinical importance, chest X-ray classification is well suited to augment radiologists’ workflows. However, most existing approaches rely solely on single-view, image-level inputs, ignoring the structured clinical information and multi-image studies available at the time of reporting. In this work, we introduce CaMCheX, a multimodal transformer-based framework that aligns multi-view chest X-ray studies with structured clinical data to better reflect how clinicians make diagnostic decisions. Our architecture employs view-specific ConvNeXt encoders for frontal and lateral chest radiographs, whose features are fused with clinical indications, history, and vital signs using a transformer fusion module. This design enables the model to generate context-aware representations that mirror reasoning in clinical practice. Our results exceed the state of the art for both the original MIMIC-CXR dataset and the more recent CXR-LT benchmarks, highlighting the value of clinically grounded multimodal alignment for advancing chest X-ray classification.

Method: Proposes nnQC with a Team of Experts architecture where two experts encode 3D spatial awareness and anatomical information, combined via weighted conditional module to generate pseudo-ground truth through diffusion process. Includes fingerprint adaptation for cross-organ adaptability.

Result: Evaluated on 7 organs using 12 public datasets, nnQC consistently outperformed state-of-the-art methods, even with highly degraded or missing segmentation masks, demonstrating versatility across different organs.

Conclusion: nnQC provides an effective and versatile quality control framework that self-adapts to any input organ dataset, overcoming the limitations of organ-specific QC methods and enabling reliable automated quality assessment at scale.

Abstract: Medical image segmentation using deep learning (DL) has enabled the development of automated analysis pipelines for large-scale population studies. However, state-of-the-art DL methods are prone to hallucinations, which can result in anatomically implausible segmentations. With manual correction impractical at scale, automated quality control (QC) techniques have to address the challenge. While promising, existing QC methods are organ-specific, limiting their generalizability and usability beyond their original intended task. To overcome this limitation, we propose no-new Quality Control (nnQC), a robust QC framework based on a diffusion-generative paradigm that self-adapts to any input organ dataset. Central to nnQC is a novel Team of Experts (ToE) architecture, where two specialized experts independently encode 3D spatial awareness, represented by the relative spatial position of an axial slice, and anatomical information derived from visual features from the original image. A weighted conditional module dynamically combines the pair of independent embeddings, or opinions to condition the sampling mechanism within a diffusion process, enabling the generation of a spatially aware pseudo-ground truth for predicting QC scores. Within its framework, nnQC integrates fingerprint adaptation to ensure adaptability across organs, datasets, and imaging modalities. We evaluated nnQC on seven organs using twelve publicly available datasets. Our results demonstrate that nnQC consistently outperforms state-of-the-art methods across all experiments, including cases where segmentation masks are highly degraded or completely missing, confirming its versatility and effectiveness across different organs.

Method: Integrates shifted-window vision transformer for hierarchical volumetric representation with uncertainty-aware training pipeline that uses uncertainty estimates as prompts, plus adversarial learning for ambiguous pathological regions.

Result: Demonstrates strong generalization and robustness across 11 datasets (3,884 training cases, 694 evaluation cases), trained on 17,075 3D volume-mask pairs across multiple modalities and cancer primaries.

Conclusion: SAT3D shows effectiveness in improving segmentation accuracy under challenging scenarios and has potential as a scalable foundation model for medical image analysis, with a 3D Slicer plugin for clinical translation.

Abstract: Prompt-driven vision foundation models, such as the Segment Anything Model, have recently demonstrated remarkable adaptability in computer vision. However, their direct application to medical imaging remains challenging due to heterogeneous tissue structures, imaging artefacts, and low-contrast boundaries, particularly in tumours and cancer primaries leading to suboptimal segmentation in ambiguous or overlapping lesion regions. Here, we present Segment Any Tumour 3D (SAT3D), a lightweight volumetric foundation model designed to enable robust and generalisable tumour segmentation across diverse medical imaging modalities. SAT3D integrates a shifted-window vision transformer for hierarchical volumetric representation with an uncertainty-aware training pipeline that explicitly incorporates uncertainty estimates as prompts to guide reliable boundary prediction in low-contrast regions. Adversarial learning further enhances model performance for the ambiguous pathological regions. We benchmark SAT3D against three recent vision foundation models and nnUNet across 11 publicly available datasets, encompassing 3,884 tumour and cancer cases for training and 694 cases for in-distribution evaluation. Trained on 17,075 3D volume-mask pairs across multiple modalities and cancer primaries, SAT3D demonstrates strong generalisation and robustness. To facilitate practical use and clinical translation, we developed a 3D Slicer plugin that enables interactive, prompt-driven segmentation and visualisation using the trained SAT3D model. Extensive experiments highlight its effectiveness in improving segmentation accuracy under challenging and out-of-distribution scenarios, underscoring its potential as a scalable foundation model for medical image analysis.

Method: Developed SuperRivolution framework using temporal sequences of low-resolution images, created benchmark dataset of 9,810 images, and tested multiple strategies including ensembling, super-resolution, and end-to-end temporal models.

Result: Significantly outperformed single-image methods and baseline temporal approaches, improving F1 score from 60.9% to 80.5%, with similar improvements in river width estimation.

Conclusion: Demonstrates the potential of publicly available low-resolution satellite archives for fine-scale river monitoring, narrowing the gap with high-resolution supervised models.

Abstract: Satellite missions provide valuable optical data for monitoring rivers at diverse spatial and temporal scales. However, accessibility remains a challenge: high-resolution imagery is ideal for fine-grained monitoring but is typically scarce and expensive compared to low-resolution imagery. To address this gap, we introduce SuperRivolution, a framework that improves river segmentation resolution by leveraging information from time series of low-resolution satellite images. We contribute a new benchmark dataset of 9,810 low-resolution temporal images paired with high-resolution labels from an existing river monitoring dataset. Using this benchmark, we investigate multiple strategies for river segmentation, including ensembling single-image models, applying image super-resolution, and developing end-to-end models trained on temporal sequences. SuperRivolution significantly outperforms single-image methods and baseline temporal approaches, narrowing the gap with supervised high-resolution models. For example, the F1 score for river segmentation improves from 60.9% to 80.5%, while the state-of-the-art model operating on high-resolution images achieves 94.1%. Similar improvements are also observed in river width estimation tasks. Our results highlight the potential of publicly available low-resolution satellite archives for fine-scale river monitoring.

Method: Developed PV-DDPM, a spatially conditioned denoising diffusion probabilistic model that generates anomalous EL images across four PV cell types using binary masks for structural features and defect positions.

Result: Generated images achieved FID of 4.10 and KID of 0.0023 ± 0.0007. Training AA-CLIP on the enhanced E-SCDD dataset improved pixel-level AUC by 1.70 points and average precision by 8.34 points compared to original SCDD.

Conclusion: PV-DDPM is the first framework to jointly model multiple PV cell types with diverse anomaly generation, significantly enhancing anomaly detection performance through synthetic data augmentation.

Abstract: Reliable anomaly detection in photovoltaic (PV) modules is critical for maintaining solar energy efficiency. However, developing robust computer vision models for PV inspection is constrained by the scarcity of large-scale, diverse, and balanced datasets. This study introduces PV-DDPM, a spatially conditioned denoising diffusion probabilistic model that generates anomalous electroluminescence (EL) images across four PV cell types: multi-crystalline silicon (multi-c-Si), mono-crystalline silicon (mono-c-Si), half-cut multi-c-Si, and interdigitated back contact (IBC) with dogbone interconnect. PV-DDPM enables controlled synthesis of single-defect and multi-defect scenarios by conditioning on binary masks representing structural features and defect positions. To the best of our knowledge, this is the first framework that jointly models multiple PV cell types while supporting simultaneous generation of diverse anomaly types. We also introduce E-SCDD, an enhanced version of the SCDD dataset, comprising 1,000 pixel-wise annotated EL images spanning 30 semantic classes, and 1,768 unlabeled synthetic samples. Quantitative evaluation shows our generated images achieve a Fréchet Inception Distance (FID) of 4.10 and Kernel Inception Distance (KID) of 0.0023 $\pm$ 0.0007 across all categories. Training the vision–language anomaly detection model AA-CLIP on E-SCDD, compared to the SCDD dataset, improves pixel-level AUC and average precision by 1.70 and 8.34 points, respectively.

Method: The proposed TomoGraphView framework integrates omnidirectional volume slicing with spherical graph-based feature aggregation, addressing limitations of conventional axial/sagittal/coronal slicing and slice-wise aggregation methods.

Result: The authors have developed an accessible code base and user-friendly library for omnidirectional volume slicing, making the framework publicly available for research and clinical use.

Conclusion: TomoGraphView provides a more effective approach for leveraging 2D vision foundation models in 3D medical imaging by preserving volumetric structure and spatial coherence through omnidirectional slicing and graph-based aggregation.

Abstract: The growing number of medical tomography examinations has necessitated the development of automated methods capable of extracting comprehensive imaging features to facilitate downstream tasks such as tumor characterization, while assisting physicians in managing their growing workload. However, 3D medical image classification remains a challenging task due to the complex spatial relationships and long-range dependencies inherent in volumetric data. Training models from scratch suffers from low data regimes, and the absence of 3D large-scale multimodal datasets has limited the development of 3D medical imaging foundation models. Recent studies, however, have highlighted the potential of 2D vision foundation models, originally trained on natural images, as powerful feature extractors for medical image analysis. Despite these advances, existing approaches that apply 2D models to 3D volumes via slice-based decomposition remain suboptimal. Conventional volume slicing strategies, which rely on canonical planes such as axial, sagittal, or coronal, may inadequately capture the spatial extent of target structures when these are misaligned with standardized viewing planes. Furthermore, existing slice-wise aggregation strategies rarely account for preserving the volumetric structure, resulting in a loss of spatial coherence across slices. To overcome these limitations, we propose TomoGraphView, a novel framework that integrates omnidirectional volume slicing with spherical graph-based feature aggregation. We publicly share our accessible code base at http://github.com/compai-lab/2025-MedIA-kiechle and provide a user-friendly library for omnidirectional volume slicing at https://pypi.org/project/OmniSlicer.

Method: Uses adaptive brightness adjustment (ABA) preprocessing, multi-scale temporal consistency-aware loss, occlusion-aware masking, reverse inference strategy, and self-ensemble mechanism.

Result: Achieves state-of-the-art performance with up to 29.7% PSNR gain over prior methods, demonstrating superior perceptual quality and temporal consistency.

Conclusion: TempRetinex effectively exploits inter-frame correlations and improves robustness to diverse lighting scenarios, making it a powerful framework for unsupervised video enhancement.

Abstract: Videos inherently contain rich temporal information that provides complementary cues for low-light enhancement beyond what can be achieved with single images. We propose TempRetinex, a novel unsupervised Retinex-based framework that effectively exploits inter-frame correlations for video enhancement. To address the poor generalization of existing unsupervised methods under varying illumination, we introduce adaptive brightness adjustment (ABA) preprocessing that explicitly aligns lighting distributions across exposures. This significantly improves model robustness to diverse lighting scenarios and eases training optimization, leading to better denoising performance. For enhanced temporal coherence, we propose a multi-scale temporal consistency-aware loss to enforce multiscale similarity between consecutive frames, and an occlusion-aware masking technique to handle complex motions. We further incorporate a reverse inference strategy to refine unconverged frames and a self-ensemble (SE) mechanism to boost the denoising across diverse textures. Experiments demonstrate that TempRetinex achieves state-of-the-art performance in both perceptual quality and temporal consistency, achieving up to a 29.7% PSNR gain over prior methods.

Method: Developed a two-channel input system with user-defined trade-off between channels, integrated pixel- and FFT-based loss functions, and trained U-net model for denoising.

Result: Compared GDM with non-FFT denoising model on STM, AFM, and SEM images, showing improved preservation of physically important weak features.

Conclusion: The proposed workflow can be extended to improve other microscopy image quality and provides design flexibility for domain experts to tune preferences.

Abstract: Microscopy such as Scanning Tunneling Microscopy (STM), Atomic Force Microscopy (AFM) and Scanning Electron Microscopy (SEM) are essential tools in material imaging at micro- and nanoscale resolutions to extract physical knowledge and materials structure-property relationships. However, tuning microscopy controls (e.g. scanning speed, current setpoint, tip bias etc.) to obtain a high-quality of images is a non-trivial and time-consuming effort. On the other hand, with sub-standard images, the key features are not accurately discovered due to noise and artifacts, leading to erroneous analysis. Existing denoising models mostly build on generalizing the weak signals as noises while the strong signals are enhanced as key features, which is not always the case in microscopy images, thus can completely erase a significant amount of hidden physical information. To address these limitations, we propose a global denoising model (GDM) to smartly remove artifacts of microscopy images while preserving weaker but physically important features. The proposed model is developed based on 1) first designing a two-imaging input channel of non-pair and goal specific pre-processed images with user-defined trade-off information between two channels and 2) then integrating a loss function of pixel- and fast Fourier-transformed (FFT) based on training the U-net model. We compared the proposed GDM with the non-FFT denoising model over STM-generated images of Copper(Cu) and Silicon(Si) materials, AFM-generated Pantoea sp.YR343 bio-film images and SEM-generated plastic degradation images. We believe this proposed workflow can be extended to improve other microscopy image quality and will benefit the experimentalists with the proposed design flexibility to smartly tune via domain-experts preferences.

Method: Uses TDM-MIMO radar with phase correlation registration of BP images for precise relative positioning, then applies relative motion compensation and iterative inversion on multi-cycle MIMO data. Integrates CC-CSI into the optimization framework as RMC-CC-CSI.

Result: RMC-CC-CSI demonstrates accelerated convergence, enhanced reconstruction fidelity, and improved noise immunity compared to conventional CC-CSI for stationary targets, though with increased computational cost.

Conclusion: The proposed scheme provides an effective general framework for electromagnetic inversion of moving targets, with RMC-CC-CSI showing superior performance over traditional methods despite higher computational requirements.

Abstract: An novel electromagnetic quantitative inversion scheme for translationally moving targets via phase correlation registration of back-projection (BP) images is proposed. Based on a time division multiplexing multiple-input multiple-output (TDM-MIMO) radar architecture, the scheme first achieves high-precision relative positioning of the target, then applies relative motion compensation to perform iterative inversion on multi-cycle MIMO measurement data, thereby reconstructing the target’s electromagnetic parameters. As a general framework compatible with other mainstream inversion algorithms, we exemplify our approach by incorporating the classical cross-correlated contrast source inversion (CC-CSI) into iterative optimization step of the scheme, resulting in a new algorithm termed RMC-CC-CSI. Numerical and experimental results demonstrate that RMC-CC-CSI offers accelerated convergence, enhanced reconstruction fidelity, and improved noise immunity over conventional CC-CSI for stationary targets despite increased computational cost.

Method: A physics-informed data augmentation scheme where a trained network generates phase diverse pseudo-data based on ground truth data, leveraging implicit local correlation between phase diverse measurements in optical imaging.

Result: The approach provides high quality inverse solutions with simpler reconstruction algorithms for both incoherent and coherent optical imaging configurations using vortex phase as diversity mechanism.

Conclusion: This method may enable leaner high-fidelity computational imaging systems across various applications by reducing hardware complexity while maintaining reconstruction quality.

Abstract: Inverse problems in imaging are typically ill-posed and are usually solved by employing regularized optimization techniques. The usage of appropriate constraints can restrict the solution space, thus making it feasible for a reconstruction algorithm to find a meaningful solution. In recent years, deep network based ideas aimed at learning the end-to-end mapping between the raw measurements and the target image have gained popularity. In the learning approach, the functional relationship between the measured raw data and the solution image are learned by training a deep network with prior examples. While this approach allows one to significantly increase the real-time operational speed, it does not change the nature of the underlying ill-posed inverse problem. It is well-known that availability of diverse non-redundant data via additional measurements can generically improve the robustness of the reconstruction algorithms. The multiple data measurements, however, typically demand additional hardware and complex system setups that are not desirable. In this work, we note that in both incoherent and coherent optical imaging, the irradiance patterns corresponding to two phase diverse measurements associated with the same test object have implicit local correlation which may be learned. A physics informed data augmentation scheme is then described where a trained network is used for generating a phase diverse pseudo-data based on a ground truth data frame. The true data along with the augmented pesudo-data are observed to provide high quality inverse solutions with simpler reconstruction algorithms. We validate this approach for both incoherent and coherent optical imaging (or phase retrieval) configurations with vortex phase as a diversity mechanism. Our results may open new avenues for leaner high-fidelity computational imaging systems across a broad range of applications.

Method: Employed CNN-based approaches with focus on dataset curation, preprocessing strategies, and model architecture optimization. Implemented voting system and developed customized CNN models tailored for ROP detection.

Result: Customized CNN models outperformed pre-trained counterparts with higher accuracy and F1-scores. Voting system further enhanced performance. Models demonstrated reduced computational burden and feasibility for deployment in clinical software/hardware configurations.

Conclusion: The study successfully demonstrates the efficacy of deep learning models in improving ROP diagnostic precision and efficiency, providing valuable diagnostic aids for clinical applications.

Abstract: This paper encompasses an in-depth examination of Retinopathy of Prematurity (ROP) diagnosis, employing advanced deep learning methodologies. Our focus centers on refining and evaluating CNN-based approaches for precise and efficient ROP detection. We navigate the complexities of dataset curation, preprocessing strategies, and model architecture, aligning with research objectives encompassing model effectiveness, computational cost analysis, and time complexity assessment. Results underscore the supremacy of tailored CNN models over pre-trained counterparts, evident in heightened accuracy and F1-scores. Implementation of a voting system further enhances performance. Additionally, our study reveals the potential of the proposed customized CNN model to alleviate computational burdens associated with deep neural networks. Furthermore, we showcase the feasibility of deploying these models within dedicated software and hardware configurations, highlighting their utility as valuable diagnostic aids in clinical settings. In summary, our discourse significantly contributes to ROP diagnosis, unveiling the efficacy of deep learning models in enhancing diagnostic precision and efficiency.

Method: Developed Equivariant Regularization by Denoising (ERED) and Equivariant Plug-and-Play (EPnP) frameworks based on equivariant denoisers and stochastic optimization.

Result: The authors analyze the convergence of their proposed algorithms and discuss their practical benefits for image restoration tasks.

Conclusion: The proposed equivariant frameworks provide a systematic way to incorporate transformation invariance into image restoration, overcoming limitations of standard deep architectures.

Abstract: One key ingredient of image restoration is to define a realistic prior on clean images to complete the missing information in the observation. State-of-the-art restoration methods rely on a neural network to encode this prior. Typical image distributions are invariant to some set of transformations, such as rotations or flips. However, most deep architectures are not designed to represent an invariant image distribution. Recent works have proposed to overcome this difficulty by including equivariance properties within a Plug-and-Play paradigm. In this work, we propose two unified frameworks named Equivariant Regularization by Denoising (ERED) and Equivariant Plug-and-Play (EPnP) based on equivariant denoisers and stochastic optimization. We analyze the convergence of the proposed algorithms and discuss their practical benefit.

Method: Uses shallow discriminator architectures with smaller receptive fields to learn pristine-to-distorted mapping through unpaired learning, focusing on accurately reproducing local distortion characteristics.

Result: Achieves mean average precision increases up to 0.15 for individual distortions and 0.16 for camera-specific characteristics, with 20x parameter reduction compared to some state-of-the-art methods.

Conclusion: Smaller receptive fields in shallow architectures effectively learn unknown image distortions with superior efficiency without compromising performance in domain adaptation tasks.

Abstract: Each image acquisition setup leads to its own camera-specific image characteristics degrading the image quality. In learning-based perception algorithms, characteristics occurring during the application phase, but absent in the training data, lead to a domain gap impeding the performance. Previously, pixel-level domain adaptation through unpaired learning of the pristine-to-distorted mapping function has been proposed. In this work, we propose shallow discriminator architectures to address limitations of these approaches. We show that a smaller receptive field size improves learning of unknown image distortions by more accurately reproducing local distortion characteristics at a low network complexity. In a domain adaptation setup for instance segmentation, we achieve mean average precision increases over previous methods of up to 0.15 for individual distortions and up to 0.16 for camera-specific image characteristics in a simplified camera model. In terms of number of parameters, our approach matches the complexity of one state of the art method while reducing complexity by a factor of 20 compared to another, demonstrating superior efficiency without compromising performance.

Method: Applied self-supervision principle in projection domain where noise is element-wise independent, optimizing FBP filtering and denoiser neural network parameters using low-dose sinograms as training targets.

Result: Outperforms both conventional and compressed sensing based iterative reconstruction methods qualitatively and quantitatively in reconstruction of analytic CT phantoms and real-world CT images.

Conclusion: Self-supervised training scheme enables effective low-dose CT reconstruction without requiring labeled data, achieving superior performance compared to existing methods.

Abstract: Ionizing radiation has been the biggest concern in CT imaging. To reduce the dose level without compromising the image quality, low-dose CT reconstruction has been offered with the availability of compressed sensing based reconstruction methods. Recently, data-driven methods got attention with the rise of deep learning, the availability of high computational power, and big datasets. Deep learning based methods have also been used in low-dose CT reconstruction problem in different manners. Usually, the success of these methods depends on labeled data. However, recent studies showed that training can be achieved successfully with noisy datasets. In this study, we defined a training scheme to use low-dose sinograms as their own training targets. We applied the self-supervision principle in the projection domain where the noise is element-wise independent which is a requirement for self-supervised training methods. Using the self-supervised training, the filtering part of the FBP method and the parameters of a denoiser neural network are optimized. We demonstrate that our method outperforms both conventional and compressed sensing based iterative reconstruction methods qualitatively and quantitatively in the reconstruction of analytic CT phantoms and real-world CT images in low-dose CT reconstruction task.

Method: Uses randomly initialized generative CNNs as priors since they generate patterns easier than noise. Implements different loss function variants for regularization without training.

Result: Outperforms FBP, SART, and TV-regularized SART methods both qualitatively and quantitatively on analytical phantoms and human CT images with different views.

Conclusion: Randomly initialized CNNs serve as effective priors for CT reconstruction, enabling data-free regularization that achieves superior performance over conventional methods.

Abstract: Low-dose CT imaging requires reconstruction from noisy indirect measurements which can be defined as an ill-posed linear inverse problem. In addition to conventional FBP method in CT imaging, recent compressed sensing based methods exploit handcrafted priors which are mostly simplistic and hard to determine. More recently, deep learning (DL) based methods have become popular in medical imaging field. In CT imaging, DL based methods try to learn a function that maps low-dose images to normal-dose images. Although the results of these methods are promising, their success mostly depends on the availability of high-quality massive datasets. In this study, we proposed a method that does not require any training data or a learning process. Our method exploits such an approach that deep convolutional neural networks (CNNs) generate patterns easier than the noise, therefore randomly initialized generative neural networks can be suitable priors to be used in regularizing the reconstruction. In the experiments, the proposed method is implemented with different loss function variants. Both analytical CT phantoms and human CT images are used with different views. Conventional FBP method, a popular iterative method (SART), and TV regularized SART are used in the comparisons. We demonstrated that our method with different loss function variants outperforms the other methods both qualitatively and quantitatively.

Method: Developed an approximate Bayesian estimation using a probabilistic model that accounts for spatially correlated label errors. Introduced ECCD (ELBO-Computable Correlated Discrete Distribution) which represents discrete dependencies through continuous latent Gaussian fields with KMS-structured covariance for scalable variational inference.

Result: Experimental results on multiple segmentation tasks show significant performance improvements by leveraging spatial correlation of label errors. In lung segmentation tasks, the method achieves performance comparable to training with clean labels under moderate noise levels.

Conclusion: The proposed framework effectively handles spatially correlated label errors in semantic segmentation through efficient Bayesian inference, demonstrating practical value for scenarios where obtaining clean annotations is challenging.

Abstract: In semantic segmentation, the accuracy of models heavily depends on the high-quality annotations. However, in many practical scenarios, such as medical imaging and remote sensing, obtaining true annotations is not straightforward and usually requires significant human labor. Relying on human labor often introduces annotation errors, including mislabeling, omissions, and inconsistency between annotators. In the case of remote sensing, differences in procurement time can lead to misaligned ground-truth annotations. These label errors are not independently distributed, and instead usually appear in spatially connected regions where adjacent pixels are more likely to share the same errors. To address these issues, we propose an approximate Bayesian estimation based on a probabilistic model that assumes training data include label errors, incorporating the tendency for these errors to occur with spatial correlations between adjacent pixels. However, Bayesian inference for such spatially correlated discrete variables is notoriously intractable. To overcome this fundamental challenge, we introduce a novel class of probabilistic models, which we term the ELBO-Computable Correlated Discrete Distribution (ECCD). By representing the discrete dependencies through a continuous latent Gaussian field with a Kac-Murdock-Szegö (KMS) structured covariance, our framework enables scalable and efficient variational inference for problems previously considered computationally prohibitive. Through experiments on multiple segmentation tasks, we confirm that leveraging the spatial correlation of label errors significantly improves performance. Notably, in specific tasks such as lung segmentation, the proposed method achieves performance comparable to training with clean labels under moderate noise levels. Code is available at https://github.com/pfnet-research/Bayesian_SpatialCorr.

Method: Uses Cosmos video tokenizer to extract discrete and continuous tokens. Discrete tokens are partially masked and compressed losslessly with discrete checkerboard context model. Continuous tokens are quantized and compressed with continuous checkerboard context model. Both streams are fused using ControlNet-based multi-scale integration module.

Result: The framework enables effective video compression at ultra-low bit rates while maintaining high perceptual quality and stable fidelity in reconstruction.

Conclusion: This work demonstrates the practicality of tokenized video compression and opens new directions for semantics-aware, token-native approaches in video compression.

Abstract: Tokenized visual representations have shown promise in image compression, yet their extension to video remains underexplored due to the challenges posed by complex temporal dynamics and stringent bit rate constraints. In this paper, we present tokenized video compression (TVC), a token-based dual-stream framework designed to operate effectively at ultra-low bit rates. TVC leverages the Cosmos video tokenizer to extract both discrete and continuous token streams. The discrete tokens are partially masked using a strategic masking scheme and then compressed losslessly with a discrete checkerboard context model to reduce transmission overhead. The masked tokens are reconstructed by a decoder-only Transformer with spatiotemporal token prediction. In parallel, the continuous tokens are quantized and compressed using a continuous checkerboard context model, providing complementary continuous information at ultra-low bit rates. At the decoder side, the two streams are fused with a ControlNet-based multi-scale integration module, ensuring high perceptual quality alongside stable fidelity in reconstruction. Overall, this work illustrates the practicality of tokenized video compression and points to new directions for semantics-aware, token-native approaches.

Method: The challenge provides a large publicly available dataset, baseline model, and evaluation framework. Submitted methods include state space models, recurrent models, registration-driven volume refinement, attention mechanisms, and physics-informed models.

Result: The challenge attracted 43 registered teams with 6 teams submitting 21 valid dockerized solutions. Comprehensive comparative analysis was performed across multiple evaluation metrics to highlight progress and limitations of current approaches.

Conclusion: The challenge serves as a live and evolving benchmark that will be continuously iterated and improved, with plans to continue at MICCAI 2025, reflecting sustained commitment to advancing trackerless freehand ultrasound reconstruction research.

Abstract: Trackerless freehand ultrasound reconstruction aims to reconstruct 3D volumes from sequences of 2D ultrasound images without relying on external tracking systems. By eliminating the need for optical or electromagnetic trackers, this approach offers a low-cost, portable, and widely deployable alternative to more expensive volumetric ultrasound imaging systems, particularly valuable in resource-constrained clinical settings. However, predicting long-distance transformations and handling complex probe trajectories remain challenging. The TUS-REC2024 Challenge establishes the first benchmark for trackerless 3D freehand ultrasound reconstruction by providing a large publicly available dataset, along with a baseline model and a rigorous evaluation framework. By the submission deadline, the Challenge had attracted 43 registered teams, of which 6 teams submitted 21 valid dockerized solutions. The submitted methods span a wide range of approaches, including the state space model, the recurrent model, the registration-driven volume refinement, the attention mechanism, and the physics-informed model. This paper provides a comprehensive background introduction and literature review in the field, presents an overview of the challenge design and dataset, and offers a comparative analysis of submitted methods across multiple evaluation metrics. These analyses highlight both the progress and the current limitations of state-of-the-art approaches in this domain and provide insights for future research directions. All data and code are publicly available to facilitate ongoing development and reproducibility. As a live and evolving benchmark, it is designed to be continuously iterated and improved. The Challenge was held at MICCAI 2024 and is organised again at MICCAI 2025, reflecting its sustained commitment to advancing this field.

Method: Proposes Topology Graph Consistency (TGC) framework that integrates graph-theoretic constraints by aligning Laplacian spectra, component counts, and adjacency statistics between prediction graphs and references.

Result: Experiments on GlaS and CRAG datasets show TGC achieves state-of-the-art performance under 5-10% supervision and significantly narrows the gap to full supervision.

Conclusion: TGC effectively enforces global topology in semi-supervised semantic segmentation, improving accuracy and reducing the need for extensive annotations in computational pathology.

Abstract: Semi-supervised semantic segmentation (SSSS) is vital in computational pathology, where dense annotations are costly and limited. Existing methods often rely on pixel-level consistency, which propagates noisy pseudo-labels and produces fragmented or topologically invalid masks. We propose Topology Graph Consistency (TGC), a framework that integrates graph-theoretic constraints by aligning Laplacian spectra, component counts, and adjacency statistics between prediction graphs and references. This enforces global topology and improves segmentation accuracy. Experiments on GlaS and CRAG demonstrate that TGC achieves state-of-the-art performance under 5-10% supervision and significantly narrows the gap to full supervision.